Lines Matching full:params

19   union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)],  in xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params()
24params->fp32_scalar_fmagic.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params()
25params->fp32_scalar_fmagic.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params()
26 params->fp32_scalar_fmagic.magic_bias = 12582912.0f; in xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params()
27params->fp32_scalar_fmagic.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) out… in xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params()
28 return sizeof(params->fp32_scalar_fmagic); in xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params()
32 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params()
39 params->fp32_scalar_imagic.magic_bias = 12582912.0f; in xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params()
40params->fp32_scalar_imagic.magic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zer… in xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params()
41params->fp32_scalar_imagic.magic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zer… in xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params()
42params->fp32_scalar_imagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zer… in xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params()
43 return sizeof(params->fp32_scalar_imagic); in xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params()
47 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params()
52params->fp32_scalar_lrintf.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params()
53params->fp32_scalar_lrintf.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params()
54 params->fp32_scalar_lrintf.output_zero_point = (int32_t) output_zero_point; in xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params()
55 return sizeof(params->fp32_scalar_lrintf); in xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params()
60 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_sse2_params()
67 params->fp32_sse2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qc8_conv_minmax_fp32_sse2_params()
70 params->fp32_sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qc8_conv_minmax_fp32_sse2_params()
71 params->fp32_sse2.output_min[i] = (int16_t) output_min; in xnn_init_qc8_conv_minmax_fp32_sse2_params()
73 return sizeof(params->fp32_sse2); in xnn_init_qc8_conv_minmax_fp32_sse2_params()
77 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_sse4_params()
84 params->fp32_sse4.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qc8_conv_minmax_fp32_sse4_params()
87 params->fp32_sse4.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qc8_conv_minmax_fp32_sse4_params()
90 params->fp32_sse4.output_min[i] = output_min; in xnn_init_qc8_conv_minmax_fp32_sse4_params()
92 return sizeof(params->fp32_sse4); in xnn_init_qc8_conv_minmax_fp32_sse4_params()
96 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_avx2_params()
103 params->fp32_avx2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qc8_conv_minmax_fp32_avx2_params()
106 params->fp32_avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qc8_conv_minmax_fp32_avx2_params()
109 params->fp32_avx2.output_min[i] = output_min; in xnn_init_qc8_conv_minmax_fp32_avx2_params()
111 return sizeof(params->fp32_avx2); in xnn_init_qc8_conv_minmax_fp32_avx2_params()
115 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_avx512_params()
122 params->fp32_avx512.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qc8_conv_minmax_fp32_avx512_params()
125 params->fp32_avx512.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qc8_conv_minmax_fp32_avx512_params()
128 params->fp32_avx512.output_min[i] = output_min; in xnn_init_qc8_conv_minmax_fp32_avx512_params()
130 return sizeof(params->fp32_avx512); in xnn_init_qc8_conv_minmax_fp32_avx512_params()
136 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_armsimd32_params()
141 params->fp32_armsimd32.magic_bias = 12582912.0f; in xnn_init_qc8_conv_minmax_fp32_armsimd32_params()
142params->fp32_armsimd32.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_po… in xnn_init_qc8_conv_minmax_fp32_armsimd32_params()
143 params->fp32_armsimd32.output_min = (uint32_t) (uint8_t) output_min * UINT32_C(0x01010101); in xnn_init_qc8_conv_minmax_fp32_armsimd32_params()
144 params->fp32_armsimd32.output_max = (uint32_t) (uint8_t) output_max * UINT32_C(0x01010101); in xnn_init_qc8_conv_minmax_fp32_armsimd32_params()
145 return sizeof(params->fp32_armsimd32); in xnn_init_qc8_conv_minmax_fp32_armsimd32_params()
151 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_neon_params()
156 params->fp32_neon.magic_bias = 12582912.0f; in xnn_init_qc8_conv_minmax_fp32_neon_params()
157params->fp32_neon.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_… in xnn_init_qc8_conv_minmax_fp32_neon_params()
158 params->fp32_neon.output_min = output_min; in xnn_init_qc8_conv_minmax_fp32_neon_params()
159 params->fp32_neon.output_max = output_max; in xnn_init_qc8_conv_minmax_fp32_neon_params()
160 return sizeof(params->fp32_neon); in xnn_init_qc8_conv_minmax_fp32_neon_params()
164 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_neonv8_params()
169 params->fp32_neonv8.output_zero_point = (int16_t) output_zero_point; in xnn_init_qc8_conv_minmax_fp32_neonv8_params()
170 params->fp32_neonv8.output_min = output_min; in xnn_init_qc8_conv_minmax_fp32_neonv8_params()
171 params->fp32_neonv8.output_max = output_max; in xnn_init_qc8_conv_minmax_fp32_neonv8_params()
172 return sizeof(params->fp32_neonv8); in xnn_init_qc8_conv_minmax_fp32_neonv8_params()
178 union xnn_qc8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qc8_conv_minmax_fp32_wasmsimd_params()
187 params->fp32_wasmsimd.magic_bias[i] = 12582912.0f; in xnn_init_qc8_conv_minmax_fp32_wasmsimd_params()
188 params->fp32_wasmsimd.magic_min[i] = magic_min; in xnn_init_qc8_conv_minmax_fp32_wasmsimd_params()
189 params->fp32_wasmsimd.magic_bias_less_output_zero_point[i] = magic_bias_less_zero_point; in xnn_init_qc8_conv_minmax_fp32_wasmsimd_params()
192 params->fp32_wasmsimd.output_max[i] = output_max; in xnn_init_qc8_conv_minmax_fp32_wasmsimd_params()
194 return sizeof(params->fp32_wasmsimd); in xnn_init_qc8_conv_minmax_fp32_wasmsimd_params()
199 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params()
208 params->fp32_scalar_fmagic.scale = scale; in xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params()
209params->fp32_scalar_fmagic.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params()
210params->fp32_scalar_fmagic.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params()
211 params->fp32_scalar_fmagic.magic_bias = 12582912.0f; in xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params()
212params->fp32_scalar_fmagic.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) out… in xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params()
213 return sizeof(params->fp32_scalar_fmagic); in xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params()
217 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params()
228 params->fp32_scalar_imagic.scale = scale; in xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params()
229 params->fp32_scalar_imagic.magic_bias = 12582912.0f; in xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params()
230params->fp32_scalar_imagic.magic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zer… in xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params()
231params->fp32_scalar_imagic.magic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zer… in xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params()
232params->fp32_scalar_imagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zer… in xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params()
233 return sizeof(params->fp32_scalar_imagic); in xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params()
237 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params()
246 params->fp32_scalar_lrintf.scale = scale; in xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params()
247params->fp32_scalar_lrintf.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params()
248params->fp32_scalar_lrintf.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params()
249 params->fp32_scalar_lrintf.output_zero_point = (int32_t) output_zero_point; in xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params()
250 return sizeof(params->fp32_scalar_lrintf); in xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params()
255 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_sse2_params()
266 params->fp32_sse2.scale[i] = scale; in xnn_init_qs8_conv_minmax_fp32_sse2_params()
267 params->fp32_sse2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qs8_conv_minmax_fp32_sse2_params()
270 params->fp32_sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_conv_minmax_fp32_sse2_params()
271 params->fp32_sse2.output_min[i] = (int16_t) output_min; in xnn_init_qs8_conv_minmax_fp32_sse2_params()
273 return sizeof(params->fp32_sse2); in xnn_init_qs8_conv_minmax_fp32_sse2_params()
277 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_sse4_params()
288 params->fp32_sse4.scale[i] = scale; in xnn_init_qs8_conv_minmax_fp32_sse4_params()
289 params->fp32_sse4.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qs8_conv_minmax_fp32_sse4_params()
292 params->fp32_sse4.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_conv_minmax_fp32_sse4_params()
295 params->fp32_sse4.output_min[i] = output_min; in xnn_init_qs8_conv_minmax_fp32_sse4_params()
297 return sizeof(params->fp32_sse4); in xnn_init_qs8_conv_minmax_fp32_sse4_params()
301 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_avx2_params()
312 params->fp32_avx2.scale[i] = scale; in xnn_init_qs8_conv_minmax_fp32_avx2_params()
313 params->fp32_avx2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qs8_conv_minmax_fp32_avx2_params()
316 params->fp32_avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_conv_minmax_fp32_avx2_params()
319 params->fp32_avx2.output_min[i] = output_min; in xnn_init_qs8_conv_minmax_fp32_avx2_params()
321 return sizeof(params->fp32_avx2); in xnn_init_qs8_conv_minmax_fp32_avx2_params()
325 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_avx512_params()
336 params->fp32_avx512.scale[i] = scale; in xnn_init_qs8_conv_minmax_fp32_avx512_params()
337 params->fp32_avx512.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qs8_conv_minmax_fp32_avx512_params()
340 params->fp32_avx512.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_conv_minmax_fp32_avx512_params()
343 params->fp32_avx512.output_min[i] = output_min; in xnn_init_qs8_conv_minmax_fp32_avx512_params()
345 return sizeof(params->fp32_avx512); in xnn_init_qs8_conv_minmax_fp32_avx512_params()
351 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_armsimd32_params()
360 params->fp32_armsimd32.scale = scale; in xnn_init_qs8_conv_minmax_fp32_armsimd32_params()
361 params->fp32_armsimd32.magic_bias = 12582912.0f; in xnn_init_qs8_conv_minmax_fp32_armsimd32_params()
362params->fp32_armsimd32.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_po… in xnn_init_qs8_conv_minmax_fp32_armsimd32_params()
363 params->fp32_armsimd32.output_min = (uint32_t) (uint8_t) output_min * UINT32_C(0x01010101); in xnn_init_qs8_conv_minmax_fp32_armsimd32_params()
364 params->fp32_armsimd32.output_max = (uint32_t) (uint8_t) output_max * UINT32_C(0x01010101); in xnn_init_qs8_conv_minmax_fp32_armsimd32_params()
365 return sizeof(params->fp32_armsimd32); in xnn_init_qs8_conv_minmax_fp32_armsimd32_params()
371 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_neon_params()
380 params->fp32_neon.scale = scale; in xnn_init_qs8_conv_minmax_fp32_neon_params()
381 params->fp32_neon.magic_bias = 12582912.0f; in xnn_init_qs8_conv_minmax_fp32_neon_params()
382params->fp32_neon.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_… in xnn_init_qs8_conv_minmax_fp32_neon_params()
383 params->fp32_neon.output_min = output_min; in xnn_init_qs8_conv_minmax_fp32_neon_params()
384 params->fp32_neon.output_max = output_max; in xnn_init_qs8_conv_minmax_fp32_neon_params()
385 return sizeof(params->fp32_neon); in xnn_init_qs8_conv_minmax_fp32_neon_params()
389 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_neonv8_params()
398 params->fp32_neonv8.scale = scale; in xnn_init_qs8_conv_minmax_fp32_neonv8_params()
399 params->fp32_neonv8.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_conv_minmax_fp32_neonv8_params()
400 params->fp32_neonv8.output_min = output_min; in xnn_init_qs8_conv_minmax_fp32_neonv8_params()
401 params->fp32_neonv8.output_max = output_max; in xnn_init_qs8_conv_minmax_fp32_neonv8_params()
402 return sizeof(params->fp32_neonv8); in xnn_init_qs8_conv_minmax_fp32_neonv8_params()
406 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_rndnu_neon_params()
432 params->rndnu_neon.right_pre_shift = -pre_shift; in xnn_init_qs8_conv_minmax_rndnu_neon_params()
433 params->rndnu_neon.multiplier = multiplier; in xnn_init_qs8_conv_minmax_rndnu_neon_params()
434 params->rndnu_neon.right_post_shift = -post_shift; in xnn_init_qs8_conv_minmax_rndnu_neon_params()
435 params->rndnu_neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_conv_minmax_rndnu_neon_params()
436 params->rndnu_neon.output_min = output_min; in xnn_init_qs8_conv_minmax_rndnu_neon_params()
437 params->rndnu_neon.output_max = output_max; in xnn_init_qs8_conv_minmax_rndnu_neon_params()
438 return sizeof(params->rndnu_neon); in xnn_init_qs8_conv_minmax_rndnu_neon_params()
444 union xnn_qs8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_conv_minmax_fp32_wasmsimd_params()
457 params->fp32_wasmsimd.scale[i] = scale; in xnn_init_qs8_conv_minmax_fp32_wasmsimd_params()
458 params->fp32_wasmsimd.magic_bias[i] = 12582912.0f; in xnn_init_qs8_conv_minmax_fp32_wasmsimd_params()
459 params->fp32_wasmsimd.magic_min[i] = magic_min; in xnn_init_qs8_conv_minmax_fp32_wasmsimd_params()
460 params->fp32_wasmsimd.magic_bias_less_output_zero_point[i] = magic_bias_less_zero_point; in xnn_init_qs8_conv_minmax_fp32_wasmsimd_params()
463 params->fp32_wasmsimd.output_max[i] = output_max; in xnn_init_qs8_conv_minmax_fp32_wasmsimd_params()
465 return sizeof(params->fp32_wasmsimd); in xnn_init_qs8_conv_minmax_fp32_wasmsimd_params()
470 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params()
480 params->fp32_scalar_fmagic.kernel_zero_point = (int32_t) kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params()
481 params->fp32_scalar_fmagic.scale = scale; in xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params()
482params->fp32_scalar_fmagic.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params()
483params->fp32_scalar_fmagic.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params()
484 params->fp32_scalar_fmagic.magic_bias = 12582912.0f; in xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params()
485params->fp32_scalar_fmagic.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) out… in xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params()
486 return sizeof(params->fp32_scalar_fmagic); in xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params()
490 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params()
502 params->fp32_scalar_imagic.kernel_zero_point = (int32_t) kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params()
503 params->fp32_scalar_imagic.scale = scale; in xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params()
504 params->fp32_scalar_imagic.magic_bias = 12582912.0f; in xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params()
505params->fp32_scalar_imagic.magic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zer… in xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params()
506params->fp32_scalar_imagic.magic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zer… in xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params()
507params->fp32_scalar_imagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zer… in xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params()
508 return sizeof(params->fp32_scalar_imagic); in xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params()
512 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params()
522 params->fp32_scalar_lrintf.kernel_zero_point = (int32_t) kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params()
523 params->fp32_scalar_lrintf.scale = scale; in xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params()
524params->fp32_scalar_lrintf.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params()
525params->fp32_scalar_lrintf.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params()
526 params->fp32_scalar_lrintf.output_zero_point = (int32_t) output_zero_point; in xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params()
527 return sizeof(params->fp32_scalar_lrintf); in xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params()
532 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_sse2_params()
544 params->fp32_sse2.scale[i] = scale; in xnn_init_qu8_conv_minmax_fp32_sse2_params()
545 params->fp32_sse2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qu8_conv_minmax_fp32_sse2_params()
548 params->fp32_sse2.kernel_zero_point[i] = (int16_t) kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_sse2_params()
549 params->fp32_sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qu8_conv_minmax_fp32_sse2_params()
552 params->fp32_sse2.output_min[i] = output_min; in xnn_init_qu8_conv_minmax_fp32_sse2_params()
554 return sizeof(params->fp32_sse2); in xnn_init_qu8_conv_minmax_fp32_sse2_params()
558 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_avx2_params()
570 params->fp32_avx2.scale[i] = scale; in xnn_init_qu8_conv_minmax_fp32_avx2_params()
571 params->fp32_avx2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qu8_conv_minmax_fp32_avx2_params()
574 params->fp32_avx2.kernel_zero_point[i] = (int16_t) kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_avx2_params()
575 params->fp32_avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qu8_conv_minmax_fp32_avx2_params()
578 params->fp32_avx2.output_min[i] = output_min; in xnn_init_qu8_conv_minmax_fp32_avx2_params()
580 return sizeof(params->fp32_avx2); in xnn_init_qu8_conv_minmax_fp32_avx2_params()
584 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_avx512_params()
596 params->fp32_avx512.scale[i] = scale; in xnn_init_qu8_conv_minmax_fp32_avx512_params()
597 params->fp32_avx512.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qu8_conv_minmax_fp32_avx512_params()
600 params->fp32_avx512.kernel_zero_point[i] = (int16_t) (uint16_t) kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_avx512_params()
601 params->fp32_avx512.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_conv_minmax_fp32_avx512_params()
604 params->fp32_avx512.output_min[i] = output_min; in xnn_init_qu8_conv_minmax_fp32_avx512_params()
606 return sizeof(params->fp32_avx512); in xnn_init_qu8_conv_minmax_fp32_avx512_params()
612 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_armsimd32_params()
623 params->fp32_armsimd32.scale = scale; in xnn_init_qu8_conv_minmax_fp32_armsimd32_params()
624 params->fp32_armsimd32.magic_bias = 12582912.0f; in xnn_init_qu8_conv_minmax_fp32_armsimd32_params()
625params->fp32_armsimd32.minus_kernel_zero_point = (uint32_t) (uint16_t) minus_kernel_zero_point * U… in xnn_init_qu8_conv_minmax_fp32_armsimd32_params()
626params->fp32_armsimd32.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_po… in xnn_init_qu8_conv_minmax_fp32_armsimd32_params()
627 params->fp32_armsimd32.output_min = (uint32_t) output_min * UINT32_C(0x01010101); in xnn_init_qu8_conv_minmax_fp32_armsimd32_params()
628 params->fp32_armsimd32.output_max = (uint32_t) output_max * UINT32_C(0x01010101); in xnn_init_qu8_conv_minmax_fp32_armsimd32_params()
629 return sizeof(params->fp32_armsimd32); in xnn_init_qu8_conv_minmax_fp32_armsimd32_params()
635 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_neon_params()
645 params->fp32_neon.kernel_zero_point[0] = kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_neon_params()
646 params->fp32_neon.kernel_zero_point[1] = kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_neon_params()
647 params->fp32_neon.kernel_zero_point[2] = kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_neon_params()
648 params->fp32_neon.kernel_zero_point[3] = kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_neon_params()
649 params->fp32_neon.scale = scale; in xnn_init_qu8_conv_minmax_fp32_neon_params()
650 params->fp32_neon.magic_bias = 12582912.0f; in xnn_init_qu8_conv_minmax_fp32_neon_params()
651params->fp32_neon.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_… in xnn_init_qu8_conv_minmax_fp32_neon_params()
652 params->fp32_neon.output_min = output_min; in xnn_init_qu8_conv_minmax_fp32_neon_params()
653 params->fp32_neon.output_max = output_max; in xnn_init_qu8_conv_minmax_fp32_neon_params()
654 return sizeof(params->fp32_neon); in xnn_init_qu8_conv_minmax_fp32_neon_params()
658 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
668 params->fp32_neonv8.kernel_zero_point[0] = kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
669 params->fp32_neonv8.kernel_zero_point[1] = kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
670 params->fp32_neonv8.kernel_zero_point[2] = kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
671 params->fp32_neonv8.kernel_zero_point[3] = kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
672 params->fp32_neonv8.scale = scale; in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
673 params->fp32_neonv8.output_zero_point = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
674 params->fp32_neonv8.output_min = output_min; in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
675 params->fp32_neonv8.output_max = output_max; in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
676 return sizeof(params->fp32_neonv8); in xnn_init_qu8_conv_minmax_fp32_neonv8_params()
680 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_rndnu_neon_params()
707 params->rndnu_neon.kernel_zero_point[0] = kernel_zero_point; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
708 params->rndnu_neon.kernel_zero_point[1] = kernel_zero_point; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
709 params->rndnu_neon.kernel_zero_point[2] = kernel_zero_point; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
710 params->rndnu_neon.kernel_zero_point[3] = kernel_zero_point; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
711 params->rndnu_neon.right_pre_shift = -pre_shift; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
712 params->rndnu_neon.multiplier = multiplier; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
713 params->rndnu_neon.right_post_shift = -post_shift; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
714 params->rndnu_neon.output_zero_point = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
715 params->rndnu_neon.output_min = output_min; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
716 params->rndnu_neon.output_max = output_max; in xnn_init_qu8_conv_minmax_rndnu_neon_params()
717 return sizeof(params->rndnu_neon); in xnn_init_qu8_conv_minmax_rndnu_neon_params()
723 union xnn_qu8_conv_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_conv_minmax_fp32_wasmsimd_params()
737 params->fp32_wasmsimd.kernel_zero_point[i] = (int16_t) (uint16_t) kernel_zero_point; in xnn_init_qu8_conv_minmax_fp32_wasmsimd_params()
740 params->fp32_wasmsimd.scale[i] = scale; in xnn_init_qu8_conv_minmax_fp32_wasmsimd_params()
741 params->fp32_wasmsimd.magic_bias[i] = 12582912.0f; in xnn_init_qu8_conv_minmax_fp32_wasmsimd_params()
742 params->fp32_wasmsimd.magic_min[i] = magic_min; in xnn_init_qu8_conv_minmax_fp32_wasmsimd_params()
743 params->fp32_wasmsimd.magic_bias_less_output_zero_point[i] = magic_bias_less_zero_point; in xnn_init_qu8_conv_minmax_fp32_wasmsimd_params()
746 params->fp32_wasmsimd.output_max[i] = output_max; in xnn_init_qu8_conv_minmax_fp32_wasmsimd_params()
748 return sizeof(params->fp32_wasmsimd); in xnn_init_qu8_conv_minmax_fp32_wasmsimd_params()
769 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
779 params->fp32_scalar_fmagic.init_bias = init_bias; in xnn_init_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
780 params->fp32_scalar_fmagic.scale = scale; in xnn_init_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
781params->fp32_scalar_fmagic.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
782params->fp32_scalar_fmagic.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
783 params->fp32_scalar_fmagic.magic_bias = 12582912.0f; in xnn_init_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
784params->fp32_scalar_fmagic.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) out… in xnn_init_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
785 return sizeof(params->fp32_scalar_fmagic); in xnn_init_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
789 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
796 params->fp32_scalar_fmagic.init_bias = init_bias; in xnn_update_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
797 params->fp32_scalar_fmagic.scale = scale; in xnn_update_qs8_avgpool_minmax_fp32_scalar_fmagic_params()
801 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params()
813 params->fp32_scalar_imagic.init_bias = init_bias; in xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params()
814 params->fp32_scalar_imagic.scale = scale; in xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params()
815 params->fp32_scalar_imagic.magic_bias = 12582912.0f; in xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params()
816params->fp32_scalar_imagic.magic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zer… in xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params()
817params->fp32_scalar_imagic.magic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zer… in xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params()
818params->fp32_scalar_imagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zer… in xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params()
819 return sizeof(params->fp32_scalar_imagic); in xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params()
823 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_fp32_scalar_imagic_params()
830 params->fp32_scalar_imagic.init_bias = init_bias; in xnn_update_qs8_avgpool_minmax_fp32_scalar_imagic_params()
831 params->fp32_scalar_imagic.scale = scale; in xnn_update_qs8_avgpool_minmax_fp32_scalar_imagic_params()
835 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
845 params->fp32_scalar_lrintf.init_bias = init_bias; in xnn_init_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
846 params->fp32_scalar_lrintf.scale = scale; in xnn_init_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
847params->fp32_scalar_lrintf.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
848params->fp32_scalar_lrintf.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
849 params->fp32_scalar_lrintf.output_zero_point = (int32_t) output_zero_point; in xnn_init_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
850 return sizeof(params->fp32_scalar_lrintf); in xnn_init_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
854 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
861 params->fp32_scalar_lrintf.init_bias = init_bias; in xnn_update_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
862 params->fp32_scalar_lrintf.scale = scale; in xnn_update_qs8_avgpool_minmax_fp32_scalar_lrintf_params()
867 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_fp32_sse2_params()
879 params->fp32_sse2.init_bias[i] = init_bias; in xnn_init_qs8_avgpool_minmax_fp32_sse2_params()
880 params->fp32_sse2.scale[i] = scale; in xnn_init_qs8_avgpool_minmax_fp32_sse2_params()
881 params->fp32_sse2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qs8_avgpool_minmax_fp32_sse2_params()
884 params->fp32_sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_avgpool_minmax_fp32_sse2_params()
885 params->fp32_sse2.output_min[i] = (int16_t) output_min; in xnn_init_qs8_avgpool_minmax_fp32_sse2_params()
887 return sizeof(params->fp32_sse2); in xnn_init_qs8_avgpool_minmax_fp32_sse2_params()
891 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_fp32_sse2_params()
899 params->fp32_sse2.init_bias[i] = init_bias; in xnn_update_qs8_avgpool_minmax_fp32_sse2_params()
900 params->fp32_sse2.scale[i] = scale; in xnn_update_qs8_avgpool_minmax_fp32_sse2_params()
905 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_fp32_sse4_params()
917 params->fp32_sse4.init_bias[i] = init_bias; in xnn_init_qs8_avgpool_minmax_fp32_sse4_params()
918 params->fp32_sse4.scale[i] = scale; in xnn_init_qs8_avgpool_minmax_fp32_sse4_params()
919 params->fp32_sse4.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qs8_avgpool_minmax_fp32_sse4_params()
922 params->fp32_sse4.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_avgpool_minmax_fp32_sse4_params()
925 params->fp32_sse4.output_min[i] = output_min; in xnn_init_qs8_avgpool_minmax_fp32_sse4_params()
927 return sizeof(params->fp32_sse4); in xnn_init_qs8_avgpool_minmax_fp32_sse4_params()
931 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_fp32_sse4_params()
939 params->fp32_sse4.init_bias[i] = init_bias; in xnn_update_qs8_avgpool_minmax_fp32_sse4_params()
940 params->fp32_sse4.scale[i] = scale; in xnn_update_qs8_avgpool_minmax_fp32_sse4_params()
947 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_fp32_neon_params()
957 params->fp32_neon.init_bias = init_bias; in xnn_init_qs8_avgpool_minmax_fp32_neon_params()
958 params->fp32_neon.scale = scale; in xnn_init_qs8_avgpool_minmax_fp32_neon_params()
959 params->fp32_neon.magic_bias = 12582912.0f; in xnn_init_qs8_avgpool_minmax_fp32_neon_params()
960params->fp32_neon.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_… in xnn_init_qs8_avgpool_minmax_fp32_neon_params()
961 params->fp32_neon.output_min = output_min; in xnn_init_qs8_avgpool_minmax_fp32_neon_params()
962 params->fp32_neon.output_max = output_max; in xnn_init_qs8_avgpool_minmax_fp32_neon_params()
963 return sizeof(params->fp32_neon); in xnn_init_qs8_avgpool_minmax_fp32_neon_params()
967 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_fp32_neon_params()
974 params->fp32_neon.init_bias = init_bias; in xnn_update_qs8_avgpool_minmax_fp32_neon_params()
975 params->fp32_neon.scale = scale; in xnn_update_qs8_avgpool_minmax_fp32_neon_params()
979 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_fp32_neonv8_params()
989 params->fp32_neonv8.init_bias = init_bias; in xnn_init_qs8_avgpool_minmax_fp32_neonv8_params()
990 params->fp32_neonv8.scale = scale; in xnn_init_qs8_avgpool_minmax_fp32_neonv8_params()
991 params->fp32_neonv8.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_avgpool_minmax_fp32_neonv8_params()
992 params->fp32_neonv8.output_min = output_min; in xnn_init_qs8_avgpool_minmax_fp32_neonv8_params()
993 params->fp32_neonv8.output_max = output_max; in xnn_init_qs8_avgpool_minmax_fp32_neonv8_params()
994 return sizeof(params->fp32_neonv8); in xnn_init_qs8_avgpool_minmax_fp32_neonv8_params()
998 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_fp32_neonv8_params()
1005 params->fp32_neonv8.init_bias = init_bias; in xnn_update_qs8_avgpool_minmax_fp32_neonv8_params()
1006 params->fp32_neonv8.scale = scale; in xnn_update_qs8_avgpool_minmax_fp32_neonv8_params()
1010 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1037 params->rndnu_neon.init_bias = init_bias; in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1038 params->rndnu_neon.left_pre_shift = -pre_shift; in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1039 params->rndnu_neon.multiplier = multiplier; in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1040 params->rndnu_neon.left_post_shift = -post_shift; in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1041 params->rndnu_neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1042 params->rndnu_neon.output_min = output_min; in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1043 params->rndnu_neon.output_max = output_max; in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1044 return sizeof(params->rndnu_neon); in xnn_init_qs8_avgpool_minmax_rndnu_neon_params()
1048 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_rndnu_neon_params()
1072 params->rndnu_neon.init_bias = init_bias; in xnn_update_qs8_avgpool_minmax_rndnu_neon_params()
1073 params->rndnu_neon.left_pre_shift = -pre_shift; in xnn_update_qs8_avgpool_minmax_rndnu_neon_params()
1074 params->rndnu_neon.multiplier = multiplier; in xnn_update_qs8_avgpool_minmax_rndnu_neon_params()
1075 params->rndnu_neon.left_post_shift = -post_shift; in xnn_update_qs8_avgpool_minmax_rndnu_neon_params()
1081 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params()
1095 params->fp32_wasmsimd.init_bias[i] = init_bias; in xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params()
1096 params->fp32_wasmsimd.scale[i] = scale; in xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params()
1097 params->fp32_wasmsimd.magic_bias[i] = 12582912.0f; in xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params()
1098 params->fp32_wasmsimd.magic_min[i] = magic_min; in xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params()
1099 params->fp32_wasmsimd.magic_bias_less_output_zero_point[i] = magic_bias_less_zero_point; in xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params()
1102 params->fp32_wasmsimd.output_max[i] = output_max; in xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params()
1104 return sizeof(params->fp32_wasmsimd); in xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params()
1108 union xnn_qs8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qs8_avgpool_minmax_fp32_wasmsimd_params()
1116 params->fp32_wasmsimd.init_bias[i] = init_bias; in xnn_update_qs8_avgpool_minmax_fp32_wasmsimd_params()
1117 params->fp32_wasmsimd.scale[i] = scale; in xnn_update_qs8_avgpool_minmax_fp32_wasmsimd_params()
1123 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1133 params->fp32_scalar_fmagic.init_bias = init_bias; in xnn_init_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1134 params->fp32_scalar_fmagic.scale = scale; in xnn_init_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1135params->fp32_scalar_fmagic.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1136params->fp32_scalar_fmagic.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1137 params->fp32_scalar_fmagic.magic_bias = 12582912.0f; in xnn_init_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1138params->fp32_scalar_fmagic.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) out… in xnn_init_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1139 return sizeof(params->fp32_scalar_fmagic); in xnn_init_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1143 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1150 params->fp32_scalar_fmagic.init_bias = init_bias; in xnn_update_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1151 params->fp32_scalar_fmagic.scale = scale; in xnn_update_qu8_avgpool_minmax_fp32_scalar_fmagic_params()
1155 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1167 params->fp32_scalar_imagic.init_bias = init_bias; in xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1168 params->fp32_scalar_imagic.scale = scale; in xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1169 params->fp32_scalar_imagic.magic_bias = 12582912.0f; in xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1170params->fp32_scalar_imagic.magic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zer… in xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1171params->fp32_scalar_imagic.magic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zer… in xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1172params->fp32_scalar_imagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zer… in xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1173 return sizeof(params->fp32_scalar_imagic); in xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1177 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1184 params->fp32_scalar_imagic.init_bias = init_bias; in xnn_update_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1185 params->fp32_scalar_imagic.scale = scale; in xnn_update_qu8_avgpool_minmax_fp32_scalar_imagic_params()
1189 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1199 params->fp32_scalar_lrintf.init_bias = init_bias; in xnn_init_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1200 params->fp32_scalar_lrintf.scale = scale; in xnn_init_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1201params->fp32_scalar_lrintf.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) … in xnn_init_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1202params->fp32_scalar_lrintf.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) … in xnn_init_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1203 params->fp32_scalar_lrintf.output_zero_point = (int32_t) output_zero_point; in xnn_init_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1204 return sizeof(params->fp32_scalar_lrintf); in xnn_init_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1208 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1215 params->fp32_scalar_lrintf.init_bias = init_bias; in xnn_update_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1216 params->fp32_scalar_lrintf.scale = scale; in xnn_update_qu8_avgpool_minmax_fp32_scalar_lrintf_params()
1221 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_fp32_sse2_params()
1233 params->fp32_sse2.init_bias[i] = init_bias; in xnn_init_qu8_avgpool_minmax_fp32_sse2_params()
1234 params->fp32_sse2.scale[i] = scale; in xnn_init_qu8_avgpool_minmax_fp32_sse2_params()
1235 params->fp32_sse2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qu8_avgpool_minmax_fp32_sse2_params()
1238 params->fp32_sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qu8_avgpool_minmax_fp32_sse2_params()
1241 params->fp32_sse2.output_min[i] = output_min; in xnn_init_qu8_avgpool_minmax_fp32_sse2_params()
1243 return sizeof(params->fp32_sse2); in xnn_init_qu8_avgpool_minmax_fp32_sse2_params()
1247 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_fp32_sse2_params()
1255 params->fp32_sse2.init_bias[i] = init_bias; in xnn_update_qu8_avgpool_minmax_fp32_sse2_params()
1256 params->fp32_sse2.scale[i] = scale; in xnn_update_qu8_avgpool_minmax_fp32_sse2_params()
1261 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_fp32_sse4_params()
1273 params->fp32_sse4.init_bias[i] = init_bias; in xnn_init_qu8_avgpool_minmax_fp32_sse4_params()
1274 params->fp32_sse4.scale[i] = scale; in xnn_init_qu8_avgpool_minmax_fp32_sse4_params()
1275 params->fp32_sse4.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_qu8_avgpool_minmax_fp32_sse4_params()
1278 params->fp32_sse4.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qu8_avgpool_minmax_fp32_sse4_params()
1281 params->fp32_sse4.output_min[i] = output_min; in xnn_init_qu8_avgpool_minmax_fp32_sse4_params()
1283 return sizeof(params->fp32_sse4); in xnn_init_qu8_avgpool_minmax_fp32_sse4_params()
1287 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_fp32_sse4_params()
1295 params->fp32_sse4.init_bias[i] = init_bias; in xnn_update_qu8_avgpool_minmax_fp32_sse4_params()
1296 params->fp32_sse4.scale[i] = scale; in xnn_update_qu8_avgpool_minmax_fp32_sse4_params()
1303 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_fp32_neon_params()
1313 params->fp32_neon.init_bias = init_bias; in xnn_init_qu8_avgpool_minmax_fp32_neon_params()
1314 params->fp32_neon.scale = scale; in xnn_init_qu8_avgpool_minmax_fp32_neon_params()
1315 params->fp32_neon.magic_bias = 12582912.0f; in xnn_init_qu8_avgpool_minmax_fp32_neon_params()
1316params->fp32_neon.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_… in xnn_init_qu8_avgpool_minmax_fp32_neon_params()
1317 params->fp32_neon.output_min = output_min; in xnn_init_qu8_avgpool_minmax_fp32_neon_params()
1318 params->fp32_neon.output_max = output_max; in xnn_init_qu8_avgpool_minmax_fp32_neon_params()
1319 return sizeof(params->fp32_neon); in xnn_init_qu8_avgpool_minmax_fp32_neon_params()
1323 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_fp32_neon_params()
1330 params->fp32_neon.init_bias = init_bias; in xnn_update_qu8_avgpool_minmax_fp32_neon_params()
1331 params->fp32_neon.scale = scale; in xnn_update_qu8_avgpool_minmax_fp32_neon_params()
1335 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_fp32_neonv8_params()
1345 params->fp32_neonv8.init_bias = init_bias; in xnn_init_qu8_avgpool_minmax_fp32_neonv8_params()
1346 params->fp32_neonv8.scale = scale; in xnn_init_qu8_avgpool_minmax_fp32_neonv8_params()
1347 params->fp32_neonv8.output_zero_point = (int16_t) output_zero_point; in xnn_init_qu8_avgpool_minmax_fp32_neonv8_params()
1348 params->fp32_neonv8.output_min = output_min; in xnn_init_qu8_avgpool_minmax_fp32_neonv8_params()
1349 params->fp32_neonv8.output_max = output_max; in xnn_init_qu8_avgpool_minmax_fp32_neonv8_params()
1350 return sizeof(params->fp32_neonv8); in xnn_init_qu8_avgpool_minmax_fp32_neonv8_params()
1354 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_fp32_neonv8_params()
1361 params->fp32_neonv8.init_bias = init_bias; in xnn_update_qu8_avgpool_minmax_fp32_neonv8_params()
1362 params->fp32_neonv8.scale = scale; in xnn_update_qu8_avgpool_minmax_fp32_neonv8_params()
1366 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1393 params->rndnu_neon.init_bias = init_bias; in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1394 params->rndnu_neon.left_pre_shift = -pre_shift; in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1395 params->rndnu_neon.multiplier = multiplier; in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1396 params->rndnu_neon.left_post_shift = -post_shift; in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1397 params->rndnu_neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1398 params->rndnu_neon.output_min = output_min; in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1399 params->rndnu_neon.output_max = output_max; in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1400 return sizeof(params->rndnu_neon); in xnn_init_qu8_avgpool_minmax_rndnu_neon_params()
1404 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_rndnu_neon_params()
1428 params->rndnu_neon.init_bias = init_bias; in xnn_update_qu8_avgpool_minmax_rndnu_neon_params()
1429 params->rndnu_neon.left_pre_shift = -pre_shift; in xnn_update_qu8_avgpool_minmax_rndnu_neon_params()
1430 params->rndnu_neon.multiplier = multiplier; in xnn_update_qu8_avgpool_minmax_rndnu_neon_params()
1431 params->rndnu_neon.left_post_shift = -post_shift; in xnn_update_qu8_avgpool_minmax_rndnu_neon_params()
1437 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params()
1451 params->fp32_wasmsimd.init_bias[i] = init_bias; in xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params()
1452 params->fp32_wasmsimd.scale[i] = scale; in xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params()
1453 params->fp32_wasmsimd.magic_bias[i] = 12582912.0f; in xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params()
1454 params->fp32_wasmsimd.magic_min[i] = magic_min; in xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params()
1455 params->fp32_wasmsimd.magic_bias_less_output_zero_point[i] = magic_bias_less_zero_point; in xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params()
1458 params->fp32_wasmsimd.output_max[i] = output_max; in xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params()
1460 return sizeof(params->fp32_wasmsimd); in xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params()
1464 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_qu8_avgpool_minmax_fp32_wasmsimd_params()
1472 params->fp32_wasmsimd.init_bias[i] = init_bias; in xnn_update_qu8_avgpool_minmax_fp32_wasmsimd_params()
1473 params->fp32_wasmsimd.scale[i] = scale; in xnn_update_qu8_avgpool_minmax_fp32_wasmsimd_params()
1479 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_scalar_params()
1503 params->scalar.bias = bias; in xnn_init_qu8_avgpool_minmax_scalar_params()
1504 params->scalar.rounding = rounding; in xnn_init_qu8_avgpool_minmax_scalar_params()
1505 params->scalar.multiplier = multiplier; in xnn_init_qu8_avgpool_minmax_scalar_params()
1506 params->scalar.right_shift = right_shift; in xnn_init_qu8_avgpool_minmax_scalar_params()
1507 params->scalar.output_min_less_zero_point = in xnn_init_qu8_avgpool_minmax_scalar_params()
1509 params->scalar.output_max_less_zero_point = in xnn_init_qu8_avgpool_minmax_scalar_params()
1511 params->scalar.output_zero_point = (int32_t) (uint32_t) output_zero_point; in xnn_init_qu8_avgpool_minmax_scalar_params()
1512 return sizeof(params->scalar); in xnn_init_qu8_avgpool_minmax_scalar_params()
1517 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_neon_params()
1539 params->neon.bias = bias; in xnn_init_qu8_avgpool_minmax_neon_params()
1540 params->neon.multiplier = multiplier; in xnn_init_qu8_avgpool_minmax_neon_params()
1541 params->neon.left_shift = (int64_t) -shift; in xnn_init_qu8_avgpool_minmax_neon_params()
1542 params->neon.output_zero_point = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_avgpool_minmax_neon_params()
1543 params->neon.output_min = output_min; in xnn_init_qu8_avgpool_minmax_neon_params()
1544 params->neon.output_max = output_max; in xnn_init_qu8_avgpool_minmax_neon_params()
1545 return sizeof(params->neon); in xnn_init_qu8_avgpool_minmax_neon_params()
1551 union xnn_qu8_avgpool_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_avgpool_minmax_sse2_params()
1575 params->sse2.bias[0] = bias; in xnn_init_qu8_avgpool_minmax_sse2_params()
1576 params->sse2.bias[1] = bias; in xnn_init_qu8_avgpool_minmax_sse2_params()
1577 params->sse2.bias[2] = bias; in xnn_init_qu8_avgpool_minmax_sse2_params()
1578 params->sse2.bias[3] = bias; in xnn_init_qu8_avgpool_minmax_sse2_params()
1579 params->sse2.multiplier[0] = (uint32_t) multiplier; in xnn_init_qu8_avgpool_minmax_sse2_params()
1580 params->sse2.multiplier[1] = (uint32_t) multiplier; in xnn_init_qu8_avgpool_minmax_sse2_params()
1581 params->sse2.multiplier[2] = (uint32_t) multiplier; in xnn_init_qu8_avgpool_minmax_sse2_params()
1582 params->sse2.multiplier[3] = (uint32_t) multiplier; in xnn_init_qu8_avgpool_minmax_sse2_params()
1583 params->sse2.rounding[0] = rounding; in xnn_init_qu8_avgpool_minmax_sse2_params()
1584 params->sse2.rounding[1] = rounding; in xnn_init_qu8_avgpool_minmax_sse2_params()
1585 params->sse2.right_shift[0] = (uint64_t) right_shift; in xnn_init_qu8_avgpool_minmax_sse2_params()
1586 params->sse2.right_shift[1] = (uint64_t) right_shift; in xnn_init_qu8_avgpool_minmax_sse2_params()
1588 params->sse2.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_avgpool_minmax_sse2_params()
1591 params->sse2.output_min[i] = output_min; in xnn_init_qu8_avgpool_minmax_sse2_params()
1592 params->sse2.output_max[i] = output_max; in xnn_init_qu8_avgpool_minmax_sse2_params()
1594 return sizeof(params->sse2); in xnn_init_qu8_avgpool_minmax_sse2_params()
1599 union xnn_qu8_avgpool_minmax_params* params, in xnn_update_qu8_avgpool_minmax_scalar_params() argument
1619 params->scalar.bias = bias; in xnn_update_qu8_avgpool_minmax_scalar_params()
1620 params->scalar.multiplier = multiplier; in xnn_update_qu8_avgpool_minmax_scalar_params()
1621 params->scalar.rounding = rounding; in xnn_update_qu8_avgpool_minmax_scalar_params()
1622 params->scalar.right_shift = (uint32_t) shift; in xnn_update_qu8_avgpool_minmax_scalar_params()
1627 union xnn_qu8_avgpool_minmax_params* params, in xnn_update_qu8_avgpool_minmax_neon_params() argument
1646 params->neon.bias = bias; in xnn_update_qu8_avgpool_minmax_neon_params()
1647 params->neon.multiplier = multiplier; in xnn_update_qu8_avgpool_minmax_neon_params()
1648 params->neon.left_shift = (int64_t) -shift; in xnn_update_qu8_avgpool_minmax_neon_params()
1654 union xnn_qu8_avgpool_minmax_params* params, in xnn_update_qu8_avgpool_minmax_sse2_params() argument
1674 params->sse2.bias[0] = bias; in xnn_update_qu8_avgpool_minmax_sse2_params()
1675 params->sse2.bias[1] = bias; in xnn_update_qu8_avgpool_minmax_sse2_params()
1676 params->sse2.bias[2] = bias; in xnn_update_qu8_avgpool_minmax_sse2_params()
1677 params->sse2.bias[3] = bias; in xnn_update_qu8_avgpool_minmax_sse2_params()
1678 params->sse2.multiplier[0] = (uint32_t) multiplier; in xnn_update_qu8_avgpool_minmax_sse2_params()
1679 params->sse2.multiplier[1] = (uint32_t) multiplier; in xnn_update_qu8_avgpool_minmax_sse2_params()
1680 params->sse2.multiplier[2] = (uint32_t) multiplier; in xnn_update_qu8_avgpool_minmax_sse2_params()
1681 params->sse2.multiplier[3] = (uint32_t) multiplier; in xnn_update_qu8_avgpool_minmax_sse2_params()
1682 params->sse2.rounding[0] = rounding; in xnn_update_qu8_avgpool_minmax_sse2_params()
1683 params->sse2.rounding[1] = rounding; in xnn_update_qu8_avgpool_minmax_sse2_params()
1684 params->sse2.right_shift[0] = (uint64_t) (uint32_t) shift; in xnn_update_qu8_avgpool_minmax_sse2_params()
1685 params->sse2.right_shift[1] = (uint64_t) (uint32_t) shift; in xnn_update_qu8_avgpool_minmax_sse2_params()
1690 union xnn_f32_scaleminmax_params* params, in xnn_update_f32_scaleminmax_scalar_params() argument
1693 params->scalar.scale = scale; in xnn_update_f32_scaleminmax_scalar_params()
1698 union xnn_f32_scaleminmax_params* params, in xnn_update_f32_scaleminmax_sse_params() argument
1702 params->sse.scale[i] = scale; in xnn_update_f32_scaleminmax_sse_params()
1709 union xnn_f16_scaleminmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_scaleminmax_neon_params()
1714 params->neon.scale = scale; in xnn_init_f16_scaleminmax_neon_params()
1715 params->neon.min = min; in xnn_init_f16_scaleminmax_neon_params()
1716 params->neon.max = max; in xnn_init_f16_scaleminmax_neon_params()
1717 return sizeof(params->neon); in xnn_init_f16_scaleminmax_neon_params()
1723 union xnn_f16_scaleminmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_scaleminmax_avx_params()
1732 params->avx.scale[i] = scale_f32; in xnn_init_f16_scaleminmax_avx_params()
1733 params->avx.min[i] = min_f32; in xnn_init_f16_scaleminmax_avx_params()
1734 params->avx.max[i] = max_f32; in xnn_init_f16_scaleminmax_avx_params()
1736 return sizeof(params->avx); in xnn_init_f16_scaleminmax_avx_params()
1742 union xnn_f16_scaleminmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_f16_scaleminmax_neon_params()
1745 params->neon.scale = scale; in xnn_update_f16_scaleminmax_neon_params()
1751 union xnn_f16_scaleminmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_update_f16_scaleminmax_avx_params()
1756 params->avx.scale[i] = scale_f32; in xnn_update_f16_scaleminmax_avx_params()
1762 union xnn_f32_scaleminmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_scaleminmax_scalar_params()
1767 params->scalar.scale = scale; in xnn_init_f32_scaleminmax_scalar_params()
1768 params->scalar.min = min; in xnn_init_f32_scaleminmax_scalar_params()
1769 params->scalar.max = max; in xnn_init_f32_scaleminmax_scalar_params()
1770 return sizeof(params->scalar); in xnn_init_f32_scaleminmax_scalar_params()
1775 union xnn_f32_scaleminmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_scaleminmax_sse_params()
1781 params->sse.scale[i] = scale; in xnn_init_f32_scaleminmax_sse_params()
1782 params->sse.min[i] = min; in xnn_init_f32_scaleminmax_sse_params()
1783 params->sse.max[i] = max; in xnn_init_f32_scaleminmax_sse_params()
1785 return sizeof(params->sse); in xnn_init_f32_scaleminmax_sse_params()
1790 union xnn_f32_gavgpool_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_gavgpool_params()
1798 params->sse.multiplier[i] = multiplier; in xnn_init_f32_gavgpool_params()
1799 params->sse.output_min[i] = output_min; in xnn_init_f32_gavgpool_params()
1800 params->sse.output_max[i] = output_max; in xnn_init_f32_gavgpool_params()
1804 params->sse.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_gavgpool_params()
1805 params->sse.mask[1] = -(uint32_t) (w >= 1); in xnn_init_f32_gavgpool_params()
1806 params->sse.mask[2] = -(uint32_t) (w >= 2); in xnn_init_f32_gavgpool_params()
1807 params->sse.mask[3] = -(uint32_t) (w >= 3); in xnn_init_f32_gavgpool_params()
1808 return sizeof(params->sse); in xnn_init_f32_gavgpool_params()
1810 params->neon.multiplier = multiplier; in xnn_init_f32_gavgpool_params()
1811 params->neon.output_min = output_min; in xnn_init_f32_gavgpool_params()
1812 params->neon.output_max = output_max; in xnn_init_f32_gavgpool_params()
1815 params->neon.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_gavgpool_params()
1816 params->neon.mask[1] = -(uint32_t) (w >= 1); in xnn_init_f32_gavgpool_params()
1817 params->neon.mask[2] = -(uint32_t) (w >= 2); in xnn_init_f32_gavgpool_params()
1818 params->neon.mask[3] = -(uint32_t) (w >= 3); in xnn_init_f32_gavgpool_params()
1819 return sizeof(params->neon); in xnn_init_f32_gavgpool_params()
1821 params->scalar.multiplier = multiplier; in xnn_init_f32_gavgpool_params()
1822 params->scalar.output_min = output_min; in xnn_init_f32_gavgpool_params()
1823 params->scalar.output_max = output_max; in xnn_init_f32_gavgpool_params()
1826 params->scalar.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_gavgpool_params()
1827 params->scalar.mask[1] = -(int32_t) (w >= 1); in xnn_init_f32_gavgpool_params()
1828 params->scalar.mask[2] = -(int32_t) (w >= 2); in xnn_init_f32_gavgpool_params()
1829 params->scalar.mask[3] = -(int32_t) (w >= 3); in xnn_init_f32_gavgpool_params()
1830 return sizeof(params->scalar); in xnn_init_f32_gavgpool_params()
1835 union xnn_f16_gavgpool_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1842 params->neonfp16arith.multiplier = multiplier; in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1843 params->neonfp16arith.output_min = output_min; in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1844 params->neonfp16arith.output_max = output_max; in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1847 params->neonfp16arith.mask[0] = UINT16_C(0xFFFF); in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1848 params->neonfp16arith.mask[1] = -(uint16_t) (w >= 1); in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1849 params->neonfp16arith.mask[2] = -(uint16_t) (w >= 2); in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1850 params->neonfp16arith.mask[3] = -(uint16_t) (w >= 3); in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1851 return sizeof(params->neonfp16arith); in xnn_init_f16_gavgpool_neonfp16arith_x4_params()
1858 union xnn_f16_gavgpool_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1865 params->neonfp16arith.multiplier = multiplier; in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1866 params->neonfp16arith.output_min = output_min; in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1867 params->neonfp16arith.output_max = output_max; in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1870 params->neonfp16arith.mask[0] = UINT16_C(0xFFFF); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1871 params->neonfp16arith.mask[1] = -(uint16_t) (w >= 1); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1872 params->neonfp16arith.mask[2] = -(uint16_t) (w >= 2); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1873 params->neonfp16arith.mask[3] = -(uint16_t) (w >= 3); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1874 params->neonfp16arith.mask[4] = -(uint16_t) (w >= 4); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1875 params->neonfp16arith.mask[5] = -(uint16_t) (w >= 5); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1876 params->neonfp16arith.mask[6] = -(uint16_t) (w >= 6); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1877 params->neonfp16arith.mask[7] = -(uint16_t) (w >= 7); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1878 return sizeof(params->neonfp16arith); in xnn_init_f16_gavgpool_neonfp16arith_x8_params()
1885 union xnn_f32_gavgpool_params* params, in xnn_update_f32_gavgpool_params() argument
1891 params->sse.multiplier[i] = multiplier; in xnn_update_f32_gavgpool_params()
1895 params->sse.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_gavgpool_params()
1896 params->sse.mask[1] = -(uint32_t) (w >= 1); in xnn_update_f32_gavgpool_params()
1897 params->sse.mask[2] = -(uint32_t) (w >= 2); in xnn_update_f32_gavgpool_params()
1898 params->sse.mask[3] = -(uint32_t) (w >= 3); in xnn_update_f32_gavgpool_params()
1900 params->neon.multiplier = multiplier; in xnn_update_f32_gavgpool_params()
1903 params->neon.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_gavgpool_params()
1904 params->neon.mask[1] = -(uint32_t) (w >= 1); in xnn_update_f32_gavgpool_params()
1905 params->neon.mask[2] = -(uint32_t) (w >= 2); in xnn_update_f32_gavgpool_params()
1906 params->neon.mask[3] = -(uint32_t) (w >= 3); in xnn_update_f32_gavgpool_params()
1908 params->scalar.multiplier = multiplier; in xnn_update_f32_gavgpool_params()
1911 params->scalar.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_gavgpool_params()
1912 params->scalar.mask[1] = -(int32_t) (w >= 1); in xnn_update_f32_gavgpool_params()
1913 params->scalar.mask[2] = -(int32_t) (w >= 2); in xnn_update_f32_gavgpool_params()
1914 params->scalar.mask[3] = -(int32_t) (w >= 3); in xnn_update_f32_gavgpool_params()
1919 union xnn_f32_gavgpool_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_scalar_f32_gavgpool_params()
1925 params->scalar.multiplier = multiplier; in xnn_init_scalar_f32_gavgpool_params()
1926 params->scalar.output_min = output_min; in xnn_init_scalar_f32_gavgpool_params()
1927 params->scalar.output_max = output_max; in xnn_init_scalar_f32_gavgpool_params()
1930 params->scalar.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_init_scalar_f32_gavgpool_params()
1931 params->scalar.mask[1] = -(int32_t) (w >= 1); in xnn_init_scalar_f32_gavgpool_params()
1932 params->scalar.mask[2] = -(int32_t) (w >= 2); in xnn_init_scalar_f32_gavgpool_params()
1933 params->scalar.mask[3] = -(int32_t) (w >= 3); in xnn_init_scalar_f32_gavgpool_params()
1934 return sizeof(params->scalar); in xnn_init_scalar_f32_gavgpool_params()
1938 union xnn_bf16_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_bf16_minmax_scalar_params()
1942 params->scalar.min = uint32_as_float((uint32_t) output_min << 16); in xnn_init_bf16_minmax_scalar_params()
1943 params->scalar.max = uint32_as_float((uint32_t) output_max << 16); in xnn_init_bf16_minmax_scalar_params()
1944 return sizeof(params->scalar); in xnn_init_bf16_minmax_scalar_params()
1949 union xnn_f16_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_minmax_neon_params()
1953 params->neon.min = min; in xnn_init_f16_minmax_neon_params()
1954 params->neon.max = max; in xnn_init_f16_minmax_neon_params()
1955 return sizeof(params->neon); in xnn_init_f16_minmax_neon_params()
1961 union xnn_f16_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_minmax_avx_params()
1968 params->avx.min[i] = min_f32; in xnn_init_f16_minmax_avx_params()
1969 params->avx.max[i] = max_f32; in xnn_init_f16_minmax_avx_params()
1971 return sizeof(params->avx); in xnn_init_f16_minmax_avx_params()
1977 union xnn_f32_default_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_default_avx_params()
1980 params->avx.mask_table[i] = -1; in xnn_init_f32_default_avx_params()
1983 params->avx.mask_table[i] = 0; in xnn_init_f32_default_avx_params()
1985 return sizeof(params->avx); in xnn_init_f32_default_avx_params()
1990 union xnn_f32_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_minmax_params()
1996 params->sse.min[i] = output_min; in xnn_init_f32_minmax_params()
1997 params->sse.max[i] = output_max; in xnn_init_f32_minmax_params()
1999 return sizeof(params->sse); in xnn_init_f32_minmax_params()
2001 params->wasmsimd.min[0] = output_min; in xnn_init_f32_minmax_params()
2002 params->wasmsimd.min[1] = output_min; in xnn_init_f32_minmax_params()
2003 params->wasmsimd.max[0] = output_max; in xnn_init_f32_minmax_params()
2004 params->wasmsimd.max[1] = output_max; in xnn_init_f32_minmax_params()
2005 return sizeof(params->wasmsimd); in xnn_init_f32_minmax_params()
2007 params->scalar.min = output_min; in xnn_init_f32_minmax_params()
2008 params->scalar.max = output_max; in xnn_init_f32_minmax_params()
2009 return sizeof(params->scalar); in xnn_init_f32_minmax_params()
2015 union xnn_f32_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_minmax_sse_params()
2020 params->sse.min[i] = output_min; in xnn_init_f32_minmax_sse_params()
2021 params->sse.max[i] = output_max; in xnn_init_f32_minmax_sse_params()
2023 return sizeof(params->sse); in xnn_init_f32_minmax_sse_params()
2027 union xnn_f32_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_minmax_avx_params()
2032 params->avx.min[i] = output_min; in xnn_init_f32_minmax_avx_params()
2033 params->avx.max[i] = output_max; in xnn_init_f32_minmax_avx_params()
2036 params->avx.mask_table[i] = -1; in xnn_init_f32_minmax_avx_params()
2039 params->avx.mask_table[i] = 0; in xnn_init_f32_minmax_avx_params()
2041 return sizeof(params->avx); in xnn_init_f32_minmax_avx_params()
2047 union xnn_f32_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_minmax_wasmsimd_params()
2051 params->wasmsimd.min[0] = output_min; in xnn_init_f32_minmax_wasmsimd_params()
2052 params->wasmsimd.min[1] = output_min; in xnn_init_f32_minmax_wasmsimd_params()
2053 params->wasmsimd.max[0] = output_max; in xnn_init_f32_minmax_wasmsimd_params()
2054 params->wasmsimd.max[1] = output_max; in xnn_init_f32_minmax_wasmsimd_params()
2055 return sizeof(params->wasmsimd); in xnn_init_f32_minmax_wasmsimd_params()
2060 union xnn_f32_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_minmax_scalar_params()
2064 params->scalar.min = output_min; in xnn_init_f32_minmax_scalar_params()
2065 params->scalar.max = output_max; in xnn_init_f32_minmax_scalar_params()
2066 return sizeof(params->scalar); in xnn_init_f32_minmax_scalar_params()
2071 union xnn_f16_hswish_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_hswish_neon_params()
2073 params->neon.sixth = UINT16_C(0x3155); in xnn_init_f16_hswish_neon_params()
2074 params->neon.three = UINT16_C(0x4200); in xnn_init_f16_hswish_neon_params()
2075 params->neon.six = UINT16_C(0x4600); in xnn_init_f16_hswish_neon_params()
2076 return sizeof(params->neon); in xnn_init_f16_hswish_neon_params()
2082 union xnn_f16_hswish_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_hswish_avx_params()
2085 params->avx.sixth[i] = 0x1.554000p-3f; in xnn_init_f16_hswish_avx_params()
2086 params->avx.three[i] = 3.0f; in xnn_init_f16_hswish_avx_params()
2087 params->avx.six[i] = UINT16_C(0x4600); in xnn_init_f16_hswish_avx_params()
2089 return sizeof(params->avx); in xnn_init_f16_hswish_avx_params()
2094 union xnn_f32_hswish_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_hswish_scalar_params()
2096 params->scalar.sixth = 0x1.555556p-3f; in xnn_init_f32_hswish_scalar_params()
2097 params->scalar.three = 3.0f; in xnn_init_f32_hswish_scalar_params()
2098 params->scalar.six = 6.0f; in xnn_init_f32_hswish_scalar_params()
2099 return sizeof(params->scalar); in xnn_init_f32_hswish_scalar_params()
2104 union xnn_f32_hswish_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_hswish_sse_params()
2107 params->sse.sixth[i] = 0x1.555556p-3f; in xnn_init_f32_hswish_sse_params()
2108 params->sse.half[i] = 0.5f; in xnn_init_f32_hswish_sse_params()
2109 params->sse.one[i] = 1.0f; in xnn_init_f32_hswish_sse_params()
2111 return sizeof(params->sse); in xnn_init_f32_hswish_sse_params()
2115 union xnn_f32_hswish_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_hswish_avx_params()
2118 params->avx.sixth[i] = 0x1.555556p-3f; in xnn_init_f32_hswish_avx_params()
2119 params->avx.half[i] = 0.5f; in xnn_init_f32_hswish_avx_params()
2120 params->avx.one[i] = 1.0f; in xnn_init_f32_hswish_avx_params()
2123 params->avx.mask_table[i] = -1; in xnn_init_f32_hswish_avx_params()
2126 params->avx.mask_table[i] = 0; in xnn_init_f32_hswish_avx_params()
2128 return sizeof(params->avx); in xnn_init_f32_hswish_avx_params()
2132 union xnn_f32_hswish_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_hswish_avx512_params()
2134 params->avx512.sixth = 0x1.555556p-3f; in xnn_init_f32_hswish_avx512_params()
2135 params->avx512.half = 0.5f; in xnn_init_f32_hswish_avx512_params()
2136 params->avx512.one = 1.0f; in xnn_init_f32_hswish_avx512_params()
2137 return sizeof(params->avx512); in xnn_init_f32_hswish_avx512_params()
2143 union xnn_f32_hswish_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_hswish_wasmsimd_params()
2146 params->wasmsimd.sixth[i] = 0x1.555556p-3f; in xnn_init_f32_hswish_wasmsimd_params()
2147 params->wasmsimd.three[i] = 3.0f; in xnn_init_f32_hswish_wasmsimd_params()
2148 params->wasmsimd.six[i] = 6.0f; in xnn_init_f32_hswish_wasmsimd_params()
2150 return sizeof(params->wasmsimd); in xnn_init_f32_hswish_wasmsimd_params()
2156 union xnn_f16_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2158 params->neonfp16arith_rr2_p2.magic_bias = UINT16_C(0x660F); // 0x1.83Cp+10h in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2159 params->neonfp16arith_rr2_p2.minus_log2e = UINT16_C(0xBDC5); // -0x1.714p+0h in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2160 params->neonfp16arith_rr2_p2.ln2_hi = UINT16_C(0x398C); // 0x1.630p-1h in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2161 params->neonfp16arith_rr2_p2.ln2_lo = UINT16_C(0x8AF4); // -0x1.BD0p-13h in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2162 params->neonfp16arith_rr2_p2.c2 = UINT16_C(0x37F9); // 0x1.FE4p-2h in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2163 params->neonfp16arith_rr2_p2.c1 = UINT16_C(0xBC0E); // -0x1.038p+0h in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2164 params->neonfp16arith_rr2_p2.denorm_cutoff = UINT16_C(0xC8DA); // -0x1.368p+3h in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2165 return sizeof(params->neonfp16arith_rr2_p2); in xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params()
2171 union xnn_f16_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2174 params->avx2_rr1_p2.sign_mask[i] = -0.0f; in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2175 params->avx2_rr1_p2.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2176 params->avx2_rr1_p2.log2e[i] = 0x1.715476p0f; in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2177 params->avx2_rr1_p2.minus_ln2[i] = -0x1.62E43p-1f; in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2178 params->avx2_rr1_p2.c2[i] = 0x1.FF3A32p-2f; in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2179 params->avx2_rr1_p2.c1[i] = 0x1.039E10p+0f; in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2180 params->avx2_rr1_p2.one[i] = 1.0f; in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2181 params->avx2_rr1_p2.denorm_cutoff[i] = -0x1.368000p+3f; in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2183 return sizeof(params->avx2_rr1_p2); in xnn_init_f16_sigmoid_avx2_rr1_p2_params()
2188 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2190 params->scalar_rr2_lut64_p2.magic_bias = 0x1.800000p17f; in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2191 params->scalar_rr2_lut64_p2.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2192 params->scalar_rr2_lut64_p2.ln2_hi = 0x1.630000p-1f; in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2193 params->scalar_rr2_lut64_p2.ln2_lo = -0x1.BD0106p-13f; in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2194 params->scalar_rr2_lut64_p2.c2 = 0x1.FFFF0Ap-2f; in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2195 params->scalar_rr2_lut64_p2.one = 1.0f; in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2196 params->scalar_rr2_lut64_p2.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2197 return sizeof(params->scalar_rr2_lut64_p2); in xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params()
2201 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2203 params->scalar_rr2_lut2048_p1.magic_bias = 0x1.800000p12f; in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2204 params->scalar_rr2_lut2048_p1.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2205 params->scalar_rr2_lut2048_p1.ln2_hi = 0x1.600000p-1f; in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2206 params->scalar_rr2_lut2048_p1.ln2_lo = 0x1.7217F8p-8f; in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2207 params->scalar_rr2_lut2048_p1.c1 = -0x1.FFFFFEp-1f; in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2208 params->scalar_rr2_lut2048_p1.one = 1.0f; in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2209 params->scalar_rr2_lut2048_p1.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2210 return sizeof(params->scalar_rr2_lut2048_p1); in xnn_init_f32_sigmoid_scalar_rr2_lut2048_p1_params()
2214 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2216 params->scalar_rr2_p5.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2217 params->scalar_rr2_p5.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2218 params->scalar_rr2_p5.ln2_hi = 0x1.62E400p-1f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2219 params->scalar_rr2_p5.ln2_lo = 0x1.7F7D1Cp-20f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2220 params->scalar_rr2_p5.c5 = -0x1.0F9F9Cp-7f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2221 params->scalar_rr2_p5.c4 = 0x1.573A1Ap-5f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2222 params->scalar_rr2_p5.c3 = -0x1.555A80p-3f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2223 params->scalar_rr2_p5.c2 = 0x1.FFFDC6p-2f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2224 params->scalar_rr2_p5.c1 = -0x1.FFFFF6p-1f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2225 params->scalar_rr2_p5.one = 1.0f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2226 params->scalar_rr2_p5.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2227 return sizeof(params->scalar_rr2_p5); in xnn_init_f32_sigmoid_scalar_rr2_p5_params()
2232 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params()
2234 params->neon_rr2_lut64_p2.magic_bias = 0x1.800000p17f; in xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params()
2235 params->neon_rr2_lut64_p2.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params()
2236 params->neon_rr2_lut64_p2.ln2_hi = 0x1.630000p-1f; in xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params()
2237 params->neon_rr2_lut64_p2.ln2_lo = -0x1.BD0106p-13f; in xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params()
2238 params->neon_rr2_lut64_p2.c2 = 0x1.FFFF0Ap-2f; in xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params()
2239 params->neon_rr2_lut64_p2.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params()
2240 return sizeof(params->neon_rr2_lut64_p2); in xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params()
2244 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_neon_rr2_lut2048_p1_params()
2246 params->neon_rr2_lut2048_p1.magic_bias = 0x1.800000p12f; in xnn_init_f32_sigmoid_neon_rr2_lut2048_p1_params()
2247 params->neon_rr2_lut2048_p1.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_neon_rr2_lut2048_p1_params()
2248 params->neon_rr2_lut2048_p1.ln2_hi = 0x1.600000p-1f; in xnn_init_f32_sigmoid_neon_rr2_lut2048_p1_params()
2249 params->neon_rr2_lut2048_p1.ln2_lo = 0x1.7217F8p-8f; in xnn_init_f32_sigmoid_neon_rr2_lut2048_p1_params()
2250 params->neon_rr2_lut2048_p1.c1 = -0x1.FFFFFEp-1f; in xnn_init_f32_sigmoid_neon_rr2_lut2048_p1_params()
2251 params->neon_rr2_lut2048_p1.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_neon_rr2_lut2048_p1_params()
2252 return sizeof(params->neon_rr2_lut2048_p1); in xnn_init_f32_sigmoid_neon_rr2_lut2048_p1_params()
2256 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2258 params->neon_rr2_p5.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2259 params->neon_rr2_p5.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2260 params->neon_rr2_p5.ln2_hi = 0x1.62E400p-1f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2261 params->neon_rr2_p5.ln2_lo = 0x1.7F7D1Cp-20f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2262 params->neon_rr2_p5.c5 = -0x1.0F9F9Cp-7f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2263 params->neon_rr2_p5.c4 = 0x1.573A1Ap-5f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2264 params->neon_rr2_p5.c3 = -0x1.555A80p-3f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2265 params->neon_rr2_p5.c2 = 0x1.FFFDC6p-2f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2266 params->neon_rr2_p5.c1 = -0x1.FFFFF6p-1f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2267 params->neon_rr2_p5.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2268 return sizeof(params->neon_rr2_p5); in xnn_init_f32_sigmoid_neon_rr2_p5_params()
2272 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params()
2274 params->neonfma_rr1_lut2048_p1.magic_bias = 0x1.800000p12f; in xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params()
2275 params->neonfma_rr1_lut2048_p1.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params()
2276 params->neonfma_rr1_lut2048_p1.ln2 = 0x1.62E430p-1f; in xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params()
2277 params->neonfma_rr1_lut2048_p1.c1 = -0x1.FFFFFEp-1f; in xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params()
2278 params->neonfma_rr1_lut2048_p1.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params()
2279 return sizeof(params->neonfma_rr1_lut2048_p1); in xnn_init_f32_sigmoid_neonfma_rr1_lut2048_p1_params()
2283 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params()
2285 params->neonfma_rr1_lut64_p2.magic_bias = 0x1.800000p17f; in xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params()
2286 params->neonfma_rr1_lut64_p2.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params()
2287 params->neonfma_rr1_lut64_p2.ln2 = 0x1.62E430p-1f; in xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params()
2288 params->neonfma_rr1_lut64_p2.c2 = 0x1.FFFF0Ap-2f; in xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params()
2289 params->neonfma_rr1_lut64_p2.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params()
2290 return sizeof(params->neonfma_rr1_lut64_p2); in xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params()
2294 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2296 params->neonfma_rr1_p5.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2297 params->neonfma_rr1_p5.minus_log2e = -0x1.715476p0f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2298 params->neonfma_rr1_p5.ln2 = 0x1.62E430p-1f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2299 params->neonfma_rr1_p5.c5 = -0x1.0F9F9Cp-7f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2300 params->neonfma_rr1_p5.c4 = 0x1.573A1Ap-5f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2301 params->neonfma_rr1_p5.c3 = -0x1.555A80p-3f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2302 params->neonfma_rr1_p5.c2 = 0x1.FFFDC6p-2f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2303 params->neonfma_rr1_p5.c1 = -0x1.FFFFF6p-1f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2304 params->neonfma_rr1_p5.denorm_cutoff = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2305 return sizeof(params->neonfma_rr1_p5); in xnn_init_f32_sigmoid_neonfma_rr1_p5_params()
2311 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2314 params->sse2_rr2_lut64_p2.sign_mask[i] = -0.0f; in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2315 params->sse2_rr2_lut64_p2.magic_bias[i] = 0x1.800000p17f; in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2316 params->sse2_rr2_lut64_p2.log2e[i] = 0x1.715476p0f; in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2317 params->sse2_rr2_lut64_p2.index_mask[i] = UINT32_C(0x3F); in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2318 params->sse2_rr2_lut64_p2.minus_ln2_hi[i] = -0x1.630000p-1f; in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2319 params->sse2_rr2_lut64_p2.minus_ln2_lo[i] = 0x1.BD0106p-13f; in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2320 params->sse2_rr2_lut64_p2.c2[i] = 0x1.FFFF0Ap-2f; in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2321 params->sse2_rr2_lut64_p2.one[i] = 1.0f; in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2322 params->sse2_rr2_lut64_p2.denorm_cutoff[i] = -0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2324 return sizeof(params->sse2_rr2_lut64_p2); in xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params()
2328 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2331 params->sse2_rr2_p5.sign_mask[i] = -0.0f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2332 params->sse2_rr2_p5.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2333 params->sse2_rr2_p5.log2e[i] = 0x1.715476p0f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2334 params->sse2_rr2_p5.minus_ln2_hi[i] = -0x1.62E400p-1f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2335 params->sse2_rr2_p5.minus_ln2_lo[i] = -0x1.7F7D1Cp-20f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2336 params->sse2_rr2_p5.c5[i] = 0x1.0F9F9Cp-7f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2337 params->sse2_rr2_p5.c4[i] = 0x1.573A1Ap-5f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2338 params->sse2_rr2_p5.c3[i] = 0x1.555A80p-3f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2339 params->sse2_rr2_p5.c2[i] = 0x1.FFFDC6p-2f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2340 params->sse2_rr2_p5.c1[i] = 0x1.FFFFF6p-1f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2341 params->sse2_rr2_p5.one[i] = 1.0f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2342 params->sse2_rr2_p5.denorm_cutoff[i] = -0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2344 return sizeof(params->sse2_rr2_p5); in xnn_init_f32_sigmoid_sse2_rr2_p5_params()
2348 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2351 params->avx_rr2_p5.sign_mask[i] = -0.0f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2352 params->avx_rr2_p5.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2353 params->avx_rr2_p5.log2e[i] = 0x1.715476p0f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2354 params->avx_rr2_p5.minus_ln2_hi[i] = -0x1.62E400p-1f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2355 params->avx_rr2_p5.minus_ln2_lo[i] = -0x1.7F7D1Cp-20f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2356 params->avx_rr2_p5.c5[i] = 0x1.0F9F9Cp-7f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2357 params->avx_rr2_p5.c4[i] = 0x1.573A1Ap-5f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2358 params->avx_rr2_p5.c3[i] = 0x1.555A80p-3f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2359 params->avx_rr2_p5.c2[i] = 0x1.FFFDC6p-2f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2360 params->avx_rr2_p5.c1[i] = 0x1.FFFFF6p-1f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2361 params->avx_rr2_p5.one[i] = 1.0f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2362 params->avx_rr2_p5.two[i] = 2.0f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2363 params->avx_rr2_p5.denorm_cutoff[i] = -0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2366 params->avx_rr2_p5.mask_table[i] = -1; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2369 params->avx_rr2_p5.mask_table[i] = 0; in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2371 return sizeof(params->avx_rr2_p5); in xnn_init_f32_sigmoid_avx_rr2_p5_params()
2375 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2378 params->avx2_rr1_p5.sign_mask[i] = -0.0f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2379 params->avx2_rr1_p5.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2380 params->avx2_rr1_p5.log2e[i] = 0x1.715476p0f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2381 params->avx2_rr1_p5.minus_ln2[i] = -0x1.62E430p-1f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2382 params->avx2_rr1_p5.c5[i] = 0x1.0F9F9Cp-7f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2383 params->avx2_rr1_p5.c4[i] = 0x1.573A1Ap-5f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2384 params->avx2_rr1_p5.c3[i] = 0x1.555A80p-3f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2385 params->avx2_rr1_p5.c2[i] = 0x1.FFFDC6p-2f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2386 params->avx2_rr1_p5.c1[i] = 0x1.FFFFF6p-1f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2387 params->avx2_rr1_p5.one[i] = 1.0f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2388 params->avx2_rr1_p5.denorm_cutoff[i] = -0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2391 params->avx2_rr1_p5.mask_table[i] = -1; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2394 params->avx2_rr1_p5.mask_table[i] = 0; in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2396 return sizeof(params->avx2_rr1_p5); in xnn_init_f32_sigmoid_avx2_rr1_p5_params()
2400 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2402 params->avx512_rr1_lut16_p3.sign_mask = UINT32_C(0x80000000); in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2403 params->avx512_rr1_lut16_p3.magic_bias = 0x1.800000p19f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2404 params->avx512_rr1_lut16_p3.log2e = 0x1.715476p0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2405 params->avx512_rr1_lut16_p3.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2406 params->avx512_rr1_lut16_p3.c3 = 0x1.55559Ap-3f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2407 params->avx512_rr1_lut16_p3.c2 = 0x1.00021Ep-1f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2408 params->avx512_rr1_lut16_p3.one = 1.0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2409 params->avx512_rr1_lut16_p3.table[ 0] = 0x1.000000p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2410 params->avx512_rr1_lut16_p3.table[ 1] = 0x1.0B5586p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2411 params->avx512_rr1_lut16_p3.table[ 2] = 0x1.172B84p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2412 params->avx512_rr1_lut16_p3.table[ 3] = 0x1.2387A6p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2413 params->avx512_rr1_lut16_p3.table[ 4] = 0x1.306FE0p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2414 params->avx512_rr1_lut16_p3.table[ 5] = 0x1.3DEA64p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2415 params->avx512_rr1_lut16_p3.table[ 6] = 0x1.4BFDAEp+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2416 params->avx512_rr1_lut16_p3.table[ 7] = 0x1.5AB07Ep+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2417 params->avx512_rr1_lut16_p3.table[ 8] = 0x1.6A09E6p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2418 params->avx512_rr1_lut16_p3.table[ 9] = 0x1.7A1148p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2419 params->avx512_rr1_lut16_p3.table[10] = 0x1.8ACE54p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2420 params->avx512_rr1_lut16_p3.table[11] = 0x1.9C4918p+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2421 params->avx512_rr1_lut16_p3.table[12] = 0x1.AE89FAp+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2422 params->avx512_rr1_lut16_p3.table[13] = 0x1.C199BEp+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2423 params->avx512_rr1_lut16_p3.table[14] = 0x1.D5818Ep+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2424 params->avx512_rr1_lut16_p3.table[15] = 0x1.EA4AFAp+0f; in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2425 return sizeof(params->avx512_rr1_lut16_p3); in xnn_init_f32_sigmoid_avx512_rr1_lut16_p3_params()
2429 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2431 params->avx512_rr2_lut32_p2.sign_mask = UINT32_C(0x80000000); in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2432 params->avx512_rr2_lut32_p2.magic_bias = 0x1.800000p18f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2433 params->avx512_rr2_lut32_p2.log2e = 0x1.715476p0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2434 params->avx512_rr2_lut32_p2.minus_ln2_hi = -0x1.62E430p-1f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2435 params->avx512_rr2_lut32_p2.minus_ln2_lo = 0x1.05C61p-29f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2436 params->avx512_rr2_lut32_p2.c2 = 0x1.000000p-1f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2437 params->avx512_rr2_lut32_p2.c1 = 0x1.0000F6p-0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2438 params->avx512_rr2_lut32_p2.one = 1.0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2440 params->avx512_rr2_lut32_p2.table_lo[ 0] = 0x1.000000p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2441 params->avx512_rr2_lut32_p2.table_lo[ 1] = 0x1.059B0Ep+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2442 params->avx512_rr2_lut32_p2.table_lo[ 2] = 0x1.0B5586p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2443 params->avx512_rr2_lut32_p2.table_lo[ 3] = 0x1.11301Ep+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2444 params->avx512_rr2_lut32_p2.table_lo[ 4] = 0x1.172B84p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2445 params->avx512_rr2_lut32_p2.table_lo[ 5] = 0x1.1D4874p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2446 params->avx512_rr2_lut32_p2.table_lo[ 6] = 0x1.2387A6p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2447 params->avx512_rr2_lut32_p2.table_lo[ 7] = 0x1.29E9E0p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2448 params->avx512_rr2_lut32_p2.table_lo[ 8] = 0x1.306FE0p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2449 params->avx512_rr2_lut32_p2.table_lo[ 9] = 0x1.371A74p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2450 params->avx512_rr2_lut32_p2.table_lo[10] = 0x1.3DEA64p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2451 params->avx512_rr2_lut32_p2.table_lo[11] = 0x1.44E086p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2452 params->avx512_rr2_lut32_p2.table_lo[12] = 0x1.4BFDAEp+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2453 params->avx512_rr2_lut32_p2.table_lo[13] = 0x1.5342B6p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2454 params->avx512_rr2_lut32_p2.table_lo[14] = 0x1.5AB07Ep+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2455 params->avx512_rr2_lut32_p2.table_lo[15] = 0x1.6247ECp+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2457 params->avx512_rr2_lut32_p2.table_hi[ 0] = 0x1.6A09E6p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2458 params->avx512_rr2_lut32_p2.table_hi[ 1] = 0x1.71F75Ep+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2459 params->avx512_rr2_lut32_p2.table_hi[ 2] = 0x1.7A1148p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2460 params->avx512_rr2_lut32_p2.table_hi[ 3] = 0x1.82589Ap+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2461 params->avx512_rr2_lut32_p2.table_hi[ 4] = 0x1.8ACE54p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2462 params->avx512_rr2_lut32_p2.table_hi[ 5] = 0x1.93737Cp+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2463 params->avx512_rr2_lut32_p2.table_hi[ 6] = 0x1.9C4918p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2464 params->avx512_rr2_lut32_p2.table_hi[ 7] = 0x1.A5503Cp+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2465 params->avx512_rr2_lut32_p2.table_hi[ 8] = 0x1.AE89FAp+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2466 params->avx512_rr2_lut32_p2.table_hi[ 9] = 0x1.B7F770p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2467 params->avx512_rr2_lut32_p2.table_hi[10] = 0x1.C199BEp+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2468 params->avx512_rr2_lut32_p2.table_hi[11] = 0x1.CB720Ep+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2469 params->avx512_rr2_lut32_p2.table_hi[12] = 0x1.D5818Ep+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2470 params->avx512_rr2_lut32_p2.table_hi[13] = 0x1.DFC974p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2471 params->avx512_rr2_lut32_p2.table_hi[14] = 0x1.EA4AFAp+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2472 params->avx512_rr2_lut32_p2.table_hi[15] = 0x1.F50766p+0f; in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2473 return sizeof(params->avx512_rr2_lut32_p2); in xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params()
2477 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2479 params->avx512_rr1_p5.sign_mask = UINT32_C(0x80000000); in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2480 params->avx512_rr1_p5.log2e = 0x1.715476p0f; in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2481 params->avx512_rr1_p5.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2482 params->avx512_rr1_p5.c5 = 0x1.0F9F9Cp-7f; in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2483 params->avx512_rr1_p5.c4 = 0x1.573A1Ap-5f; in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2484 params->avx512_rr1_p5.c3 = 0x1.555A80p-3f; in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2485 params->avx512_rr1_p5.c2 = 0x1.FFFDC6p-2f; in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2486 params->avx512_rr1_p5.c1 = 0x1.FFFFF6p-1f; in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2487 params->avx512_rr1_p5.one = 1.0f; in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2488 return sizeof(params->avx512_rr1_p5); in xnn_init_f32_sigmoid_avx512_rr1_p5_params()
2494 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2497 params->wasmsimd_rr2_lut64_p2.magic_bias[i] = 0x1.800000p17f; in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2498 params->wasmsimd_rr2_lut64_p2.minus_log2e[i] = -0x1.715476p0f; in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2499 params->wasmsimd_rr2_lut64_p2.index_mask[i] = UINT32_C(0x3F); in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2500 params->wasmsimd_rr2_lut64_p2.ln2_hi[i] = 0x1.630000p-1f; in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2501 params->wasmsimd_rr2_lut64_p2.ln2_lo[i] = -0x1.BD0106p-13f; in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2502 params->wasmsimd_rr2_lut64_p2.c2[i] = 0x1.FFFF0Ap-2f; in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2503 params->wasmsimd_rr2_lut64_p2.one[i] = 1.0f; in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2504 params->wasmsimd_rr2_lut64_p2.denorm_cutoff[i] = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2506 return sizeof(params->wasmsimd_rr2_lut64_p2); in xnn_init_f32_sigmoid_wasmsimd_rr2_lut64_p2_params()
2510 union xnn_f32_sigmoid_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2513 params->wasmsimd_rr2_p5.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2514 params->wasmsimd_rr2_p5.minus_log2e[i] = -0x1.715476p+0f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2515 params->wasmsimd_rr2_p5.ln2_hi[i] = 0x1.62E400p-1f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2516 params->wasmsimd_rr2_p5.ln2_lo[i] = 0x1.7F7D1Cp-20f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2517 params->wasmsimd_rr2_p5.c5[i] = -0x1.0F9F9Cp-7f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2518 params->wasmsimd_rr2_p5.c4[i] = 0x1.573A1Ap-5f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2519 params->wasmsimd_rr2_p5.c3[i] = -0x1.555A80p-3f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2520 params->wasmsimd_rr2_p5.c2[i] = 0x1.FFFDC6p-2f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2521 params->wasmsimd_rr2_p5.c1[i] = -0x1.FFFFF6p-1f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2522 params->wasmsimd_rr2_p5.one[i] = 1.0f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2523 params->wasmsimd_rr2_p5.denorm_cutoff[i] = 0x1.5D589Ep+6f; in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2525 return sizeof(params->wasmsimd_rr2_p5); in xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params()
2531 union xnn_f16_abs_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_abs_sse_params()
2534 params->sse.nonsign_mask[i] = UINT16_C(0x7FFF); in xnn_init_f16_abs_sse_params()
2536 return sizeof(params->sse); in xnn_init_f16_abs_sse_params()
2542 union xnn_f32_abs_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_abs_sse_params()
2545 params->sse.nonsign_mask[i] = math_nonsign_mask_f32(); in xnn_init_f32_abs_sse_params()
2547 return sizeof(params->sse); in xnn_init_f32_abs_sse_params()
2551 union xnn_f32_abs_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_abs_avx_params()
2554 params->avx.nonsign_mask[i] = math_nonsign_mask_f32(); in xnn_init_f32_abs_avx_params()
2557 params->avx.mask_table[i] = -1; in xnn_init_f32_abs_avx_params()
2560 params->avx.mask_table[i] = 0; in xnn_init_f32_abs_avx_params()
2562 return sizeof(params->avx); in xnn_init_f32_abs_avx_params()
2566 union xnn_f32_abs_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_abs_avx512_params()
2568 params->avx512.nonsign_mask = UINT32_C(0x7FFFFFFF); in xnn_init_f32_abs_avx512_params()
2569 return sizeof(params->avx512); in xnn_init_f32_abs_avx512_params()
2575 union xnn_f32_abs_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_abs_wasmsimd_params()
2577 params->wasmsimd.nonsign_mask[0] = math_nonsign_mask_f32(); in xnn_init_f32_abs_wasmsimd_params()
2578 params->wasmsimd.nonsign_mask[1] = math_nonsign_mask_f32(); in xnn_init_f32_abs_wasmsimd_params()
2579 return sizeof(params->wasmsimd); in xnn_init_f32_abs_wasmsimd_params()
2585 union xnn_f16_neg_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_neg_sse_params()
2588 params->sse.sign_mask[i] = UINT16_C(0x8000); in xnn_init_f16_neg_sse_params()
2590 return sizeof(params->sse); in xnn_init_f16_neg_sse_params()
2596 union xnn_f32_neg_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_neg_sse_params()
2599 params->sse.sign_mask[i] = -0.0f; in xnn_init_f32_neg_sse_params()
2601 return sizeof(params->sse); in xnn_init_f32_neg_sse_params()
2605 union xnn_f32_neg_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_neg_avx_params()
2608 params->avx.sign_mask[i] = -0.0f; in xnn_init_f32_neg_avx_params()
2611 params->avx.mask_table[i] = -1; in xnn_init_f32_neg_avx_params()
2614 params->avx.mask_table[i] = 0; in xnn_init_f32_neg_avx_params()
2616 return sizeof(params->avx); in xnn_init_f32_neg_avx_params()
2620 union xnn_f32_neg_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_neg_avx512_params()
2622 params->avx512.sign_mask = UINT32_C(0x80000000); in xnn_init_f32_neg_avx512_params()
2623 return sizeof(params->avx512); in xnn_init_f32_neg_avx512_params()
2629 union xnn_f32_neg_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_neg_wasmsimd_params()
2631 params->wasmsimd.sign_mask[0] = -0.0f; in xnn_init_f32_neg_wasmsimd_params()
2632 params->wasmsimd.sign_mask[1] = -0.0f; in xnn_init_f32_neg_wasmsimd_params()
2633 return sizeof(params->wasmsimd); in xnn_init_f32_neg_wasmsimd_params()
2639 union xnn_f32_rnd_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_rnd_sse2_params()
2642 params->sse2.sign_mask[i] = -0.0f; in xnn_init_f32_rnd_sse2_params()
2643 params->sse2.one[i] = 1.0f; in xnn_init_f32_rnd_sse2_params()
2645 return sizeof(params->sse2); in xnn_init_f32_rnd_sse2_params()
2649 union xnn_f32_rnd_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_rnd_avx_params()
2652 params->avx.mask_table[i] = -1; in xnn_init_f32_rnd_avx_params()
2655 params->avx.mask_table[i] = 0; in xnn_init_f32_rnd_avx_params()
2657 return sizeof(params->avx); in xnn_init_f32_rnd_avx_params()
2663 union xnn_f16_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2668 params->neonfp16arith_rr1_p3.prescale = prescale; in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2669 params->neonfp16arith_rr1_p3.sat_cutoff = UINT16_C(0xC829); // -0x1.0A4p+3h; in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2670 params->neonfp16arith_rr1_p3.magic_bias = UINT16_C(0x660F); // 0x1.83Cp+10h in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2671 params->neonfp16arith_rr1_p3.log2e = UINT16_C(0x3DC5); // 0x1.714p+0h in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2672 params->neonfp16arith_rr1_p3.minus_ln2 = UINT16_C(0xB98C); // -0x1.62E430p-1h in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2673 params->neonfp16arith_rr1_p3.c3 = UINT16_C(0x315B); // 0x1.56Cp-3h in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2674 params->neonfp16arith_rr1_p3.c2 = UINT16_C(0x3808); // 0x1.020p-1h in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2675 params->neonfp16arith_rr1_p3.minus_alpha = alpha ^ UINT16_C(0x8000); in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2676 params->neonfp16arith_rr1_p3.beta = beta; in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2677 return sizeof(params->neonfp16arith_rr1_p3); in xnn_init_f16_elu_neonfp16arith_rr1_p3_params()
2683 union xnn_f16_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_elu_avx2_rr1_p3_params()
2689 params->avx2_rr1_p3.prescale[i] = fp16_ieee_to_fp32_value(prescale); in xnn_init_f16_elu_avx2_rr1_p3_params()
2690 params->avx2_rr1_p3.sat_cutoff[i] = -0x1.0A4000p+3f; in xnn_init_f16_elu_avx2_rr1_p3_params()
2691 params->avx2_rr1_p3.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f16_elu_avx2_rr1_p3_params()
2692 params->avx2_rr1_p3.log2e[i] = 0x1.715476p+0f; in xnn_init_f16_elu_avx2_rr1_p3_params()
2693 params->avx2_rr1_p3.minus_ln2[i] = -0x1.62E430p-1f; in xnn_init_f16_elu_avx2_rr1_p3_params()
2694 params->avx2_rr1_p3.c3[i] = 0x1.5554DCp-3f; in xnn_init_f16_elu_avx2_rr1_p3_params()
2695 params->avx2_rr1_p3.c2[i] = 0x1.01EBB2p-1f; in xnn_init_f16_elu_avx2_rr1_p3_params()
2696 params->avx2_rr1_p3.c1[i] = 0x1.0002F2p+0f; in xnn_init_f16_elu_avx2_rr1_p3_params()
2697 params->avx2_rr1_p3.alpha[i] = fp16_ieee_to_fp32_value(alpha); in xnn_init_f16_elu_avx2_rr1_p3_params()
2698 params->avx2_rr1_p3.beta[i] = fp16_ieee_to_fp32_value(beta); in xnn_init_f16_elu_avx2_rr1_p3_params()
2700 return sizeof(params->avx2_rr1_p3); in xnn_init_f16_elu_avx2_rr1_p3_params()
2705 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2710 params->scalar_rr2_lut16_p3.prescale = prescale; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2711 params->scalar_rr2_lut16_p3.alpha = alpha; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2712 params->scalar_rr2_lut16_p3.beta = beta; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2713 params->scalar_rr2_lut16_p3.sat_cutoff = -0x1.154246p+4f; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2714 params->scalar_rr2_lut16_p3.magic_bias = 0x1.800000p19f; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2715 params->scalar_rr2_lut16_p3.log2e = 0x1.715476p+0f; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2716 params->scalar_rr2_lut16_p3.minus_ln2_hi = -0x1.62E400p-1f; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2717 params->scalar_rr2_lut16_p3.minus_ln2_lo = -0x1.7F7D1Cp-20f; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2718 params->scalar_rr2_lut16_p3.c3 = 0x1.55561Cp-3f; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2719 params->scalar_rr2_lut16_p3.c2 = 0x1.0001ECp-1f; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2720 params->scalar_rr2_lut16_p3.one = 1.0f; in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2721 return sizeof(params->scalar_rr2_lut16_p3); in xnn_init_f32_elu_scalar_rr2_lut16_p3_params()
2725 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_scalar_rr2_p6_params()
2730 params->scalar_rr2_p6.prescale = prescale; in xnn_init_f32_elu_scalar_rr2_p6_params()
2731 params->scalar_rr2_p6.alpha = alpha; in xnn_init_f32_elu_scalar_rr2_p6_params()
2732 params->scalar_rr2_p6.beta = beta; in xnn_init_f32_elu_scalar_rr2_p6_params()
2733 params->scalar_rr2_p6.sat_cutoff = -0x1.154246p+4f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2734 params->scalar_rr2_p6.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2735 params->scalar_rr2_p6.log2e = 0x1.715476p+0f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2736 params->scalar_rr2_p6.minus_ln2_hi = -0x1.62E440p-1f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2737 params->scalar_rr2_p6.minus_ln2_lo = 0x1.0105C6p-21f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2738 params->scalar_rr2_p6.c6 = 0x1.6b7338p-10f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2739 params->scalar_rr2_p6.c5 = 0x1.12278Ep-7f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2740 params->scalar_rr2_p6.c4 = 0x1.555716p-5f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2741 params->scalar_rr2_p6.c3 = 0x1.5554B0p-3f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2742 params->scalar_rr2_p6.c2 = 0x1.FFFFFEp-2f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2743 params->scalar_rr2_p6.one = 1.0f; in xnn_init_f32_elu_scalar_rr2_p6_params()
2744 return sizeof(params->scalar_rr2_p6); in xnn_init_f32_elu_scalar_rr2_p6_params()
2749 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2754 params->neon_rr2_lut16_p3.prescale = prescale; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2755 params->neon_rr2_lut16_p3.alpha = alpha; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2756 params->neon_rr2_lut16_p3.beta = beta; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2757 params->neon_rr2_lut16_p3.sat_cutoff = -0x1.154246p+4f; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2758 params->neon_rr2_lut16_p3.magic_bias = 0x1.800000p19f; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2759 params->neon_rr2_lut16_p3.log2e = 0x1.715476p+0f; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2760 params->neon_rr2_lut16_p3.minus_ln2_hi = -0x1.62E400p-1f; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2761 params->neon_rr2_lut16_p3.minus_ln2_lo = -0x1.7F7D1Cp-20f; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2762 params->neon_rr2_lut16_p3.c3 = 0x1.55561Cp-3f; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2763 params->neon_rr2_lut16_p3.c2 = 0x1.0001ECp-1f; in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2764 return sizeof(params->neon_rr2_lut16_p3); in xnn_init_f32_elu_neon_rr2_lut16_p3_params()
2768 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_neon_rr2_p6_params()
2773 params->neon_rr2_p6.prescale = prescale; in xnn_init_f32_elu_neon_rr2_p6_params()
2774 params->neon_rr2_p6.alpha = alpha; in xnn_init_f32_elu_neon_rr2_p6_params()
2775 params->neon_rr2_p6.beta = beta; in xnn_init_f32_elu_neon_rr2_p6_params()
2776 params->neon_rr2_p6.sat_cutoff = -0x1.154246p+4f; in xnn_init_f32_elu_neon_rr2_p6_params()
2777 params->neon_rr2_p6.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_elu_neon_rr2_p6_params()
2778 params->neon_rr2_p6.log2e = 0x1.715476p+0f; in xnn_init_f32_elu_neon_rr2_p6_params()
2779 params->neon_rr2_p6.minus_ln2_hi = -0x1.62E440p-1f; in xnn_init_f32_elu_neon_rr2_p6_params()
2780 params->neon_rr2_p6.minus_ln2_lo = 0x1.0105C6p-21f; in xnn_init_f32_elu_neon_rr2_p6_params()
2781 params->neon_rr2_p6.c6 = 0x1.6b7338p-10f; in xnn_init_f32_elu_neon_rr2_p6_params()
2782 params->neon_rr2_p6.c5 = 0x1.12278Ep-7f; in xnn_init_f32_elu_neon_rr2_p6_params()
2783 params->neon_rr2_p6.c4 = 0x1.555716p-5f; in xnn_init_f32_elu_neon_rr2_p6_params()
2784 params->neon_rr2_p6.c3 = 0x1.5554B0p-3f; in xnn_init_f32_elu_neon_rr2_p6_params()
2785 params->neon_rr2_p6.c2 = 0x1.FFFFFEp-2f; in xnn_init_f32_elu_neon_rr2_p6_params()
2786 return sizeof(params->neon_rr2_p6); in xnn_init_f32_elu_neon_rr2_p6_params()
2790 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2795 params->neonfma_rr1_lut16_p3.prescale = prescale; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2796 params->neonfma_rr1_lut16_p3.alpha = alpha; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2797 params->neonfma_rr1_lut16_p3.beta = beta; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2798 params->neonfma_rr1_lut16_p3.sat_cutoff = -0x1.154246p+4f; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2799 params->neonfma_rr1_lut16_p3.magic_bias = 0x1.800000p19f; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2800 params->neonfma_rr1_lut16_p3.log2e = 0x1.715476p+0f; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2801 params->neonfma_rr1_lut16_p3.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2802 params->neonfma_rr1_lut16_p3.c3 = 0x1.55561Cp-3f; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2803 params->neonfma_rr1_lut16_p3.c2 = 0x1.0001ECp-1f; in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2804 return sizeof(params->neonfma_rr1_lut16_p3); in xnn_init_f32_elu_neonfma_rr1_lut16_p3_params()
2808 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_neonfma_rr1_p6_params()
2813 params->neonfma_rr1_p6.prescale = prescale; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2814 params->neonfma_rr1_p6.alpha = alpha; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2815 params->neonfma_rr1_p6.beta = beta; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2816 params->neonfma_rr1_p6.sat_cutoff = -0x1.154246p+4f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2817 params->neonfma_rr1_p6.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2818 params->neonfma_rr1_p6.log2e = 0x1.715476p+0f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2819 params->neonfma_rr1_p6.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2820 params->neonfma_rr1_p6.c6 = 0x1.6b7338p-10f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2821 params->neonfma_rr1_p6.c5 = 0x1.12278Ep-7f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2822 params->neonfma_rr1_p6.c4 = 0x1.555716p-5f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2823 params->neonfma_rr1_p6.c3 = 0x1.5554B0p-3f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2824 params->neonfma_rr1_p6.c2 = 0x1.FFFFFEp-2f; in xnn_init_f32_elu_neonfma_rr1_p6_params()
2825 return sizeof(params->neonfma_rr1_p6); in xnn_init_f32_elu_neonfma_rr1_p6_params()
2831 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2837 params->sse2_rr2_lut16_p3.prescale[i] = prescale; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2838 params->sse2_rr2_lut16_p3.alpha[i] = alpha; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2839 params->sse2_rr2_lut16_p3.beta[i] = beta; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2840 params->sse2_rr2_lut16_p3.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2841 params->sse2_rr2_lut16_p3.magic_bias[i] = 0x1.800000p19f; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2842 params->sse2_rr2_lut16_p3.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2843 params->sse2_rr2_lut16_p3.index_mask[i] = UINT32_C(0xF); in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2844 params->sse2_rr2_lut16_p3.minus_ln2_hi[i] = -0x1.62E400p-1f; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2845 params->sse2_rr2_lut16_p3.minus_ln2_lo[i] = -0x1.7F7D1Cp-20f; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2846 params->sse2_rr2_lut16_p3.c3[i] = 0x1.55561Cp-3f; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2847 params->sse2_rr2_lut16_p3.c2[i] = 0x1.0001ECp-1f; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2848 params->sse2_rr2_lut16_p3.one[i] = 1.0f; in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2850 return sizeof(params->sse2_rr2_lut16_p3); in xnn_init_f32_elu_sse2_rr2_lut16_p3_params()
2854 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_sse2_rr2_p6_params()
2860 params->sse2_rr2_p6.prescale[i] = prescale; in xnn_init_f32_elu_sse2_rr2_p6_params()
2861 params->sse2_rr2_p6.alpha[i] = alpha; in xnn_init_f32_elu_sse2_rr2_p6_params()
2862 params->sse2_rr2_p6.beta[i] = beta; in xnn_init_f32_elu_sse2_rr2_p6_params()
2863 params->sse2_rr2_p6.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2864 params->sse2_rr2_p6.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2865 params->sse2_rr2_p6.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2866 params->sse2_rr2_p6.minus_ln2_hi[i] = -0x1.62E440p-1f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2867 params->sse2_rr2_p6.minus_ln2_lo[i] = 0x1.0105C6p-21f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2868 params->sse2_rr2_p6.c6[i] = 0x1.6b7338p-10f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2869 params->sse2_rr2_p6.c5[i] = 0x1.12278Ep-7f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2870 params->sse2_rr2_p6.c4[i] = 0x1.555716p-5f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2871 params->sse2_rr2_p6.c3[i] = 0x1.5554B0p-3f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2872 params->sse2_rr2_p6.c2[i] = 0x1.FFFFFEp-2f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2873 params->sse2_rr2_p6.one[i] = 1.0f; in xnn_init_f32_elu_sse2_rr2_p6_params()
2875 return sizeof(params->sse2_rr2_p6); in xnn_init_f32_elu_sse2_rr2_p6_params()
2879 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2885 params->avx_rr2_lut16_p3.prescale[i] = prescale; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2886 params->avx_rr2_lut16_p3.alpha[i] = alpha; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2887 params->avx_rr2_lut16_p3.beta[i] = beta; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2888 params->avx_rr2_lut16_p3.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2889 params->avx_rr2_lut16_p3.magic_bias[i] = 0x1.800000p19f; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2890 params->avx_rr2_lut16_p3.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2891 params->avx_rr2_lut16_p3.index_mask[i] = UINT32_C(0xF); in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2892 params->avx_rr2_lut16_p3.minus_ln2_hi[i] = -0x1.62E400p-1f; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2893 params->avx_rr2_lut16_p3.minus_ln2_lo[i] = -0x1.7F7D1Cp-20f; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2894 params->avx_rr2_lut16_p3.c3[i] = 0x1.55561Cp-3f; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2895 params->avx_rr2_lut16_p3.c2[i] = 0x1.0001ECp-1f; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2896 params->avx_rr2_lut16_p3.one[i] = 1.0f; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2899 params->avx_rr2_lut16_p3.mask_table[i] = -1; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2902 params->avx_rr2_lut16_p3.mask_table[i] = 0; in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2904 return sizeof(params->avx_rr2_lut16_p3); in xnn_init_f32_elu_avx_rr2_lut16_p3_params()
2908 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2914 params->avx_rr2_lut4_p4.prescale[i] = prescale; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2915 params->avx_rr2_lut4_p4.alpha[i] = alpha; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2916 params->avx_rr2_lut4_p4.beta[i] = beta; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2917 params->avx_rr2_lut4_p4.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2918 params->avx_rr2_lut4_p4.magic_bias[i] = 0x1.8003F8p21f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2919 params->avx_rr2_lut4_p4.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2920 params->avx_rr2_lut4_p4.index_mask[i] = UINT32_C(0x3); in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2922 params->avx_rr2_lut4_p4.table[0] = 0x1.000000p+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2923 params->avx_rr2_lut4_p4.table[1] = 0x1.306FE0p+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2924 params->avx_rr2_lut4_p4.table[2] = 0x1.6A09E6p+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2925 params->avx_rr2_lut4_p4.table[3] = 0x1.AE89FAp+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2926 params->avx_rr2_lut4_p4.table[4] = 0x1.000000p+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2927 params->avx_rr2_lut4_p4.table[5] = 0x1.306FE0p+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2928 params->avx_rr2_lut4_p4.table[6] = 0x1.6A09E6p+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2929 params->avx_rr2_lut4_p4.table[7] = 0x1.AE89FAp+0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2931 params->avx_rr2_lut4_p4.minus_ln2_hi[i] = -0x1.62E400p-1f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2932 params->avx_rr2_lut4_p4.minus_ln2_lo[i] = -0x1.7F7D1Cp-20f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2933 params->avx_rr2_lut4_p4.c4[i] = 0x1.554F9Ap-5f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2934 params->avx_rr2_lut4_p4.c3[i] = 0x1.557082p-3f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2935 params->avx_rr2_lut4_p4.c2[i] = 0x1.000002p-1f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2936 params->avx_rr2_lut4_p4.one[i] = 1.0f; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2939 params->avx_rr2_lut4_p4.mask_table[i] = -1; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2942 params->avx_rr2_lut4_p4.mask_table[i] = 0; in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2944 return sizeof(params->avx_rr2_lut4_p4); in xnn_init_f32_elu_avx_rr2_lut4_p4_params()
2948 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx_rr2_p6_params()
2954 params->avx_rr2_p6.prescale[i] = prescale; in xnn_init_f32_elu_avx_rr2_p6_params()
2955 params->avx_rr2_p6.alpha[i] = alpha; in xnn_init_f32_elu_avx_rr2_p6_params()
2956 params->avx_rr2_p6.beta[i] = beta; in xnn_init_f32_elu_avx_rr2_p6_params()
2957 params->avx_rr2_p6.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_avx_rr2_p6_params()
2958 params->avx_rr2_p6.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_elu_avx_rr2_p6_params()
2959 params->avx_rr2_p6.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_avx_rr2_p6_params()
2960 params->avx_rr2_p6.minus_ln2_hi[i] = -0x1.62E440p-1f; in xnn_init_f32_elu_avx_rr2_p6_params()
2961 params->avx_rr2_p6.minus_ln2_lo[i] = 0x1.0105C6p-21f; in xnn_init_f32_elu_avx_rr2_p6_params()
2962 params->avx_rr2_p6.c6[i] = 0x1.6b7338p-10f; in xnn_init_f32_elu_avx_rr2_p6_params()
2963 params->avx_rr2_p6.c5[i] = 0x1.12278Ep-7f; in xnn_init_f32_elu_avx_rr2_p6_params()
2964 params->avx_rr2_p6.c4[i] = 0x1.555716p-5f; in xnn_init_f32_elu_avx_rr2_p6_params()
2965 params->avx_rr2_p6.c3[i] = 0x1.5554B0p-3f; in xnn_init_f32_elu_avx_rr2_p6_params()
2966 params->avx_rr2_p6.c2[i] = 0x1.FFFFFEp-2f; in xnn_init_f32_elu_avx_rr2_p6_params()
2967 params->avx_rr2_p6.one[i] = 1.0f; in xnn_init_f32_elu_avx_rr2_p6_params()
2970 params->avx_rr2_p6.mask_table[i] = -1; in xnn_init_f32_elu_avx_rr2_p6_params()
2973 params->avx_rr2_p6.mask_table[i] = 0; in xnn_init_f32_elu_avx_rr2_p6_params()
2975 return sizeof(params->avx_rr2_p6); in xnn_init_f32_elu_avx_rr2_p6_params()
2979 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2985 params->avx2_rr1_lut16_p3.prescale[i] = prescale; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2986 params->avx2_rr1_lut16_p3.alpha[i] = alpha; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2987 params->avx2_rr1_lut16_p3.beta[i] = beta; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2988 params->avx2_rr1_lut16_p3.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2989 params->avx2_rr1_lut16_p3.magic_bias[i] = 0x1.800000p19f; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2990 params->avx2_rr1_lut16_p3.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2991 params->avx2_rr1_lut16_p3.index_mask[i] = UINT32_C(0xF); in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2992 params->avx2_rr1_lut16_p3.minus_ln2[i] = -0x1.62E430p-1f; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2993 params->avx2_rr1_lut16_p3.c3[i] = 0x1.55561Cp-3f; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2994 params->avx2_rr1_lut16_p3.c2[i] = 0x1.0001ECp-1f; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
2997 params->avx2_rr1_lut16_p3.mask_table[i] = -1; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
3000 params->avx2_rr1_lut16_p3.mask_table[i] = 0; in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
3002 return sizeof(params->avx2_rr1_lut16_p3); in xnn_init_f32_elu_avx2_rr1_lut16_p3_params()
3006 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3012 params->avx2_rr1_lut8_p4.prescale[i] = prescale; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3013 params->avx2_rr1_lut8_p4.alpha[i] = alpha; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3014 params->avx2_rr1_lut8_p4.beta[i] = beta; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3015 params->avx2_rr1_lut8_p4.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3016 params->avx2_rr1_lut8_p4.magic_bias[i] = 0x1.800000p20f; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3017 params->avx2_rr1_lut8_p4.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3019 params->avx2_rr1_lut8_p4.table[0] = UINT32_C(0x3F800000); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3020 params->avx2_rr1_lut8_p4.table[1] = UINT32_C(0x3F7B95C2); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3021 params->avx2_rr1_lut8_p4.table[2] = UINT32_C(0x3F7837F0); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3022 params->avx2_rr1_lut8_p4.table[3] = UINT32_C(0x3F75FED7); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3023 params->avx2_rr1_lut8_p4.table[4] = UINT32_C(0x3F7504F3); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3024 params->avx2_rr1_lut8_p4.table[5] = UINT32_C(0x3F75672A); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3025 params->avx2_rr1_lut8_p4.table[6] = UINT32_C(0x3F7744FD); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3026 params->avx2_rr1_lut8_p4.table[7] = UINT32_C(0x3F7AC0C7); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3028 params->avx2_rr1_lut8_p4.minus_ln2[i] = -0x1.62E430p-1f; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3029 params->avx2_rr1_lut8_p4.c4[i] = 0x1.5558ECp-5f; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3030 params->avx2_rr1_lut8_p4.c3[i] = 0x1.555C20p-3f; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3031 params->avx2_rr1_lut8_p4.c2[i] = 0x1.000000p-1f; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3034 params->avx2_rr1_lut8_p4.mask_table[i] = -1; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3037 params->avx2_rr1_lut8_p4.mask_table[i] = 0; in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3039 return sizeof(params->avx2_rr1_lut8_p4); in xnn_init_f32_elu_avx2_rr1_lut8_p4_params()
3043 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3049 params->avx2_rr1_lut4_p4.prescale[i] = prescale; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3050 params->avx2_rr1_lut4_p4.alpha[i] = alpha; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3051 params->avx2_rr1_lut4_p4.beta[i] = beta; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3052 params->avx2_rr1_lut4_p4.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3053 params->avx2_rr1_lut4_p4.magic_bias[i] = 0x1.800000p21f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3054 params->avx2_rr1_lut4_p4.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3056 params->avx2_rr1_lut4_p4.table[0] = 0x1.000000p+0f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3057 params->avx2_rr1_lut4_p4.table[1] = 0x1.F06FE0p-1f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3058 params->avx2_rr1_lut4_p4.table[2] = 0x1.EA09E6p-1f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3059 params->avx2_rr1_lut4_p4.table[3] = 0x1.EE89FAp-1f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3060 params->avx2_rr1_lut4_p4.table[4] = 0x1.000000p+0f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3061 params->avx2_rr1_lut4_p4.table[5] = 0x1.F06FE0p-1f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3062 params->avx2_rr1_lut4_p4.table[6] = 0x1.EA09E6p-1f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3063 params->avx2_rr1_lut4_p4.table[7] = 0x1.EE89FAp-1f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3065 params->avx2_rr1_lut4_p4.minus_ln2[i] = -0x1.62E430p-1f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3066 params->avx2_rr1_lut4_p4.c4[i] = 0x1.554F9Ap-5f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3067 params->avx2_rr1_lut4_p4.c3[i] = 0x1.557082p-3f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3068 params->avx2_rr1_lut4_p4.c2[i] = 0x1.000002p-1f; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3071 params->avx2_rr1_lut4_p4.mask_table[i] = -1; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3074 params->avx2_rr1_lut4_p4.mask_table[i] = 0; in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3076 return sizeof(params->avx2_rr1_lut4_p4); in xnn_init_f32_elu_avx2_rr1_lut4_p4_params()
3080 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx2_rr1_p6_params()
3086 params->avx2_rr1_p6.prescale[i] = prescale; in xnn_init_f32_elu_avx2_rr1_p6_params()
3087 params->avx2_rr1_p6.alpha[i] = alpha; in xnn_init_f32_elu_avx2_rr1_p6_params()
3088 params->avx2_rr1_p6.beta[i] = beta; in xnn_init_f32_elu_avx2_rr1_p6_params()
3089 params->avx2_rr1_p6.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3090 params->avx2_rr1_p6.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3091 params->avx2_rr1_p6.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3092 params->avx2_rr1_p6.minus_ln2[i] = -0x1.62E430p-1f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3093 params->avx2_rr1_p6.c6[i] = 0x1.6B7338p-10f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3094 params->avx2_rr1_p6.c5[i] = 0x1.12278Ep-7f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3095 params->avx2_rr1_p6.c4[i] = 0x1.555716p-5f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3096 params->avx2_rr1_p6.c3[i] = 0x1.5554B0p-3f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3097 params->avx2_rr1_p6.c2[i] = 0x1.FFFFFEp-2f; in xnn_init_f32_elu_avx2_rr1_p6_params()
3100 params->avx2_rr1_p6.mask_table[i] = -1; in xnn_init_f32_elu_avx2_rr1_p6_params()
3103 params->avx2_rr1_p6.mask_table[i] = 0; in xnn_init_f32_elu_avx2_rr1_p6_params()
3105 return sizeof(params->avx2_rr1_p6); in xnn_init_f32_elu_avx2_rr1_p6_params()
3109 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3114 params->avx512_rr1_lut16_p3.prescale = prescale; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3115 params->avx512_rr1_lut16_p3.alpha = alpha; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3116 params->avx512_rr1_lut16_p3.beta = beta; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3117 params->avx512_rr1_lut16_p3.sat_cutoff = -0x1.154246p+4f; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3118 params->avx512_rr1_lut16_p3.magic_bias = 0x1.800000p19f; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3119 params->avx512_rr1_lut16_p3.log2e = 0x1.715476p+0f; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3120 params->avx512_rr1_lut16_p3.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3121 params->avx512_rr1_lut16_p3.c3 = 0x1.55561Cp-3f; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3122 params->avx512_rr1_lut16_p3.c2 = 0x1.0001ECp-1f; in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3123 params->avx512_rr1_lut16_p3.table[ 0] = UINT32_C(0x3F800000); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3124 params->avx512_rr1_lut16_p3.table[ 1] = UINT32_C(0x3F7DAAC3); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3125 params->avx512_rr1_lut16_p3.table[ 2] = UINT32_C(0x3F7B95C2); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3126 params->avx512_rr1_lut16_p3.table[ 3] = UINT32_C(0x3F79C3D3); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3127 params->avx512_rr1_lut16_p3.table[ 4] = UINT32_C(0x3F7837F0); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3128 params->avx512_rr1_lut16_p3.table[ 5] = UINT32_C(0x3F76F532); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3129 params->avx512_rr1_lut16_p3.table[ 6] = UINT32_C(0x3F75FED7); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3130 params->avx512_rr1_lut16_p3.table[ 7] = UINT32_C(0x3F75583F); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3131 params->avx512_rr1_lut16_p3.table[ 8] = UINT32_C(0x3F7504F3); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3132 params->avx512_rr1_lut16_p3.table[ 9] = UINT32_C(0x3F7508A4); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3133 params->avx512_rr1_lut16_p3.table[10] = UINT32_C(0x3F75672A); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3134 params->avx512_rr1_lut16_p3.table[11] = UINT32_C(0x3F76248C); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3135 params->avx512_rr1_lut16_p3.table[12] = UINT32_C(0x3F7744FD); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3136 params->avx512_rr1_lut16_p3.table[13] = UINT32_C(0x3F78CCDF); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3137 params->avx512_rr1_lut16_p3.table[14] = UINT32_C(0x3F7AC0C7); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3138 params->avx512_rr1_lut16_p3.table[15] = UINT32_C(0x3F7D257D); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3139 return sizeof(params->avx512_rr1_lut16_p3); in xnn_init_f32_elu_avx512_rr1_lut16_p3_params()
3143 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_avx512_rr1_p6_params()
3148 params->avx512_rr1_p6.prescale = prescale; in xnn_init_f32_elu_avx512_rr1_p6_params()
3149 params->avx512_rr1_p6.alpha = alpha; in xnn_init_f32_elu_avx512_rr1_p6_params()
3150 params->avx512_rr1_p6.beta = beta; in xnn_init_f32_elu_avx512_rr1_p6_params()
3151 params->avx512_rr1_p6.sat_cutoff = -0x1.154246p+4f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3152 params->avx512_rr1_p6.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3153 params->avx512_rr1_p6.log2e = 0x1.715476p+0f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3154 params->avx512_rr1_p6.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3155 params->avx512_rr1_p6.c6 = 0x1.6B7338p-10f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3156 params->avx512_rr1_p6.c5 = 0x1.12278Ep-7f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3157 params->avx512_rr1_p6.c4 = 0x1.555716p-5f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3158 params->avx512_rr1_p6.c3 = 0x1.5554B0p-3f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3159 params->avx512_rr1_p6.c2 = 0x1.FFFFFEp-2f; in xnn_init_f32_elu_avx512_rr1_p6_params()
3160 return sizeof(params->avx512_rr1_p6); in xnn_init_f32_elu_avx512_rr1_p6_params()
3166 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3172 params->wasmsimd_rr2_lut16_p3.prescale[i] = prescale; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3173 params->wasmsimd_rr2_lut16_p3.alpha[i] = alpha; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3174 params->wasmsimd_rr2_lut16_p3.beta[i] = beta; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3175 params->wasmsimd_rr2_lut16_p3.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3176 params->wasmsimd_rr2_lut16_p3.magic_bias[i] = 0x1.800000p19f; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3177 params->wasmsimd_rr2_lut16_p3.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3178 params->wasmsimd_rr2_lut16_p3.index_mask[i] = UINT32_C(0xF); in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3179 params->wasmsimd_rr2_lut16_p3.minus_ln2_hi[i] = -0x1.62E400p-1f; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3180 params->wasmsimd_rr2_lut16_p3.minus_ln2_lo[i] = -0x1.7F7D1Cp-20f; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3181 params->wasmsimd_rr2_lut16_p3.c3[i] = 0x1.55561Cp-3f; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3182 params->wasmsimd_rr2_lut16_p3.c2[i] = 0x1.0001ECp-1f; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3183 params->wasmsimd_rr2_lut16_p3.one[i] = 1.0f; in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3185 return sizeof(params->wasmsimd_rr2_lut16_p3); in xnn_init_f32_elu_wasmsimd_rr2_lut16_p3_params()
3189 union xnn_f32_elu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3195 params->wasmsimd_rr2_p6.prescale[i] = prescale; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3196 params->wasmsimd_rr2_p6.alpha[i] = alpha; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3197 params->wasmsimd_rr2_p6.beta[i] = beta; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3198 params->wasmsimd_rr2_p6.sat_cutoff[i] = -0x1.154246p+4f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3199 params->wasmsimd_rr2_p6.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3200 params->wasmsimd_rr2_p6.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3201 params->wasmsimd_rr2_p6.minus_ln2_hi[i] = -0x1.62E440p-1f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3202 params->wasmsimd_rr2_p6.minus_ln2_lo[i] = 0x1.0105C6p-21f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3203 params->wasmsimd_rr2_p6.c6[i] = 0x1.6b7338p-10f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3204 params->wasmsimd_rr2_p6.c5[i] = 0x1.12278Ep-7f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3205 params->wasmsimd_rr2_p6.c4[i] = 0x1.555716p-5f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3206 params->wasmsimd_rr2_p6.c3[i] = 0x1.5554B0p-3f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3207 params->wasmsimd_rr2_p6.c2[i] = 0x1.FFFFFEp-2f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3208 params->wasmsimd_rr2_p6.one[i] = 1.0f; in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3210 return sizeof(params->wasmsimd_rr2_p6); in xnn_init_f32_elu_wasmsimd_rr2_p6_params()
3216 union xnn_f16_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3218 params->neonfp16arith_rr2_p2.magic_bias = UINT16_C(0x660F); // 0x1.83Cp+10h in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3219 params->neonfp16arith_rr2_p2.log2e = UINT16_C(0x3DC5); // 0x1.714p+0h in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3220 params->neonfp16arith_rr2_p2.minus_ln2_hi = UINT16_C(0xB98C); // -0x1.630p-1h in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3221 params->neonfp16arith_rr2_p2.minus_ln2_lo = UINT16_C(0x0AF4); // 0x1.BD0p-13h in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3222 params->neonfp16arith_rr2_p2.c2 = UINT16_C(0x37F9); // 0x1.FE4p-2h in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3223 params->neonfp16arith_rr2_p2.c1 = UINT16_C(0x3C0E); // 0x1.038p+0h in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3224 params->neonfp16arith_rr2_p2.denorm_cutoff = UINT16_C(0xC8DA); // -0x1.368p+3h in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3225 return sizeof(params->neonfp16arith_rr2_p2); in xnn_init_f16_expminus_neonfp16arith_rr2_p2_params()
3231 union xnn_f16_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_expminus_avx2_rr1_p2_params()
3234 params->avx2_rr1_p2.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f16_expminus_avx2_rr1_p2_params()
3235 params->avx2_rr1_p2.log2e[i] = 0x1.715476p0f; in xnn_init_f16_expminus_avx2_rr1_p2_params()
3236 params->avx2_rr1_p2.minus_ln2[i] = -0x1.62E43p-1f; in xnn_init_f16_expminus_avx2_rr1_p2_params()
3237 params->avx2_rr1_p2.c2[i] = 0x1.FF3A32p-2f; in xnn_init_f16_expminus_avx2_rr1_p2_params()
3238 params->avx2_rr1_p2.c1[i] = 0x1.039E10p+0f; in xnn_init_f16_expminus_avx2_rr1_p2_params()
3239 params->avx2_rr1_p2.denorm_cutoff[i] = -0x1.368000p+3f; in xnn_init_f16_expminus_avx2_rr1_p2_params()
3241 return sizeof(params->avx2_rr1_p2); in xnn_init_f16_expminus_avx2_rr1_p2_params()
3246 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_scalar_rr2_p5_params()
3248 params->scalar_rr2_p5.log2e = 0x1.715476p+0f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3249 params->scalar_rr2_p5.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3250 params->scalar_rr2_p5.minus_ln2_hi = -0x1.62E400p-1f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3251 params->scalar_rr2_p5.minus_ln2_lo = -0x1.7F7D1Cp-20f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3252 params->scalar_rr2_p5.c5 = 0x1.0F9F9Cp-7f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3253 params->scalar_rr2_p5.c4 = 0x1.573A1Ap-5f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3254 params->scalar_rr2_p5.c3 = 0x1.555A80p-3f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3255 params->scalar_rr2_p5.c2 = 0x1.FFFDC6p-2f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3256 params->scalar_rr2_p5.c1 = 0x1.FFFFF6p-1f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3257 params->scalar_rr2_p5.denorm_cutoff = -0x1.5D589Ep6f; in xnn_init_f32_expminus_scalar_rr2_p5_params()
3258 return sizeof(params->scalar_rr2_p5); in xnn_init_f32_expminus_scalar_rr2_p5_params()
3262 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_scalar_rr2_lut64_p2_params()
3264 params->scalar_rr2_lut64_p2.log2e = 0x1.715476p0f; in xnn_init_f32_expminus_scalar_rr2_lut64_p2_params()
3265 params->scalar_rr2_lut64_p2.magic_bias = 0x1.800000p17f; in xnn_init_f32_expminus_scalar_rr2_lut64_p2_params()
3266 params->scalar_rr2_lut64_p2.minus_ln2_hi = -0x1.630000p-1f; in xnn_init_f32_expminus_scalar_rr2_lut64_p2_params()
3267 params->scalar_rr2_lut64_p2.minus_ln2_lo = 0x1.BD0106p-13f; in xnn_init_f32_expminus_scalar_rr2_lut64_p2_params()
3268 params->scalar_rr2_lut64_p2.c2 = 0x1.FFFF0Ap-2f; in xnn_init_f32_expminus_scalar_rr2_lut64_p2_params()
3269 params->scalar_rr2_lut64_p2.denorm_cutoff = -0x1.5D589Ep6f; in xnn_init_f32_expminus_scalar_rr2_lut64_p2_params()
3270 return sizeof(params->scalar_rr2_lut64_p2); in xnn_init_f32_expminus_scalar_rr2_lut64_p2_params()
3275 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_neon_rr2_p5_params()
3277 params->neon_rr2_p5.log2e = 0x1.715476p+0f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3278 params->neon_rr2_p5.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3279 params->neon_rr2_p5.minus_ln2_hi = -0x1.62E400p-1f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3280 params->neon_rr2_p5.minus_ln2_lo = -0x1.7F7D1Cp-20f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3281 params->neon_rr2_p5.c5 = 0x1.0F9F9Cp-7f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3282 params->neon_rr2_p5.c4 = 0x1.573A1Ap-5f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3283 params->neon_rr2_p5.c3 = 0x1.555A80p-3f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3284 params->neon_rr2_p5.c2 = 0x1.FFFDC6p-2f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3285 params->neon_rr2_p5.c1 = 0x1.FFFFF6p-1f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3286 params->neon_rr2_p5.denorm_cutoff = -0x1.5D589Ep6f; in xnn_init_f32_expminus_neon_rr2_p5_params()
3287 return sizeof(params->neon_rr2_p5); in xnn_init_f32_expminus_neon_rr2_p5_params()
3291 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_neon_rr2_lut64_p2_params()
3293 params->neon_rr2_lut64_p2.log2e = 0x1.715476p+0f; in xnn_init_f32_expminus_neon_rr2_lut64_p2_params()
3294 params->neon_rr2_lut64_p2.magic_bias = 0x1.800000p17f; in xnn_init_f32_expminus_neon_rr2_lut64_p2_params()
3295 params->neon_rr2_lut64_p2.minus_ln2_hi = -0x1.62E400p-1f; in xnn_init_f32_expminus_neon_rr2_lut64_p2_params()
3296 params->neon_rr2_lut64_p2.minus_ln2_lo = -0x1.7F7D1Cp-20f; in xnn_init_f32_expminus_neon_rr2_lut64_p2_params()
3297 params->neon_rr2_lut64_p2.c2 = 0x1.FFFF0Ap-2f; in xnn_init_f32_expminus_neon_rr2_lut64_p2_params()
3298 params->neon_rr2_lut64_p2.denorm_cutoff = -0x1.5D589Ep6f; in xnn_init_f32_expminus_neon_rr2_lut64_p2_params()
3299 return sizeof(params->neon_rr2_lut64_p2); in xnn_init_f32_expminus_neon_rr2_lut64_p2_params()
3303 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3305 params->neonfma_rr1_p5.log2e = 0x1.715476p+0f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3306 params->neonfma_rr1_p5.magic_bias = 0x1.8000FEp23f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3307 params->neonfma_rr1_p5.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3308 params->neonfma_rr1_p5.c5 = 0x1.0F9F9Cp-7f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3309 params->neonfma_rr1_p5.c4 = 0x1.573A1Ap-5f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3310 params->neonfma_rr1_p5.c3 = 0x1.555A80p-3f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3311 params->neonfma_rr1_p5.c2 = 0x1.FFFDC6p-2f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3312 params->neonfma_rr1_p5.c1 = 0x1.FFFFF6p-1f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3313 params->neonfma_rr1_p5.denorm_cutoff = -0x1.5D589Ep6f; in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3314 return sizeof(params->neonfma_rr1_p5); in xnn_init_f32_expminus_neonfma_rr1_p5_params()
3318 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_neonfma_rr1_lut64_p2_params()
3320 params->neonfma_rr1_lut64_p2.log2e = 0x1.715476p+0f; in xnn_init_f32_expminus_neonfma_rr1_lut64_p2_params()
3321 params->neonfma_rr1_lut64_p2.magic_bias = 0x1.800000p17f; in xnn_init_f32_expminus_neonfma_rr1_lut64_p2_params()
3322 params->neonfma_rr1_lut64_p2.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_expminus_neonfma_rr1_lut64_p2_params()
3323 params->neonfma_rr1_lut64_p2.c2 = 0x1.FFFF0Ap-2f; in xnn_init_f32_expminus_neonfma_rr1_lut64_p2_params()
3324 params->neonfma_rr1_lut64_p2.denorm_cutoff = -0x1.5D589Ep6f; in xnn_init_f32_expminus_neonfma_rr1_lut64_p2_params()
3325 return sizeof(params->neonfma_rr1_lut64_p2); in xnn_init_f32_expminus_neonfma_rr1_lut64_p2_params()
3331 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_sse2_rr2_p5_params()
3334 params->sse2_rr2_p5.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3335 params->sse2_rr2_p5.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3336 params->sse2_rr2_p5.minus_ln2_hi[i] = -0x1.62E400p-1f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3337 params->sse2_rr2_p5.minus_ln2_lo[i] = -0x1.7F7D1Cp-20f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3338 params->sse2_rr2_p5.c5[i] = 0x1.0F9F9Cp-7f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3339 params->sse2_rr2_p5.c4[i] = 0x1.573A1Ap-5f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3340 params->sse2_rr2_p5.c3[i] = 0x1.555A80p-3f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3341 params->sse2_rr2_p5.c2[i] = 0x1.FFFDC6p-2f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3342 params->sse2_rr2_p5.c1[i] = 0x1.FFFFF6p-1f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3343 params->sse2_rr2_p5.denorm_cutoff[i] = -0x1.5D589Ep6f; in xnn_init_f32_expminus_sse2_rr2_p5_params()
3345 return sizeof(params->sse2_rr2_p5); in xnn_init_f32_expminus_sse2_rr2_p5_params()
3349 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_avx2_rr1_p5_params()
3352 params->avx2_rr1_p5.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3353 params->avx2_rr1_p5.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3354 params->avx2_rr1_p5.minus_ln2[i] = -0x1.62E430p-1f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3355 params->avx2_rr1_p5.c5[i] = 0x1.0F9F9Cp-7f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3356 params->avx2_rr1_p5.c4[i] = 0x1.573A1Ap-5f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3357 params->avx2_rr1_p5.c3[i] = 0x1.555A80p-3f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3358 params->avx2_rr1_p5.c2[i] = 0x1.FFFDC6p-2f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3359 params->avx2_rr1_p5.c1[i] = 0x1.FFFFF6p-1f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3360 params->avx2_rr1_p5.denorm_cutoff[i] = -0x1.5D589Ep6f; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3363 params->avx2_rr1_p5.mask_table[i] = -1; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3366 params->avx2_rr1_p5.mask_table[i] = 0; in xnn_init_f32_expminus_avx2_rr1_p5_params()
3368 return sizeof(params->avx2_rr1_p5); in xnn_init_f32_expminus_avx2_rr1_p5_params()
3372 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_avx512_rr1_p5_params()
3374 params->avx512_rr1_p5.log2e = 0x1.715476p+0f; in xnn_init_f32_expminus_avx512_rr1_p5_params()
3375 params->avx512_rr1_p5.minus_ln2 = -0x1.62E430p-1f; in xnn_init_f32_expminus_avx512_rr1_p5_params()
3376 params->avx512_rr1_p5.c5 = 0x1.0F9F9Cp-7f; in xnn_init_f32_expminus_avx512_rr1_p5_params()
3377 params->avx512_rr1_p5.c4 = 0x1.573A1Ap-5f; in xnn_init_f32_expminus_avx512_rr1_p5_params()
3378 params->avx512_rr1_p5.c3 = 0x1.555A80p-3f; in xnn_init_f32_expminus_avx512_rr1_p5_params()
3379 params->avx512_rr1_p5.c2 = 0x1.FFFDC6p-2f; in xnn_init_f32_expminus_avx512_rr1_p5_params()
3380 params->avx512_rr1_p5.c1 = 0x1.FFFFF6p-1f; in xnn_init_f32_expminus_avx512_rr1_p5_params()
3381 params->avx512_rr1_p5.c0 = 1.0f; in xnn_init_f32_expminus_avx512_rr1_p5_params()
3382 return sizeof(params->avx512_rr1_p5); in xnn_init_f32_expminus_avx512_rr1_p5_params()
3388 union xnn_f32_expminus_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3391 params->wasmsimd_rr2_p5.log2e[i] = 0x1.715476p+0f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3392 params->wasmsimd_rr2_p5.magic_bias[i] = 0x1.8000FEp23f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3393 params->wasmsimd_rr2_p5.minus_ln2_hi[i] = -0x1.62E400p-1f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3394 params->wasmsimd_rr2_p5.minus_ln2_lo[i] = -0x1.7F7D1Cp-20f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3395 params->wasmsimd_rr2_p5.c5[i] = 0x1.0F9F9Cp-7f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3396 params->wasmsimd_rr2_p5.c4[i] = 0x1.573A1Ap-5f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3397 params->wasmsimd_rr2_p5.c3[i] = 0x1.555A80p-3f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3398 params->wasmsimd_rr2_p5.c2[i] = 0x1.FFFDC6p-2f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3399 params->wasmsimd_rr2_p5.c1[i] = 0x1.FFFFF6p-1f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3400 params->wasmsimd_rr2_p5.denorm_cutoff[i] = -0x1.5D589Ep6f; in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3402 return sizeof(params->wasmsimd_rr2_p5); in xnn_init_f32_expminus_wasmsimd_rr2_p5_params()
3408 union xnn_f16_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_lrelu_neon_params()
3411 params->neon.slope = slope; in xnn_init_f16_lrelu_neon_params()
3412 return sizeof(params->neon); in xnn_init_f16_lrelu_neon_params()
3418 union xnn_f16_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_lrelu_avx_params()
3422 params->avx.slope[i] = fp16_ieee_to_fp32_value(slope); in xnn_init_f16_lrelu_avx_params()
3424 return sizeof(params->avx); in xnn_init_f16_lrelu_avx_params()
3429 union xnn_f32_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_lrelu_scalar_params()
3432 params->scalar.slope = slope; in xnn_init_f32_lrelu_scalar_params()
3433 return sizeof(params->scalar); in xnn_init_f32_lrelu_scalar_params()
3438 union xnn_f32_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_lrelu_sse_params()
3442 params->sse.slope[i] = slope; in xnn_init_f32_lrelu_sse_params()
3444 return sizeof(params->sse); in xnn_init_f32_lrelu_sse_params()
3448 union xnn_f32_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_lrelu_avx_params()
3452 params->avx.slope[i] = slope; in xnn_init_f32_lrelu_avx_params()
3455 params->avx.mask_table[i] = -1; in xnn_init_f32_lrelu_avx_params()
3458 params->avx.mask_table[i] = 0; in xnn_init_f32_lrelu_avx_params()
3460 return sizeof(params->avx); in xnn_init_f32_lrelu_avx_params()
3466 union xnn_f32_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_lrelu_wasmsimd_params()
3469 params->wasmsimd.slope[0] = slope; in xnn_init_f32_lrelu_wasmsimd_params()
3470 params->wasmsimd.slope[1] = slope; in xnn_init_f32_lrelu_wasmsimd_params()
3471 return sizeof(params->wasmsimd); in xnn_init_f32_lrelu_wasmsimd_params()
3476 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_scalar_select_params()
3495 params->scalar_select.input_zero_point = (int32_t) input_zero_point; in xnn_init_qs8_lrelu_scalar_select_params()
3496 params->scalar_select.positive_multiplier = (int32_t) positive_multiplier; in xnn_init_qs8_lrelu_scalar_select_params()
3497 params->scalar_select.negative_multiplier = (int32_t) negative_multiplier; in xnn_init_qs8_lrelu_scalar_select_params()
3498 params->scalar_select.bias = ((int32_t) output_zero_point << 8) + INT32_C(0x80); in xnn_init_qs8_lrelu_scalar_select_params()
3499 return sizeof(params->scalar_select); in xnn_init_qs8_lrelu_scalar_select_params()
3503 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_scalar_andxor_params()
3522 params->scalar_andxor.input_zero_point = (int32_t) input_zero_point; in xnn_init_qs8_lrelu_scalar_andxor_params()
3523 params->scalar_andxor.multiplier_base = (int32_t) positive_multiplier; in xnn_init_qs8_lrelu_scalar_andxor_params()
3524params->scalar_andxor.multiplier_diff = (int32_t) negative_multiplier ^ (int32_t) positive_multipl… in xnn_init_qs8_lrelu_scalar_andxor_params()
3525 params->scalar_andxor.bias = ((int32_t) output_zero_point << 8) + INT32_C(0x80); in xnn_init_qs8_lrelu_scalar_andxor_params()
3526 return sizeof(params->scalar_andxor); in xnn_init_qs8_lrelu_scalar_andxor_params()
3531 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_armsimd32_params()
3550params->armsimd32.input_zero_point = (uint32_t) (uint16_t) (int16_t) input_zero_point * UINT32_C(0… in xnn_init_qs8_lrelu_armsimd32_params()
3551params->armsimd32.positive_multiplier = (uint32_t) (uint16_t) (int16_t) positive_multiplier * UINT… in xnn_init_qs8_lrelu_armsimd32_params()
3552params->armsimd32.negative_multiplier = (uint32_t) (uint16_t) (int16_t) negative_multiplier * UINT… in xnn_init_qs8_lrelu_armsimd32_params()
3553 params->armsimd32.bias = ((int32_t) output_zero_point << 8) + INT32_C(0x80); in xnn_init_qs8_lrelu_armsimd32_params()
3554 return sizeof(params->armsimd32); in xnn_init_qs8_lrelu_armsimd32_params()
3560 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_neon_params()
3579 params->neon.input_zero_point = (int16_t) input_zero_point; in xnn_init_qs8_lrelu_neon_params()
3580 params->neon.positive_multiplier = (int16_t) positive_multiplier; in xnn_init_qs8_lrelu_neon_params()
3581 params->neon.negative_multiplier = (int16_t) negative_multiplier; in xnn_init_qs8_lrelu_neon_params()
3582 params->neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_lrelu_neon_params()
3583 return sizeof(params->neon); in xnn_init_qs8_lrelu_neon_params()
3589 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_sse2_params()
3611 params->sse2.input_zero_point[i] = (int16_t) input_zero_point; in xnn_init_qs8_lrelu_sse2_params()
3612 params->sse2.multiplier_diff[i] = multiplier_diff; in xnn_init_qs8_lrelu_sse2_params()
3613 params->sse2.multiplier_base[i] = multiplier_base; in xnn_init_qs8_lrelu_sse2_params()
3614 params->sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_lrelu_sse2_params()
3616 return sizeof(params->sse2); in xnn_init_qs8_lrelu_sse2_params()
3620 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_avx_params()
3640 params->avx.input_zero_point[i] = (int16_t) input_zero_point; in xnn_init_qs8_lrelu_avx_params()
3641 params->avx.positive_multiplier[i] = (int16_t) positive_multiplier; in xnn_init_qs8_lrelu_avx_params()
3642 params->avx.negative_multiplier[i] = (int16_t) negative_multiplier; in xnn_init_qs8_lrelu_avx_params()
3643 params->avx.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_lrelu_avx_params()
3645 return sizeof(params->avx); in xnn_init_qs8_lrelu_avx_params()
3649 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_avx2_params()
3669 params->avx2.input_zero_point[i] = (int16_t) input_zero_point; in xnn_init_qs8_lrelu_avx2_params()
3670 params->avx2.positive_multiplier[i] = (int16_t) positive_multiplier; in xnn_init_qs8_lrelu_avx2_params()
3671 params->avx2.negative_multiplier[i] = (int16_t) negative_multiplier; in xnn_init_qs8_lrelu_avx2_params()
3672 params->avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_lrelu_avx2_params()
3674 return sizeof(params->avx2); in xnn_init_qs8_lrelu_avx2_params()
3680 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_wasmsimd_arm_params()
3700 params->wasmsimd_arm.input_zero_point[i] = (int16_t) input_zero_point; in xnn_init_qs8_lrelu_wasmsimd_arm_params()
3701 params->wasmsimd_arm.positive_multiplier[i] = (int16_t) positive_multiplier; in xnn_init_qs8_lrelu_wasmsimd_arm_params()
3702 params->wasmsimd_arm.negative_multiplier[i] = (int16_t) negative_multiplier; in xnn_init_qs8_lrelu_wasmsimd_arm_params()
3703 params->wasmsimd_arm.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_lrelu_wasmsimd_arm_params()
3705 return sizeof(params->wasmsimd_arm); in xnn_init_qs8_lrelu_wasmsimd_arm_params()
3709 union xnn_qs8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_lrelu_wasmsimd_x86_params()
3731 params->wasmsimd_x86.input_zero_point[i] = (int16_t) input_zero_point; in xnn_init_qs8_lrelu_wasmsimd_x86_params()
3732 params->wasmsimd_x86.multiplier_diff[i] = multiplier_diff; in xnn_init_qs8_lrelu_wasmsimd_x86_params()
3733 params->wasmsimd_x86.multiplier_base[i] = multiplier_base; in xnn_init_qs8_lrelu_wasmsimd_x86_params()
3734 params->wasmsimd_x86.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_lrelu_wasmsimd_x86_params()
3736 return sizeof(params->wasmsimd_x86); in xnn_init_qs8_lrelu_wasmsimd_x86_params()
3741 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_scalar_select_params()
3760 params->scalar_select.input_zero_point = (int32_t) input_zero_point; in xnn_init_qu8_lrelu_scalar_select_params()
3761 params->scalar_select.positive_multiplier = (int32_t) positive_multiplier; in xnn_init_qu8_lrelu_scalar_select_params()
3762 params->scalar_select.negative_multiplier = (int32_t) negative_multiplier; in xnn_init_qu8_lrelu_scalar_select_params()
3763 params->scalar_select.bias = ((int32_t) output_zero_point << 8) + INT32_C(0x80); in xnn_init_qu8_lrelu_scalar_select_params()
3764 return sizeof(params->scalar_select); in xnn_init_qu8_lrelu_scalar_select_params()
3768 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_scalar_andxor_params()
3787 params->scalar_andxor.input_zero_point = (int32_t) input_zero_point; in xnn_init_qu8_lrelu_scalar_andxor_params()
3788 params->scalar_andxor.multiplier_base = (int32_t) positive_multiplier; in xnn_init_qu8_lrelu_scalar_andxor_params()
3789params->scalar_andxor.multiplier_diff = (int32_t) negative_multiplier ^ (int32_t) positive_multipl… in xnn_init_qu8_lrelu_scalar_andxor_params()
3790 params->scalar_andxor.bias = ((int32_t) output_zero_point << 8) + INT32_C(0x80); in xnn_init_qu8_lrelu_scalar_andxor_params()
3791 return sizeof(params->scalar_andxor); in xnn_init_qu8_lrelu_scalar_andxor_params()
3796 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_armsimd32_params()
3815 params->armsimd32.input_zero_point = (uint32_t) input_zero_point * UINT32_C(0x00010001); in xnn_init_qu8_lrelu_armsimd32_params()
3816params->armsimd32.positive_multiplier = (uint32_t) (uint16_t) (int16_t) positive_multiplier * UINT… in xnn_init_qu8_lrelu_armsimd32_params()
3817params->armsimd32.negative_multiplier = (uint32_t) (uint16_t) (int16_t) negative_multiplier * UINT… in xnn_init_qu8_lrelu_armsimd32_params()
3818 params->armsimd32.bias = ((int32_t) output_zero_point << 8) + INT32_C(0x80); in xnn_init_qu8_lrelu_armsimd32_params()
3819 return sizeof(params->armsimd32); in xnn_init_qu8_lrelu_armsimd32_params()
3825 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_neon_params()
3844 params->neon.input_zero_point = (uint16_t) input_zero_point; in xnn_init_qu8_lrelu_neon_params()
3845 params->neon.positive_multiplier = (int16_t) positive_multiplier; in xnn_init_qu8_lrelu_neon_params()
3846 params->neon.negative_multiplier = (int16_t) negative_multiplier; in xnn_init_qu8_lrelu_neon_params()
3847 params->neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qu8_lrelu_neon_params()
3848 return sizeof(params->neon); in xnn_init_qu8_lrelu_neon_params()
3854 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_sse2_params()
3876 params->sse2.input_zero_point[i] = (int16_t) (uint16_t) input_zero_point; in xnn_init_qu8_lrelu_sse2_params()
3877 params->sse2.multiplier_diff[i] = multiplier_diff; in xnn_init_qu8_lrelu_sse2_params()
3878 params->sse2.multiplier_base[i] = multiplier_base; in xnn_init_qu8_lrelu_sse2_params()
3879 params->sse2.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_lrelu_sse2_params()
3881 return sizeof(params->sse2); in xnn_init_qu8_lrelu_sse2_params()
3885 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_avx_params()
3905 params->avx.input_zero_point[i] = (int16_t) (uint16_t) input_zero_point; in xnn_init_qu8_lrelu_avx_params()
3906 params->avx.positive_multiplier[i] = (int16_t) positive_multiplier; in xnn_init_qu8_lrelu_avx_params()
3907 params->avx.negative_multiplier[i] = (int16_t) negative_multiplier; in xnn_init_qu8_lrelu_avx_params()
3908 params->avx.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_lrelu_avx_params()
3910 return sizeof(params->avx); in xnn_init_qu8_lrelu_avx_params()
3914 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_avx2_params()
3934 params->avx2.input_zero_point[i] = (int16_t) (uint16_t) input_zero_point; in xnn_init_qu8_lrelu_avx2_params()
3935 params->avx2.positive_multiplier[i] = (int16_t) positive_multiplier; in xnn_init_qu8_lrelu_avx2_params()
3936 params->avx2.negative_multiplier[i] = (int16_t) negative_multiplier; in xnn_init_qu8_lrelu_avx2_params()
3937 params->avx2.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_lrelu_avx2_params()
3939 return sizeof(params->avx2); in xnn_init_qu8_lrelu_avx2_params()
3945 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_wasmsimd_arm_params()
3965 params->wasmsimd_arm.input_zero_point[i] = (int16_t) (uint16_t) input_zero_point; in xnn_init_qu8_lrelu_wasmsimd_arm_params()
3966 params->wasmsimd_arm.positive_multiplier[i] = (int16_t) positive_multiplier; in xnn_init_qu8_lrelu_wasmsimd_arm_params()
3967 params->wasmsimd_arm.negative_multiplier[i] = (int16_t) negative_multiplier; in xnn_init_qu8_lrelu_wasmsimd_arm_params()
3968 params->wasmsimd_arm.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_lrelu_wasmsimd_arm_params()
3970 return sizeof(params->wasmsimd_arm); in xnn_init_qu8_lrelu_wasmsimd_arm_params()
3974 union xnn_qu8_lrelu_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_lrelu_wasmsimd_x86_params()
3996 params->wasmsimd_x86.input_zero_point[i] = (int16_t) (uint16_t) input_zero_point; in xnn_init_qu8_lrelu_wasmsimd_x86_params()
3997 params->wasmsimd_x86.multiplier_diff[i] = multiplier_diff; in xnn_init_qu8_lrelu_wasmsimd_x86_params()
3998 params->wasmsimd_x86.multiplier_base[i] = multiplier_base; in xnn_init_qu8_lrelu_wasmsimd_x86_params()
3999 params->wasmsimd_x86.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_lrelu_wasmsimd_x86_params()
4001 return sizeof(params->wasmsimd_x86); in xnn_init_qu8_lrelu_wasmsimd_x86_params()
4007 union xnn_f32_sqrt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sqrt_avx_params()
4010 params->avx.mask_table[i] = -1; in xnn_init_f32_sqrt_avx_params()
4013 params->avx.mask_table[i] = 0; in xnn_init_f32_sqrt_avx_params()
4015 return sizeof(params->avx); in xnn_init_f32_sqrt_avx_params()
4019 union xnn_f32_sqrt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sqrt_fma_params()
4022 params->fma.half[i] = 0.5f; in xnn_init_f32_sqrt_fma_params()
4025 params->fma.mask_table[i] = -1; in xnn_init_f32_sqrt_fma_params()
4028 params->fma.mask_table[i] = 0; in xnn_init_f32_sqrt_fma_params()
4030 return sizeof(params->fma); in xnn_init_f32_sqrt_fma_params()
4034 union xnn_f32_sqrt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_sqrt_avx512_params()
4036 params->avx512.half = 0.5f; in xnn_init_f32_sqrt_avx512_params()
4037 return sizeof(params->avx512); in xnn_init_f32_sqrt_avx512_params()
4042 union xnn_f32_chw_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_chw_params()
4049 params->sse.min[i] = output_min; in xnn_init_f32_chw_params()
4050 params->sse.max[i] = output_max; in xnn_init_f32_chw_params()
4054 params->sse.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_chw_params()
4055 params->sse.mask[1] = -(uint32_t) (w4 >= 1); in xnn_init_f32_chw_params()
4056 params->sse.mask[2] = -(uint32_t) (w4 >= 2); in xnn_init_f32_chw_params()
4057 params->sse.mask[3] = -(uint32_t) (w4 >= 3); in xnn_init_f32_chw_params()
4060 params->sse.mask_even[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_chw_params()
4061 params->sse.mask_even[1] = -(uint32_t) (w8 >= 2); in xnn_init_f32_chw_params()
4062 params->sse.mask_even[2] = -(uint32_t) (w8 >= 4); in xnn_init_f32_chw_params()
4063 params->sse.mask_even[3] = -(uint32_t) (w8 >= 6); in xnn_init_f32_chw_params()
4064 params->sse.mask_odd[0] = -(uint32_t) (w8 >= 1); in xnn_init_f32_chw_params()
4065 params->sse.mask_odd[1] = -(uint32_t) (w8 >= 3); in xnn_init_f32_chw_params()
4066 params->sse.mask_odd[2] = -(uint32_t) (w8 >= 5); in xnn_init_f32_chw_params()
4067 params->sse.mask_odd[3] = -(uint32_t) (w8 >= 7); in xnn_init_f32_chw_params()
4068 return sizeof(params->sse); in xnn_init_f32_chw_params()
4070 params->neon.min = output_min; in xnn_init_f32_chw_params()
4071 params->neon.max = output_max; in xnn_init_f32_chw_params()
4074 params->neon.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_chw_params()
4075 params->neon.mask[1] = -(uint32_t) (w4 >= 1); in xnn_init_f32_chw_params()
4076 params->neon.mask[2] = -(uint32_t) (w4 >= 2); in xnn_init_f32_chw_params()
4077 params->neon.mask[3] = -(uint32_t) (w4 >= 3); in xnn_init_f32_chw_params()
4080 params->neon.mask_even[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_chw_params()
4081 params->neon.mask_even[1] = -(uint32_t) (w8 >= 2); in xnn_init_f32_chw_params()
4082 params->neon.mask_even[2] = -(uint32_t) (w8 >= 4); in xnn_init_f32_chw_params()
4083 params->neon.mask_even[3] = -(uint32_t) (w8 >= 6); in xnn_init_f32_chw_params()
4084 params->neon.mask_odd[0] = -(uint32_t) (w8 >= 1); in xnn_init_f32_chw_params()
4085 params->neon.mask_odd[1] = -(uint32_t) (w8 >= 3); in xnn_init_f32_chw_params()
4086 params->neon.mask_odd[2] = -(uint32_t) (w8 >= 5); in xnn_init_f32_chw_params()
4087 params->neon.mask_odd[3] = -(uint32_t) (w8 >= 7); in xnn_init_f32_chw_params()
4088 return sizeof(params->neon); in xnn_init_f32_chw_params()
4090 params->scalar.min = output_min; in xnn_init_f32_chw_params()
4091 params->scalar.max = output_max; in xnn_init_f32_chw_params()
4094 params->scalar.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_chw_params()
4095 params->scalar.mask[1] = -(uint32_t) (w4 >= 1); in xnn_init_f32_chw_params()
4096 params->scalar.mask[2] = -(uint32_t) (w4 >= 2); in xnn_init_f32_chw_params()
4097 params->scalar.mask[3] = -(uint32_t) (w4 >= 3); in xnn_init_f32_chw_params()
4100 params->scalar.mask_even[0] = UINT32_C(0xFFFFFFFF); in xnn_init_f32_chw_params()
4101 params->scalar.mask_even[1] = -(uint32_t) (w8 >= 2); in xnn_init_f32_chw_params()
4102 params->scalar.mask_even[2] = -(uint32_t) (w8 >= 4); in xnn_init_f32_chw_params()
4103 params->scalar.mask_even[3] = -(uint32_t) (w8 >= 6); in xnn_init_f32_chw_params()
4104 params->scalar.mask_odd[0] = -(uint32_t) (w8 >= 1); in xnn_init_f32_chw_params()
4105 params->scalar.mask_odd[1] = -(uint32_t) (w8 >= 3); in xnn_init_f32_chw_params()
4106 params->scalar.mask_odd[2] = -(uint32_t) (w8 >= 5); in xnn_init_f32_chw_params()
4107 params->scalar.mask_odd[3] = -(uint32_t) (w8 >= 7); in xnn_init_f32_chw_params()
4108 return sizeof(params->scalar); in xnn_init_f32_chw_params()
4113 union xnn_f16_chw_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f16_chw_params()
4119 params->neonfp16arith.min = output_min; in xnn_init_f16_chw_params()
4120 params->neonfp16arith.max = output_max; in xnn_init_f16_chw_params()
4123 params->neonfp16arith.mask[0] = UINT16_C(0xFFFF); in xnn_init_f16_chw_params()
4124 params->neonfp16arith.mask[1] = -(uint16_t) (w4 >= 1); in xnn_init_f16_chw_params()
4125 params->neonfp16arith.mask[2] = -(uint16_t) (w4 >= 2); in xnn_init_f16_chw_params()
4126 params->neonfp16arith.mask[3] = -(uint16_t) (w4 >= 3); in xnn_init_f16_chw_params()
4129 params->neonfp16arith.maskx8[0] = UINT16_C(0xFFFF); in xnn_init_f16_chw_params()
4130 params->neonfp16arith.maskx8[1] = -(uint16_t) (w8 >= 1); in xnn_init_f16_chw_params()
4131 params->neonfp16arith.maskx8[2] = -(uint16_t) (w8 >= 2); in xnn_init_f16_chw_params()
4132 params->neonfp16arith.maskx8[3] = -(uint16_t) (w8 >= 3); in xnn_init_f16_chw_params()
4133 params->neonfp16arith.maskx8[4] = -(uint16_t) (w8 >= 4); in xnn_init_f16_chw_params()
4134 params->neonfp16arith.maskx8[5] = -(uint16_t) (w8 >= 5); in xnn_init_f16_chw_params()
4135 params->neonfp16arith.maskx8[6] = -(uint16_t) (w8 >= 6); in xnn_init_f16_chw_params()
4136 params->neonfp16arith.maskx8[7] = -(uint16_t) (w8 >= 7); in xnn_init_f16_chw_params()
4138 params->neonfp16arith.mask_even[0] = UINT16_C(0xFFFF); in xnn_init_f16_chw_params()
4139 params->neonfp16arith.mask_even[1] = -(uint16_t) (w8 >= 2); in xnn_init_f16_chw_params()
4140 params->neonfp16arith.mask_even[2] = -(uint16_t) (w8 >= 4); in xnn_init_f16_chw_params()
4141 params->neonfp16arith.mask_even[3] = -(uint16_t) (w8 >= 6); in xnn_init_f16_chw_params()
4142 params->neonfp16arith.mask_odd[0] = -(uint16_t) (w8 >= 1); in xnn_init_f16_chw_params()
4143 params->neonfp16arith.mask_odd[1] = -(uint16_t) (w8 >= 3); in xnn_init_f16_chw_params()
4144 params->neonfp16arith.mask_odd[2] = -(uint16_t) (w8 >= 5); in xnn_init_f16_chw_params()
4145 params->neonfp16arith.mask_odd[3] = -(uint16_t) (w8 >= 7); in xnn_init_f16_chw_params()
4146 return sizeof(params->neonfp16arith); in xnn_init_f16_chw_params()
4153 union xnn_f32_chw_params* params, in xnn_update_f32_chw_params() argument
4158 params->sse.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_chw_params()
4159 params->sse.mask[1] = -(uint32_t) (w4 >= 1); in xnn_update_f32_chw_params()
4160 params->sse.mask[2] = -(uint32_t) (w4 >= 2); in xnn_update_f32_chw_params()
4161 params->sse.mask[3] = -(uint32_t) (w4 >= 3); in xnn_update_f32_chw_params()
4164 params->sse.mask_even[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_chw_params()
4165 params->sse.mask_even[1] = -(uint32_t) (w8 >= 2); in xnn_update_f32_chw_params()
4166 params->sse.mask_even[2] = -(uint32_t) (w8 >= 4); in xnn_update_f32_chw_params()
4167 params->sse.mask_even[3] = -(uint32_t) (w8 >= 6); in xnn_update_f32_chw_params()
4168 params->sse.mask_odd[0] = -(uint32_t) (w8 >= 1); in xnn_update_f32_chw_params()
4169 params->sse.mask_odd[1] = -(uint32_t) (w8 >= 3); in xnn_update_f32_chw_params()
4170 params->sse.mask_odd[2] = -(uint32_t) (w8 >= 5); in xnn_update_f32_chw_params()
4171 params->sse.mask_odd[3] = -(uint32_t) (w8 >= 7); in xnn_update_f32_chw_params()
4174 params->neon.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_chw_params()
4175 params->neon.mask[1] = -(uint32_t) (w4 >= 1); in xnn_update_f32_chw_params()
4176 params->neon.mask[2] = -(uint32_t) (w4 >= 2); in xnn_update_f32_chw_params()
4177 params->neon.mask[3] = -(uint32_t) (w4 >= 3); in xnn_update_f32_chw_params()
4180 params->neon.mask_even[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_chw_params()
4181 params->neon.mask_even[1] = -(uint32_t) (w8 >= 2); in xnn_update_f32_chw_params()
4182 params->neon.mask_even[2] = -(uint32_t) (w8 >= 4); in xnn_update_f32_chw_params()
4183 params->neon.mask_even[3] = -(uint32_t) (w8 >= 6); in xnn_update_f32_chw_params()
4184 params->neon.mask_odd[0] = -(uint32_t) (w8 >= 1); in xnn_update_f32_chw_params()
4185 params->neon.mask_odd[1] = -(uint32_t) (w8 >= 3); in xnn_update_f32_chw_params()
4186 params->neon.mask_odd[2] = -(uint32_t) (w8 >= 5); in xnn_update_f32_chw_params()
4187 params->neon.mask_odd[3] = -(uint32_t) (w8 >= 7); in xnn_update_f32_chw_params()
4190 params->scalar.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_chw_params()
4191 params->scalar.mask[1] = -(uint32_t) (w4 >= 1); in xnn_update_f32_chw_params()
4192 params->scalar.mask[2] = -(uint32_t) (w4 >= 2); in xnn_update_f32_chw_params()
4193 params->scalar.mask[3] = -(uint32_t) (w4 >= 3); in xnn_update_f32_chw_params()
4196 params->scalar.mask_even[0] = UINT32_C(0xFFFFFFFF); in xnn_update_f32_chw_params()
4197 params->scalar.mask_even[1] = -(uint32_t) (w8 >= 2); in xnn_update_f32_chw_params()
4198 params->scalar.mask_even[2] = -(uint32_t) (w8 >= 4); in xnn_update_f32_chw_params()
4199 params->scalar.mask_even[3] = -(uint32_t) (w8 >= 6); in xnn_update_f32_chw_params()
4200 params->scalar.mask_odd[0] = -(uint32_t) (w8 >= 1); in xnn_update_f32_chw_params()
4201 params->scalar.mask_odd[1] = -(uint32_t) (w8 >= 3); in xnn_update_f32_chw_params()
4202 params->scalar.mask_odd[2] = -(uint32_t) (w8 >= 5); in xnn_update_f32_chw_params()
4203 params->scalar.mask_odd[3] = -(uint32_t) (w8 >= 7); in xnn_update_f32_chw_params()
4208 union xnn_f32_chw_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_scalar_f32_chw_params()
4213 params->scalar.min = output_min; in xnn_init_scalar_f32_chw_params()
4214 params->scalar.max = output_max; in xnn_init_scalar_f32_chw_params()
4217 params->scalar.mask[0] = UINT32_C(0xFFFFFFFF); in xnn_init_scalar_f32_chw_params()
4218 params->scalar.mask[1] = -(uint32_t) (w4 >= 1); in xnn_init_scalar_f32_chw_params()
4219 params->scalar.mask[2] = -(uint32_t) (w4 >= 2); in xnn_init_scalar_f32_chw_params()
4220 params->scalar.mask[3] = -(uint32_t) (w4 >= 3); in xnn_init_scalar_f32_chw_params()
4223 params->scalar.mask_even[0] = UINT32_C(0xFFFFFFFF); in xnn_init_scalar_f32_chw_params()
4224 params->scalar.mask_even[1] = -(uint32_t) (w8 >= 2); in xnn_init_scalar_f32_chw_params()
4225 params->scalar.mask_even[2] = -(uint32_t) (w8 >= 4); in xnn_init_scalar_f32_chw_params()
4226 params->scalar.mask_even[3] = -(uint32_t) (w8 >= 6); in xnn_init_scalar_f32_chw_params()
4227 params->scalar.mask_odd[0] = -(uint32_t) (w8 >= 1); in xnn_init_scalar_f32_chw_params()
4228 params->scalar.mask_odd[1] = -(uint32_t) (w8 >= 3); in xnn_init_scalar_f32_chw_params()
4229 params->scalar.mask_odd[2] = -(uint32_t) (w8 >= 5); in xnn_init_scalar_f32_chw_params()
4230 params->scalar.mask_odd[3] = -(uint32_t) (w8 >= 7); in xnn_init_scalar_f32_chw_params()
4231 return sizeof(params->scalar); in xnn_init_scalar_f32_chw_params()
4236 union xnn_s8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_s8_minmax_sse2_params()
4245 params->sse2.bias[i] = UINT8_C(0x80); in xnn_init_s8_minmax_sse2_params()
4246 params->sse2.min_with_bias[i] = output_min_with_bias; in xnn_init_s8_minmax_sse2_params()
4247 params->sse2.max_with_bias[i] = output_max_with_bias; in xnn_init_s8_minmax_sse2_params()
4249 return sizeof(params->sse2); in xnn_init_s8_minmax_sse2_params()
4253 union xnn_s8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_s8_minmax_sse4_params()
4260 params->sse4.min[i] = output_min; in xnn_init_s8_minmax_sse4_params()
4261 params->sse4.max[i] = output_max; in xnn_init_s8_minmax_sse4_params()
4263 return sizeof(params->sse4); in xnn_init_s8_minmax_sse4_params()
4269 union xnn_s8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_s8_minmax_neon_params()
4275 params->neon.min = output_min; in xnn_init_s8_minmax_neon_params()
4276 params->neon.max = output_max; in xnn_init_s8_minmax_neon_params()
4277 return sizeof(params->neon); in xnn_init_s8_minmax_neon_params()
4283 union xnn_s8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_s8_minmax_wasmsimd_params()
4290 params->wasmsimd.min[i] = output_min; in xnn_init_s8_minmax_wasmsimd_params()
4291 params->wasmsimd.max[i] = output_max; in xnn_init_s8_minmax_wasmsimd_params()
4293 return sizeof(params->wasmsimd); in xnn_init_s8_minmax_wasmsimd_params()
4298 union xnn_s8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_s8_minmax_scalar_params()
4304 params->scalar.min = (int32_t) output_min; in xnn_init_s8_minmax_scalar_params()
4305 params->scalar.max = (int32_t) output_max; in xnn_init_s8_minmax_scalar_params()
4306 return sizeof(params->scalar); in xnn_init_s8_minmax_scalar_params()
4310 union xnn_u8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_u8_minmax_params()
4318 params->sse2.min[i] = output_min; in xnn_init_u8_minmax_params()
4319 params->sse2.max[i] = output_max; in xnn_init_u8_minmax_params()
4321 return sizeof(params->sse2); in xnn_init_u8_minmax_params()
4323 params->neon.min = output_min; in xnn_init_u8_minmax_params()
4324 params->neon.max = output_max; in xnn_init_u8_minmax_params()
4325 return sizeof(params->neon); in xnn_init_u8_minmax_params()
4327 params->scalar.min = (uint32_t) output_min; in xnn_init_u8_minmax_params()
4328 params->scalar.max = (uint32_t) output_max; in xnn_init_u8_minmax_params()
4329 return sizeof(params->scalar); in xnn_init_u8_minmax_params()
4335 union xnn_u8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_u8_minmax_sse2_params()
4342 params->sse2.min[i] = output_min; in xnn_init_u8_minmax_sse2_params()
4343 params->sse2.max[i] = output_max; in xnn_init_u8_minmax_sse2_params()
4345 return sizeof(params->sse2); in xnn_init_u8_minmax_sse2_params()
4351 union xnn_u8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_u8_minmax_wasmsimd_params()
4358 params->wasmsimd.min[i] = output_min; in xnn_init_u8_minmax_wasmsimd_params()
4359 params->wasmsimd.max[i] = output_max; in xnn_init_u8_minmax_wasmsimd_params()
4361 return sizeof(params->wasmsimd); in xnn_init_u8_minmax_wasmsimd_params()
4367 union xnn_u8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_u8_minmax_neon_params()
4373 params->neon.min = output_min; in xnn_init_u8_minmax_neon_params()
4374 params->neon.max = output_max; in xnn_init_u8_minmax_neon_params()
4375 return sizeof(params->neon); in xnn_init_u8_minmax_neon_params()
4380 union xnn_u8_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_u8_minmax_scalar_params()
4386 params->scalar.min = (uint32_t) output_min; in xnn_init_u8_minmax_scalar_params()
4387 params->scalar.max = (uint32_t) output_max; in xnn_init_u8_minmax_scalar_params()
4388 return sizeof(params->scalar); in xnn_init_u8_minmax_scalar_params()
4393 union xnn_qu8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_add_minmax_sse2_params()
4434 params->sse2.bias[i] = bias; in xnn_init_qu8_add_minmax_sse2_params()
4441 params->sse2.a_multiplier_lo[i] = a_multiplier_lo; in xnn_init_qu8_add_minmax_sse2_params()
4442 params->sse2.a_multiplier_hi[i] = a_multiplier_hi; in xnn_init_qu8_add_minmax_sse2_params()
4443 params->sse2.b_multiplier_lo[i] = b_multiplier_lo; in xnn_init_qu8_add_minmax_sse2_params()
4444 params->sse2.b_multiplier_hi[i] = b_multiplier_hi; in xnn_init_qu8_add_minmax_sse2_params()
4446 params->sse2.shift = shift; in xnn_init_qu8_add_minmax_sse2_params()
4447 params->sse2.b_multiplier = (uint32_t) b_multiplier; in xnn_init_qu8_add_minmax_sse2_params()
4449 params->sse2.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_add_minmax_sse2_params()
4452 params->sse2.output_min[i] = output_min; in xnn_init_qu8_add_minmax_sse2_params()
4453 params->sse2.output_max[i] = output_max; in xnn_init_qu8_add_minmax_sse2_params()
4455 return sizeof(params->sse2); in xnn_init_qu8_add_minmax_sse2_params()
4459 union xnn_qu8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_add_minmax_sse4_params()
4500 params->sse4.bias[i] = bias; in xnn_init_qu8_add_minmax_sse4_params()
4501 params->sse4.a_multiplier[i] = a_multiplier; in xnn_init_qu8_add_minmax_sse4_params()
4502 params->sse4.b_multiplier[i] = b_multiplier; in xnn_init_qu8_add_minmax_sse4_params()
4505 params->sse4.shift[i] = (uint64_t) shift; in xnn_init_qu8_add_minmax_sse4_params()
4508 params->sse4.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_add_minmax_sse4_params()
4511 params->sse4.output_min[i] = output_min; in xnn_init_qu8_add_minmax_sse4_params()
4512 params->sse4.output_max[i] = output_max; in xnn_init_qu8_add_minmax_sse4_params()
4514 return sizeof(params->sse4); in xnn_init_qu8_add_minmax_sse4_params()
4518 union xnn_qu8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_add_minmax_avx2_params()
4559 params->avx2.bias[i] = bias; in xnn_init_qu8_add_minmax_avx2_params()
4560 params->avx2.a_multiplier[i] = a_multiplier; in xnn_init_qu8_add_minmax_avx2_params()
4561 params->avx2.b_multiplier[i] = b_multiplier; in xnn_init_qu8_add_minmax_avx2_params()
4564 params->avx2.shift[i] = (uint64_t) shift; in xnn_init_qu8_add_minmax_avx2_params()
4567 params->avx2.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_add_minmax_avx2_params()
4568 params->avx2.output_min[i] = output_min; in xnn_init_qu8_add_minmax_avx2_params()
4569 params->avx2.output_max[i] = output_max; in xnn_init_qu8_add_minmax_avx2_params()
4571 return sizeof(params->avx2); in xnn_init_qu8_add_minmax_avx2_params()
4575 union xnn_qu8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_add_minmax_avx512_params()
4616 params->avx512.bias[i] = bias; in xnn_init_qu8_add_minmax_avx512_params()
4617 params->avx512.a_multiplier[i] = a_multiplier; in xnn_init_qu8_add_minmax_avx512_params()
4618 params->avx512.b_multiplier[i] = b_multiplier; in xnn_init_qu8_add_minmax_avx512_params()
4621 params->avx512.shift[i] = (uint64_t) shift; in xnn_init_qu8_add_minmax_avx512_params()
4624 params->avx512.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_add_minmax_avx512_params()
4625 params->avx512.output_min[i] = output_min; in xnn_init_qu8_add_minmax_avx512_params()
4626 params->avx512.output_max[i] = output_max; in xnn_init_qu8_add_minmax_avx512_params()
4628 return sizeof(params->avx512); in xnn_init_qu8_add_minmax_avx512_params()
4634 union xnn_qu8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_add_minmax_neon_params()
4672 params->neon.a_zero_point = a_zero_point; in xnn_init_qu8_add_minmax_neon_params()
4673 params->neon.b_zero_point = b_zero_point; in xnn_init_qu8_add_minmax_neon_params()
4674 params->neon.a_multiplier = (int32_t) a_multiplier; in xnn_init_qu8_add_minmax_neon_params()
4675 params->neon.b_multiplier = (int32_t) b_multiplier; in xnn_init_qu8_add_minmax_neon_params()
4676 params->neon.right_shift = (int32_t) -shift; in xnn_init_qu8_add_minmax_neon_params()
4677 params->neon.output_zero_point = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_add_minmax_neon_params()
4678 params->neon.output_min = output_min; in xnn_init_qu8_add_minmax_neon_params()
4679 params->neon.output_max = output_max; in xnn_init_qu8_add_minmax_neon_params()
4680 return sizeof(params->neon); in xnn_init_qu8_add_minmax_neon_params()
4686 union xnn_qu8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_add_minmax_wasmsimd_params()
4727 params->wasmsimd.bias[i] = bias; in xnn_init_qu8_add_minmax_wasmsimd_params()
4728 params->wasmsimd.a_multiplier[i] = a_multiplier; in xnn_init_qu8_add_minmax_wasmsimd_params()
4729 params->wasmsimd.b_multiplier[i] = b_multiplier; in xnn_init_qu8_add_minmax_wasmsimd_params()
4731 params->wasmsimd.shift = shift; in xnn_init_qu8_add_minmax_wasmsimd_params()
4733 params->wasmsimd.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_add_minmax_wasmsimd_params()
4736 params->wasmsimd.output_min[i] = output_min; in xnn_init_qu8_add_minmax_wasmsimd_params()
4737 params->wasmsimd.output_max[i] = output_max; in xnn_init_qu8_add_minmax_wasmsimd_params()
4739 return sizeof(params->wasmsimd); in xnn_init_qu8_add_minmax_wasmsimd_params()
4744 union xnn_qu8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_add_minmax_scalar_params()
4783params->scalar.bias = rounding - a_multiplier * (int32_t) (uint32_t) a_zero_point - b_multiplier *… in xnn_init_qu8_add_minmax_scalar_params()
4784 params->scalar.a_multiplier = a_multiplier; in xnn_init_qu8_add_minmax_scalar_params()
4785 params->scalar.b_multiplier = b_multiplier; in xnn_init_qu8_add_minmax_scalar_params()
4786 params->scalar.shift = shift; in xnn_init_qu8_add_minmax_scalar_params()
4787params->scalar.output_min_less_zero_point = (int32_t) (uint32_t) output_min - (int32_t) (uint32_t)… in xnn_init_qu8_add_minmax_scalar_params()
4788params->scalar.output_max_less_zero_point = (int32_t) (uint32_t) output_max - (int32_t) (uint32_t)… in xnn_init_qu8_add_minmax_scalar_params()
4789 params->scalar.output_zero_point = (int32_t) (uint32_t) output_zero_point; in xnn_init_qu8_add_minmax_scalar_params()
4790 return sizeof(params->scalar); in xnn_init_qu8_add_minmax_scalar_params()
4795 union xnn_qs8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_add_minmax_sse2_params()
4836 params->sse2.bias[i] = bias; in xnn_init_qs8_add_minmax_sse2_params()
4843 params->sse2.a_multiplier_lo[i] = a_multiplier_lo; in xnn_init_qs8_add_minmax_sse2_params()
4844 params->sse2.a_multiplier_hi[i] = a_multiplier_hi; in xnn_init_qs8_add_minmax_sse2_params()
4845 params->sse2.b_multiplier_lo[i] = b_multiplier_lo; in xnn_init_qs8_add_minmax_sse2_params()
4846 params->sse2.b_multiplier_hi[i] = b_multiplier_hi; in xnn_init_qs8_add_minmax_sse2_params()
4848 params->sse2.shift = shift; in xnn_init_qs8_add_minmax_sse2_params()
4849 params->sse2.b_multiplier = (uint32_t) b_multiplier; in xnn_init_qs8_add_minmax_sse2_params()
4851 params->sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_add_minmax_sse2_params()
4852 params->sse2.output_min[i] = (int16_t) output_min; in xnn_init_qs8_add_minmax_sse2_params()
4853 params->sse2.output_max[i] = (int16_t) output_max; in xnn_init_qs8_add_minmax_sse2_params()
4855 return sizeof(params->sse2); in xnn_init_qs8_add_minmax_sse2_params()
4859 union xnn_qs8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_add_minmax_sse4_mul16_params()
4900 params->sse4_mul16.bias[i] = bias; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4907 params->sse4_mul16.a_multiplier_lo[i] = a_multiplier_lo; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4908 params->sse4_mul16.a_multiplier_hi[i] = a_multiplier_hi; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4909 params->sse4_mul16.b_multiplier_lo[i] = b_multiplier_lo; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4910 params->sse4_mul16.b_multiplier_hi[i] = b_multiplier_hi; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4912 params->sse4_mul16.shift = shift; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4913 params->sse4_mul16.b_multiplier = (uint32_t) b_multiplier; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4915 params->sse4_mul16.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4918 params->sse4_mul16.output_min[i] = output_min; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4919 params->sse4_mul16.output_max[i] = output_max; in xnn_init_qs8_add_minmax_sse4_mul16_params()
4921 return sizeof(params->sse4_mul16); in xnn_init_qs8_add_minmax_sse4_mul16_params()
4925 union xnn_qs8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_add_minmax_sse4_mul32_params()
4966 params->sse4_mul32.bias[i] = bias; in xnn_init_qs8_add_minmax_sse4_mul32_params()
4967 params->sse4_mul32.a_multiplier[i] = a_multiplier; in xnn_init_qs8_add_minmax_sse4_mul32_params()
4968 params->sse4_mul32.b_multiplier[i] = b_multiplier; in xnn_init_qs8_add_minmax_sse4_mul32_params()
4971 params->sse4_mul32.shift[i] = (uint64_t) shift; in xnn_init_qs8_add_minmax_sse4_mul32_params()
4974 params->sse4_mul32.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_add_minmax_sse4_mul32_params()
4977 params->sse4_mul32.output_min[i] = output_min; in xnn_init_qs8_add_minmax_sse4_mul32_params()
4978 params->sse4_mul32.output_max[i] = output_max; in xnn_init_qs8_add_minmax_sse4_mul32_params()
4980 return sizeof(params->sse4_mul32); in xnn_init_qs8_add_minmax_sse4_mul32_params()
4984 union xnn_qs8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_add_minmax_avx2_params()
5025 params->avx2.bias[i] = bias; in xnn_init_qs8_add_minmax_avx2_params()
5026 params->avx2.a_multiplier[i] = a_multiplier; in xnn_init_qs8_add_minmax_avx2_params()
5027 params->avx2.b_multiplier[i] = b_multiplier; in xnn_init_qs8_add_minmax_avx2_params()
5030 params->avx2.shift[i] = (uint64_t) shift; in xnn_init_qs8_add_minmax_avx2_params()
5033 params->avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_add_minmax_avx2_params()
5034 params->avx2.output_min[i] = output_min; in xnn_init_qs8_add_minmax_avx2_params()
5035 params->avx2.output_max[i] = output_max; in xnn_init_qs8_add_minmax_avx2_params()
5037 return sizeof(params->avx2); in xnn_init_qs8_add_minmax_avx2_params()
5041 union xnn_qs8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_add_minmax_avx512_params()
5082 params->avx512.bias[i] = bias; in xnn_init_qs8_add_minmax_avx512_params()
5083 params->avx512.a_multiplier[i] = a_multiplier; in xnn_init_qs8_add_minmax_avx512_params()
5084 params->avx512.b_multiplier[i] = b_multiplier; in xnn_init_qs8_add_minmax_avx512_params()
5087 params->avx512.shift[i] = (uint64_t) shift; in xnn_init_qs8_add_minmax_avx512_params()
5090 params->avx512.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_add_minmax_avx512_params()
5091 params->avx512.output_min[i] = output_min; in xnn_init_qs8_add_minmax_avx512_params()
5092 params->avx512.output_max[i] = output_max; in xnn_init_qs8_add_minmax_avx512_params()
5094 return sizeof(params->avx512); in xnn_init_qs8_add_minmax_avx512_params()
5100 union xnn_qs8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_add_minmax_neon_params()
5138 params->neon.a_zero_point = a_zero_point; in xnn_init_qs8_add_minmax_neon_params()
5139 params->neon.b_zero_point = b_zero_point; in xnn_init_qs8_add_minmax_neon_params()
5140 params->neon.a_multiplier = (int32_t) a_multiplier; in xnn_init_qs8_add_minmax_neon_params()
5141 params->neon.b_multiplier = (int32_t) b_multiplier; in xnn_init_qs8_add_minmax_neon_params()
5142 params->neon.right_shift = (int32_t) -shift; in xnn_init_qs8_add_minmax_neon_params()
5143 params->neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_add_minmax_neon_params()
5144 params->neon.output_min = output_min; in xnn_init_qs8_add_minmax_neon_params()
5145 params->neon.output_max = output_max; in xnn_init_qs8_add_minmax_neon_params()
5146 return sizeof(params->neon); in xnn_init_qs8_add_minmax_neon_params()
5152 union xnn_qs8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_add_minmax_wasmsimd_params()
5193 params->wasmsimd.bias[i] = bias; in xnn_init_qs8_add_minmax_wasmsimd_params()
5194 params->wasmsimd.a_multiplier[i] = a_multiplier; in xnn_init_qs8_add_minmax_wasmsimd_params()
5195 params->wasmsimd.b_multiplier[i] = b_multiplier; in xnn_init_qs8_add_minmax_wasmsimd_params()
5197 params->wasmsimd.shift = shift; in xnn_init_qs8_add_minmax_wasmsimd_params()
5199 params->wasmsimd.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_add_minmax_wasmsimd_params()
5202 params->wasmsimd.output_min[i] = output_min; in xnn_init_qs8_add_minmax_wasmsimd_params()
5203 params->wasmsimd.output_max[i] = output_max; in xnn_init_qs8_add_minmax_wasmsimd_params()
5205 return sizeof(params->wasmsimd); in xnn_init_qs8_add_minmax_wasmsimd_params()
5210 union xnn_qs8_add_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_add_minmax_scalar_params()
5249params->scalar.bias = rounding - a_multiplier * (int32_t) a_zero_point - b_multiplier * (int32_t) … in xnn_init_qs8_add_minmax_scalar_params()
5250 params->scalar.a_multiplier = a_multiplier; in xnn_init_qs8_add_minmax_scalar_params()
5251 params->scalar.b_multiplier = b_multiplier; in xnn_init_qs8_add_minmax_scalar_params()
5252 params->scalar.shift = shift; in xnn_init_qs8_add_minmax_scalar_params()
5253 params->scalar.output_min_less_zero_point = (int32_t) output_min - (int32_t) output_zero_point; in xnn_init_qs8_add_minmax_scalar_params()
5254 params->scalar.output_max_less_zero_point = (int32_t) output_max - (int32_t) output_zero_point; in xnn_init_qs8_add_minmax_scalar_params()
5255 params->scalar.output_zero_point = (int32_t) output_zero_point; in xnn_init_qs8_add_minmax_scalar_params()
5256 return sizeof(params->scalar); in xnn_init_qs8_add_minmax_scalar_params()
5260 union xnn_qu8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5271 params->fp32_scalar.a_zero_point = (int16_t) (uint16_t) a_zero_point; in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5272 params->fp32_scalar.b_zero_point = (int16_t) (uint16_t) b_zero_point; in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5273 params->fp32_scalar.scale = product_output_scale; in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5274params->fp32_scalar.output_min_less_zero_point = (float) (int32_t) ((uint32_t) output_min - (uint3… in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5275params->fp32_scalar.output_max_less_zero_point = (float) (int32_t) ((uint32_t) output_max - (uint3… in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5276 params->fp32_scalar.magic_bias = 12582912.0f; in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5277params->fp32_scalar.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) (uint32_t)… in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5278 return sizeof(params->fp32_scalar); in xnn_init_qu8_mul_minmax_fp32_scalar_params()
5283 union xnn_qu8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_mul_minmax_fp32_neon_params()
5294 params->fp32_neon.a_zero_point[0] = a_zero_point; in xnn_init_qu8_mul_minmax_fp32_neon_params()
5295 params->fp32_neon.a_zero_point[1] = a_zero_point; in xnn_init_qu8_mul_minmax_fp32_neon_params()
5296 params->fp32_neon.b_zero_point[0] = b_zero_point; in xnn_init_qu8_mul_minmax_fp32_neon_params()
5297 params->fp32_neon.b_zero_point[1] = b_zero_point; in xnn_init_qu8_mul_minmax_fp32_neon_params()
5298 params->fp32_neon.scale = product_output_scale; in xnn_init_qu8_mul_minmax_fp32_neon_params()
5299 params->fp32_neon.magic_bias = 12582912.0f; in xnn_init_qu8_mul_minmax_fp32_neon_params()
5300params->fp32_neon.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_… in xnn_init_qu8_mul_minmax_fp32_neon_params()
5301 params->fp32_neon.output_min = output_min; in xnn_init_qu8_mul_minmax_fp32_neon_params()
5302 params->fp32_neon.output_max = output_max; in xnn_init_qu8_mul_minmax_fp32_neon_params()
5303 return sizeof(params->fp32_neon); in xnn_init_qu8_mul_minmax_fp32_neon_params()
5307 union xnn_qu8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5318 params->fp32_neonv8.a_zero_point[0] = a_zero_point; in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5319 params->fp32_neonv8.a_zero_point[1] = a_zero_point; in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5320 params->fp32_neonv8.b_zero_point[0] = b_zero_point; in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5321 params->fp32_neonv8.b_zero_point[1] = b_zero_point; in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5322 params->fp32_neonv8.scale = product_output_scale; in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5323 params->fp32_neonv8.output_zero_point = (int16_t) output_zero_point; in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5324 params->fp32_neonv8.output_min = output_min; in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5325 params->fp32_neonv8.output_max = output_max; in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5326 return sizeof(params->fp32_neonv8); in xnn_init_qu8_mul_minmax_fp32_neonv8_params()
5330 union xnn_qu8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5358 params->rndnu_neon.a_zero_point[0] = a_zero_point; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5359 params->rndnu_neon.a_zero_point[1] = a_zero_point; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5360 params->rndnu_neon.b_zero_point[0] = b_zero_point; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5361 params->rndnu_neon.b_zero_point[1] = b_zero_point; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5362 params->rndnu_neon.left_pre_shift = -pre_shift; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5363 params->rndnu_neon.multiplier = multiplier; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5364 params->rndnu_neon.left_post_shift = -post_shift; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5365 params->rndnu_neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5366 params->rndnu_neon.output_min = output_min; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5367 params->rndnu_neon.output_max = output_max; in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5368 return sizeof(params->rndnu_neon); in xnn_init_qu8_mul_minmax_rndnu_neon_params()
5374 union xnn_qu8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_mul_minmax_fp32_sse2_params()
5386 params->fp32_sse2.a_zero_point[i] = (int16_t) (uint16_t) a_zero_point; in xnn_init_qu8_mul_minmax_fp32_sse2_params()
5387 params->fp32_sse2.b_zero_point[i] = (int16_t) (uint16_t) b_zero_point; in xnn_init_qu8_mul_minmax_fp32_sse2_params()
5390 params->fp32_sse2.scale[i] = product_output_scale; in xnn_init_qu8_mul_minmax_fp32_sse2_params()
5393 params->fp32_sse2.output_zero_point[i] = (int16_t) (uint16_t) output_zero_point; in xnn_init_qu8_mul_minmax_fp32_sse2_params()
5396 params->fp32_sse2.output_min[i] = output_min; in xnn_init_qu8_mul_minmax_fp32_sse2_params()
5397 params->fp32_sse2.output_max[i] = output_max; in xnn_init_qu8_mul_minmax_fp32_sse2_params()
5399 return sizeof(params->fp32_sse2); in xnn_init_qu8_mul_minmax_fp32_sse2_params()
5405 union xnn_qu8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5420 params->fp32_wasmsimd.a_zero_point[i] = (int16_t) a_zero_point; in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5421 params->fp32_wasmsimd.b_zero_point[i] = (int16_t) b_zero_point; in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5424 params->fp32_wasmsimd.scale[i] = product_output_scale; in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5425 params->fp32_wasmsimd.magic_bias[i] = 12582912.0f; in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5426 params->fp32_wasmsimd.magic_min[i] = magic_min; in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5427 params->fp32_wasmsimd.magic_bias_less_output_zero_point[i] = magic_bias_less_output_zero_point; in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5430 params->fp32_wasmsimd.output_max[i] = output_max; in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5432 return sizeof(params->fp32_wasmsimd); in xnn_init_qu8_mul_minmax_fp32_wasmsimd_params()
5437 union xnn_qs8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5448 params->fp32_scalar.a_zero_point = (int16_t) a_zero_point; in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5449 params->fp32_scalar.b_zero_point = (int16_t) b_zero_point; in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5450 params->fp32_scalar.scale = product_output_scale; in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5451params->fp32_scalar.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) output_… in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5452params->fp32_scalar.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) output_… in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5453 params->fp32_scalar.magic_bias = 12582912.0f; in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5454params->fp32_scalar.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) output_zer… in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5455 return sizeof(params->fp32_scalar); in xnn_init_qs8_mul_minmax_fp32_scalar_params()
5460 union xnn_qs8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_mul_minmax_fp32_neon_params()
5471 params->fp32_neon.a_zero_point[0] = a_zero_point; in xnn_init_qs8_mul_minmax_fp32_neon_params()
5472 params->fp32_neon.a_zero_point[1] = a_zero_point; in xnn_init_qs8_mul_minmax_fp32_neon_params()
5473 params->fp32_neon.b_zero_point[0] = b_zero_point; in xnn_init_qs8_mul_minmax_fp32_neon_params()
5474 params->fp32_neon.b_zero_point[1] = b_zero_point; in xnn_init_qs8_mul_minmax_fp32_neon_params()
5475 params->fp32_neon.scale = product_output_scale; in xnn_init_qs8_mul_minmax_fp32_neon_params()
5476 params->fp32_neon.magic_bias = 12582912.0f; in xnn_init_qs8_mul_minmax_fp32_neon_params()
5477params->fp32_neon.magic_bias_less_output_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_… in xnn_init_qs8_mul_minmax_fp32_neon_params()
5478 params->fp32_neon.output_min = output_min; in xnn_init_qs8_mul_minmax_fp32_neon_params()
5479 params->fp32_neon.output_max = output_max; in xnn_init_qs8_mul_minmax_fp32_neon_params()
5480 return sizeof(params->fp32_neon); in xnn_init_qs8_mul_minmax_fp32_neon_params()
5484 union xnn_qs8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5495 params->fp32_neonv8.a_zero_point[0] = a_zero_point; in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5496 params->fp32_neonv8.a_zero_point[1] = a_zero_point; in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5497 params->fp32_neonv8.b_zero_point[0] = b_zero_point; in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5498 params->fp32_neonv8.b_zero_point[1] = b_zero_point; in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5499 params->fp32_neonv8.scale = product_output_scale; in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5500 params->fp32_neonv8.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5501 params->fp32_neonv8.output_min = output_min; in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5502 params->fp32_neonv8.output_max = output_max; in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5503 return sizeof(params->fp32_neonv8); in xnn_init_qs8_mul_minmax_fp32_neonv8_params()
5507 union xnn_qs8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5535 params->rndnu_neon.a_zero_point[0] = a_zero_point; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5536 params->rndnu_neon.a_zero_point[1] = a_zero_point; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5537 params->rndnu_neon.b_zero_point[0] = b_zero_point; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5538 params->rndnu_neon.b_zero_point[1] = b_zero_point; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5539 params->rndnu_neon.left_pre_shift = -pre_shift; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5540 params->rndnu_neon.multiplier = multiplier; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5541 params->rndnu_neon.left_post_shift = -post_shift; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5542 params->rndnu_neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5543 params->rndnu_neon.output_min = output_min; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5544 params->rndnu_neon.output_max = output_max; in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5545 return sizeof(params->rndnu_neon); in xnn_init_qs8_mul_minmax_rndnu_neon_params()
5551 union xnn_qs8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_mul_minmax_fp32_sse2_params()
5563 params->fp32_sse2.a_zero_point[i] = (int16_t) a_zero_point; in xnn_init_qs8_mul_minmax_fp32_sse2_params()
5564 params->fp32_sse2.b_zero_point[i] = (int16_t) b_zero_point; in xnn_init_qs8_mul_minmax_fp32_sse2_params()
5567 params->fp32_sse2.scale[i] = product_output_scale; in xnn_init_qs8_mul_minmax_fp32_sse2_params()
5570 params->fp32_sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_mul_minmax_fp32_sse2_params()
5573 params->fp32_sse2.output_min[i] = (int16_t) output_min; in xnn_init_qs8_mul_minmax_fp32_sse2_params()
5574 params->fp32_sse2.output_max[i] = (int16_t) output_max; in xnn_init_qs8_mul_minmax_fp32_sse2_params()
5576 return sizeof(params->fp32_sse2); in xnn_init_qs8_mul_minmax_fp32_sse2_params()
5580 union xnn_qs8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_mul_minmax_fp32_sse4_params()
5592 params->fp32_sse4.a_zero_point[i] = (int16_t) a_zero_point; in xnn_init_qs8_mul_minmax_fp32_sse4_params()
5593 params->fp32_sse4.b_zero_point[i] = (int16_t) b_zero_point; in xnn_init_qs8_mul_minmax_fp32_sse4_params()
5596 params->fp32_sse4.scale[i] = product_output_scale; in xnn_init_qs8_mul_minmax_fp32_sse4_params()
5599 params->fp32_sse4.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_mul_minmax_fp32_sse4_params()
5602 params->fp32_sse4.output_min[i] = output_min; in xnn_init_qs8_mul_minmax_fp32_sse4_params()
5603 params->fp32_sse4.output_max[i] = output_max; in xnn_init_qs8_mul_minmax_fp32_sse4_params()
5605 return sizeof(params->fp32_sse4); in xnn_init_qs8_mul_minmax_fp32_sse4_params()
5611 union xnn_qs8_mul_minmax_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5626 params->fp32_wasmsimd.a_zero_point[i] = (int16_t) a_zero_point; in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5627 params->fp32_wasmsimd.b_zero_point[i] = (int16_t) b_zero_point; in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5630 params->fp32_wasmsimd.scale[i] = product_output_scale; in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5631 params->fp32_wasmsimd.magic_bias[i] = 12582912.0f; in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5632 params->fp32_wasmsimd.magic_min[i] = magic_min; in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5633 params->fp32_wasmsimd.magic_bias_less_output_zero_point[i] = magic_bias_less_output_zero_point; in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5636 params->fp32_wasmsimd.output_max[i] = output_max; in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5638 return sizeof(params->fp32_wasmsimd); in xnn_init_qs8_mul_minmax_fp32_wasmsimd_params()
5643 union xnn_f16_f32_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_f32_cvt_scalar_params()
5645 params->scalar.sign_mask = UINT32_C(0x80000000); in xnn_init_f16_f32_cvt_scalar_params()
5646 params->scalar.exp_offset = UINT32_C(0x70000000); in xnn_init_f16_f32_cvt_scalar_params()
5647 params->scalar.exp_scale = 0x1.0p-112f; in xnn_init_f16_f32_cvt_scalar_params()
5648 params->scalar.magic_mask = UINT32_C(0x3F000000); in xnn_init_f16_f32_cvt_scalar_params()
5649 params->scalar.magic_bias = 0.5f; in xnn_init_f16_f32_cvt_scalar_params()
5650 params->scalar.denorm_cutoff = UINT32_C(0x08000000); in xnn_init_f16_f32_cvt_scalar_params()
5651 return sizeof(params->scalar); in xnn_init_f16_f32_cvt_scalar_params()
5656 union xnn_f16_f32_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_f32_cvt_neon_params()
5658 params->neon.exp_scale = 0x1.0p-112f; in xnn_init_f16_f32_cvt_neon_params()
5659 return sizeof(params->neon); in xnn_init_f16_f32_cvt_neon_params()
5665 union xnn_f16_f32_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_f32_cvt_sse_int16_params()
5668 params->sse_int16.sign_mask[i] = UINT16_C(0x8000); in xnn_init_f16_f32_cvt_sse_int16_params()
5669 params->sse_int16.exp_offset[i] = UINT16_C(0x7000); in xnn_init_f16_f32_cvt_sse_int16_params()
5672 params->sse_int16.exp_scale[i] = 0x1.0p-112f; in xnn_init_f16_f32_cvt_sse_int16_params()
5675 params->sse_int16.magic_mask[i] = UINT16_C(0x3F00); in xnn_init_f16_f32_cvt_sse_int16_params()
5678 params->sse_int16.magic_bias[i] = 0.5f; in xnn_init_f16_f32_cvt_sse_int16_params()
5681 params->sse_int16.denorm_cutoff[i] = INT16_C(0x0400); in xnn_init_f16_f32_cvt_sse_int16_params()
5683 return sizeof(params->sse_int16); in xnn_init_f16_f32_cvt_sse_int16_params()
5687 union xnn_f16_f32_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_f32_cvt_sse_int32_params()
5690 params->sse_int32.sign_mask[i] = UINT32_C(0x80000000); in xnn_init_f16_f32_cvt_sse_int32_params()
5691 params->sse_int32.exp_offset[i] = UINT32_C(0x70000000); in xnn_init_f16_f32_cvt_sse_int32_params()
5692 params->sse_int32.exp_scale[i] = 0x1.0p-112f; in xnn_init_f16_f32_cvt_sse_int32_params()
5693 params->sse_int32.magic_bias[i] = UINT32_C(0x3F000000); in xnn_init_f16_f32_cvt_sse_int32_params()
5694 params->sse_int32.denorm_cutoff[i] = INT32_C(0x04000000); in xnn_init_f16_f32_cvt_sse_int32_params()
5696 return sizeof(params->sse_int32); in xnn_init_f16_f32_cvt_sse_int32_params()
5702 union xnn_f16_f32_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_f32_cvt_wasmsimd_int16_params()
5705 params->wasmsimd_int16.sign_mask[i] = UINT16_C(0x8000); in xnn_init_f16_f32_cvt_wasmsimd_int16_params()
5706 params->wasmsimd_int16.exp_offset[i] = UINT16_C(0x7000); in xnn_init_f16_f32_cvt_wasmsimd_int16_params()
5709 params->wasmsimd_int16.exp_scale[i] = 0x1.0p-112f; in xnn_init_f16_f32_cvt_wasmsimd_int16_params()
5712 params->wasmsimd_int16.magic_mask[i] = UINT16_C(0x3F00); in xnn_init_f16_f32_cvt_wasmsimd_int16_params()
5715 params->wasmsimd_int16.magic_bias[i] = 0.5f; in xnn_init_f16_f32_cvt_wasmsimd_int16_params()
5718 params->wasmsimd_int16.denorm_cutoff[i] = INT16_C(0x0400); in xnn_init_f16_f32_cvt_wasmsimd_int16_params()
5720 return sizeof(params->wasmsimd_int16); in xnn_init_f16_f32_cvt_wasmsimd_int16_params()
5724 union xnn_f16_f32_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f16_f32_cvt_wasmsimd_int32_params()
5727 params->wasmsimd_int32.sign_mask[i] = UINT32_C(0x80000000); in xnn_init_f16_f32_cvt_wasmsimd_int32_params()
5728 params->wasmsimd_int32.exp_offset[i] = UINT32_C(0x70000000); in xnn_init_f16_f32_cvt_wasmsimd_int32_params()
5729 params->wasmsimd_int32.exp_scale[i] = 0x1.0p-112f; in xnn_init_f16_f32_cvt_wasmsimd_int32_params()
5730 params->wasmsimd_int32.magic_bias[i] = UINT32_C(0x3F000000); in xnn_init_f16_f32_cvt_wasmsimd_int32_params()
5731 params->wasmsimd_int32.denorm_cutoff[i] = INT32_C(0x04000000); in xnn_init_f16_f32_cvt_wasmsimd_int32_params()
5733 return sizeof(params->wasmsimd_int32); in xnn_init_f16_f32_cvt_wasmsimd_int32_params()
5738 union xnn_f32_f16_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5740 params->scalar_bitcast.nonsign_mask = UINT32_C(0x7FFFFFFF); in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5741 params->scalar_bitcast.exp_bias = UINT32_C(0x07800000); in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5742 params->scalar_bitcast.scale_to_inf = 0x1.0p+112f; in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5743 params->scalar_bitcast.expw_max = UINT32_C(0x7F800000); in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5744 params->scalar_bitcast.scale_to_zero = 0x1.0p-110f; in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5745 params->scalar_bitcast.bias_min = UINT32_C(0x40000000); in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5746 params->scalar_bitcast.exph_mask = UINT16_C(0x7C00); in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5747 params->scalar_bitcast.manth_mask = UINT16_C(0x0FFF); in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5748 params->scalar_bitcast.nanh = UINT16_C(0x7E00); in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5749 return sizeof(params->scalar_bitcast); in xnn_init_f32_f16_cvt_scalar_bitcast_params()
5753 union xnn_f32_f16_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5755 params->scalar_fabsf.scale_to_inf = 0x1.0p+112f; in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5756 params->scalar_fabsf.exp_bias = UINT32_C(0x07800000); in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5757 params->scalar_fabsf.scale_to_zero = 0x1.0p-110f; in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5758 params->scalar_fabsf.expw_max = UINT32_C(0x7F800000); in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5759 params->scalar_fabsf.bias_min = UINT32_C(0x40000000); in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5760 params->scalar_fabsf.exph_mask = UINT16_C(0x7C00); in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5761 params->scalar_fabsf.manth_mask = UINT16_C(0x0FFF); in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5762 params->scalar_fabsf.nanh = UINT16_C(0x7E00); in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5763 return sizeof(params->scalar_fabsf); in xnn_init_f32_f16_cvt_scalar_fabsf_params()
5768 union xnn_f32_f16_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_f16_cvt_neon_params()
5770 params->neon.exp_bias = UINT32_C(0x07800000); in xnn_init_f32_f16_cvt_neon_params()
5771 params->neon.scale_to_inf = 0x1.0p+112f; in xnn_init_f32_f16_cvt_neon_params()
5772 params->neon.expw_max = UINT32_C(0x7F800000); in xnn_init_f32_f16_cvt_neon_params()
5773 params->neon.scale_to_zero = 0x1.0p-110f; in xnn_init_f32_f16_cvt_neon_params()
5774 return sizeof(params->neon); in xnn_init_f32_f16_cvt_neon_params()
5780 union xnn_f32_f16_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_f16_cvt_sse2_params()
5783 params->sse2.nonsign_mask[i] = UINT32_C(0x7FFFFFFF); in xnn_init_f32_f16_cvt_sse2_params()
5784 params->sse2.exp_bias[i] = UINT32_C(0x07800000); in xnn_init_f32_f16_cvt_sse2_params()
5785 params->sse2.scale_to_inf[i] = 0x1.0p+112f; in xnn_init_f32_f16_cvt_sse2_params()
5786 params->sse2.expw_max[i] = UINT32_C(0x7F800000); in xnn_init_f32_f16_cvt_sse2_params()
5787 params->sse2.scale_to_zero[i] = 0x1.0p-110f; in xnn_init_f32_f16_cvt_sse2_params()
5789 params->sse2.bias_min[0] = INT16_C(0x8000); in xnn_init_f32_f16_cvt_sse2_params()
5790 params->sse2.bias_min[1] = INT16_C(0x4000); in xnn_init_f32_f16_cvt_sse2_params()
5791 params->sse2.bias_min[2] = INT16_C(0x8000); in xnn_init_f32_f16_cvt_sse2_params()
5792 params->sse2.bias_min[3] = INT16_C(0x4000); in xnn_init_f32_f16_cvt_sse2_params()
5793 params->sse2.bias_min[4] = INT16_C(0x8000); in xnn_init_f32_f16_cvt_sse2_params()
5794 params->sse2.bias_min[5] = INT16_C(0x4000); in xnn_init_f32_f16_cvt_sse2_params()
5795 params->sse2.bias_min[6] = INT16_C(0x8000); in xnn_init_f32_f16_cvt_sse2_params()
5796 params->sse2.bias_min[7] = INT16_C(0x4000); in xnn_init_f32_f16_cvt_sse2_params()
5798 params->sse2.manth_mask[i] = UINT32_C(0x00000FFF); in xnn_init_f32_f16_cvt_sse2_params()
5799 params->sse2.exph_mask[i] = UINT32_C(0x00007C00); in xnn_init_f32_f16_cvt_sse2_params()
5802 params->sse2.nanh[i] = UINT16_C(0x7E00); in xnn_init_f32_f16_cvt_sse2_params()
5804 return sizeof(params->sse2); in xnn_init_f32_f16_cvt_sse2_params()
5808 union xnn_f32_f16_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_f16_cvt_f16c_params()
5811 params->f16c.mask_table[i] = -1; in xnn_init_f32_f16_cvt_f16c_params()
5814 params->f16c.mask_table[i] = 0; in xnn_init_f32_f16_cvt_f16c_params()
5816 return sizeof(params->f16c); in xnn_init_f32_f16_cvt_f16c_params()
5822 union xnn_f32_f16_cvt_params params[XNN_MIN_ELEMENTS(1)]) in xnn_init_f32_f16_cvt_wasmsimd_params()
5825 params->wasmsimd.exp_bias[i] = UINT32_C(0x07800000); in xnn_init_f32_f16_cvt_wasmsimd_params()
5826 params->wasmsimd.scale_to_inf[i] = 0x1.0p+112f; in xnn_init_f32_f16_cvt_wasmsimd_params()
5827 params->wasmsimd.expw_max[i] = UINT32_C(0x7F800000); in xnn_init_f32_f16_cvt_wasmsimd_params()
5828 params->wasmsimd.scale_to_zero[i] = 0x1.0p-110f; in xnn_init_f32_f16_cvt_wasmsimd_params()
5830 params->wasmsimd.bias_min[0] = INT16_C(0x8000); in xnn_init_f32_f16_cvt_wasmsimd_params()
5831 params->wasmsimd.bias_min[1] = INT16_C(0x4000); in xnn_init_f32_f16_cvt_wasmsimd_params()
5832 params->wasmsimd.bias_min[2] = INT16_C(0x8000); in xnn_init_f32_f16_cvt_wasmsimd_params()
5833 params->wasmsimd.bias_min[3] = INT16_C(0x4000); in xnn_init_f32_f16_cvt_wasmsimd_params()
5835 params->wasmsimd.manth_mask[i] = UINT32_C(0x00000FFF); in xnn_init_f32_f16_cvt_wasmsimd_params()
5836 params->wasmsimd.exph_mask[i] = UINT32_C(0x00007C00); in xnn_init_f32_f16_cvt_wasmsimd_params()
5839 params->wasmsimd.nanh[i] = UINT16_C(0x7E00); in xnn_init_f32_f16_cvt_wasmsimd_params()
5841 return sizeof(params->wasmsimd); in xnn_init_f32_f16_cvt_wasmsimd_params()
5846 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_scalar_fmagic_params()
5852 params->scalar_fmagic.scale = scale; in xnn_init_f32_qs8_cvt_scalar_fmagic_params()
5853params->scalar_fmagic.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) outpu… in xnn_init_f32_qs8_cvt_scalar_fmagic_params()
5854params->scalar_fmagic.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) outpu… in xnn_init_f32_qs8_cvt_scalar_fmagic_params()
5855 params->scalar_fmagic.magic_bias = 12582912.0f; in xnn_init_f32_qs8_cvt_scalar_fmagic_params()
5856params->scalar_fmagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_poi… in xnn_init_f32_qs8_cvt_scalar_fmagic_params()
5857 return sizeof(params->scalar_fmagic); in xnn_init_f32_qs8_cvt_scalar_fmagic_params()
5861 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_scalar_imagic_params()
5869 params->scalar_imagic.scale = scale; in xnn_init_f32_qs8_cvt_scalar_imagic_params()
5870 params->scalar_imagic.magic_bias = 12582912.0f; in xnn_init_f32_qs8_cvt_scalar_imagic_params()
5871params->scalar_imagic.magic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_poi… in xnn_init_f32_qs8_cvt_scalar_imagic_params()
5872params->scalar_imagic.magic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_poi… in xnn_init_f32_qs8_cvt_scalar_imagic_params()
5873params->scalar_imagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_poi… in xnn_init_f32_qs8_cvt_scalar_imagic_params()
5874 return sizeof(params->scalar_imagic); in xnn_init_f32_qs8_cvt_scalar_imagic_params()
5878 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_scalar_lrintf_params()
5884 params->scalar_lrintf.scale = scale; in xnn_init_f32_qs8_cvt_scalar_lrintf_params()
5885params->scalar_lrintf.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) outpu… in xnn_init_f32_qs8_cvt_scalar_lrintf_params()
5886params->scalar_lrintf.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) outpu… in xnn_init_f32_qs8_cvt_scalar_lrintf_params()
5887 params->scalar_lrintf.output_zero_point = (int32_t) output_zero_point; in xnn_init_f32_qs8_cvt_scalar_lrintf_params()
5888 return sizeof(params->scalar_lrintf); in xnn_init_f32_qs8_cvt_scalar_lrintf_params()
5893 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_neon_params()
5899 params->neon.scale = scale; in xnn_init_f32_qs8_cvt_neon_params()
5900 params->neon.magic_bias = 12582912.0f; in xnn_init_f32_qs8_cvt_neon_params()
5901 params->neon.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_point; in xnn_init_f32_qs8_cvt_neon_params()
5902 params->neon.output_min = output_min; in xnn_init_f32_qs8_cvt_neon_params()
5903 params->neon.output_max = output_max; in xnn_init_f32_qs8_cvt_neon_params()
5904 return sizeof(params->neon); in xnn_init_f32_qs8_cvt_neon_params()
5908 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_neonv8_params()
5914 params->neonv8.scale = scale; in xnn_init_f32_qs8_cvt_neonv8_params()
5915 params->neonv8.output_zero_point = (int16_t) output_zero_point; in xnn_init_f32_qs8_cvt_neonv8_params()
5916 params->neonv8.output_min = output_min; in xnn_init_f32_qs8_cvt_neonv8_params()
5917 params->neonv8.output_max = output_max; in xnn_init_f32_qs8_cvt_neonv8_params()
5918 return sizeof(params->neonv8); in xnn_init_f32_qs8_cvt_neonv8_params()
5924 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_sse2_params()
5932 params->sse2.scale[i] = scale; in xnn_init_f32_qs8_cvt_sse2_params()
5933 params->sse2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qs8_cvt_sse2_params()
5936 params->sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qs8_cvt_sse2_params()
5937 params->sse2.output_min[i] = (int16_t) output_min; in xnn_init_f32_qs8_cvt_sse2_params()
5939 return sizeof(params->sse2); in xnn_init_f32_qs8_cvt_sse2_params()
5943 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_sse4_params()
5951 params->sse4.scale[i] = scale; in xnn_init_f32_qs8_cvt_sse4_params()
5952 params->sse4.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qs8_cvt_sse4_params()
5955 params->sse4.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qs8_cvt_sse4_params()
5958 params->sse4.output_min[i] = output_min; in xnn_init_f32_qs8_cvt_sse4_params()
5960 return sizeof(params->sse4); in xnn_init_f32_qs8_cvt_sse4_params()
5964 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_avx_params()
5972 params->avx.scale[i] = scale; in xnn_init_f32_qs8_cvt_avx_params()
5973 params->avx.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qs8_cvt_avx_params()
5976 params->avx.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qs8_cvt_avx_params()
5979 params->avx.output_min[i] = output_min; in xnn_init_f32_qs8_cvt_avx_params()
5982 params->avx.mask_table[i] = -1; in xnn_init_f32_qs8_cvt_avx_params()
5985 params->avx.mask_table[i] = 0; in xnn_init_f32_qs8_cvt_avx_params()
5987 return sizeof(params->avx); in xnn_init_f32_qs8_cvt_avx_params()
5991 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_avx2_params()
5999 params->avx2.scale[i] = scale; in xnn_init_f32_qs8_cvt_avx2_params()
6000 params->avx2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qs8_cvt_avx2_params()
6003 params->avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qs8_cvt_avx2_params()
6005 params->avx2.shuffle_mask[0] = 0; in xnn_init_f32_qs8_cvt_avx2_params()
6006 params->avx2.shuffle_mask[1] = 4; in xnn_init_f32_qs8_cvt_avx2_params()
6007 params->avx2.shuffle_mask[2] = 1; in xnn_init_f32_qs8_cvt_avx2_params()
6008 params->avx2.shuffle_mask[3] = 5; in xnn_init_f32_qs8_cvt_avx2_params()
6009 params->avx2.shuffle_mask[4] = 2; in xnn_init_f32_qs8_cvt_avx2_params()
6010 params->avx2.shuffle_mask[5] = 6; in xnn_init_f32_qs8_cvt_avx2_params()
6011 params->avx2.shuffle_mask[6] = 3; in xnn_init_f32_qs8_cvt_avx2_params()
6012 params->avx2.shuffle_mask[7] = 7; in xnn_init_f32_qs8_cvt_avx2_params()
6014 params->avx2.output_min[i] = output_min; in xnn_init_f32_qs8_cvt_avx2_params()
6017 params->avx2.mask_table[i] = -1; in xnn_init_f32_qs8_cvt_avx2_params()
6020 params->avx2.mask_table[i] = 0; in xnn_init_f32_qs8_cvt_avx2_params()
6022 return sizeof(params->avx2); in xnn_init_f32_qs8_cvt_avx2_params()
6026 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_avx512_params()
6034 params->avx512.scale[i] = scale; in xnn_init_f32_qs8_cvt_avx512_params()
6035 params->avx512.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qs8_cvt_avx512_params()
6038 params->avx512.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qs8_cvt_avx512_params()
6041 params->avx512.output_min[i] = output_min; in xnn_init_f32_qs8_cvt_avx512_params()
6043 params->avx512.shuffle512_mask[0] = 0; in xnn_init_f32_qs8_cvt_avx512_params()
6044 params->avx512.shuffle512_mask[1] = 4; in xnn_init_f32_qs8_cvt_avx512_params()
6045 params->avx512.shuffle512_mask[2] = 8; in xnn_init_f32_qs8_cvt_avx512_params()
6046 params->avx512.shuffle512_mask[3] = 12; in xnn_init_f32_qs8_cvt_avx512_params()
6047 params->avx512.shuffle512_mask[4] = 1; in xnn_init_f32_qs8_cvt_avx512_params()
6048 params->avx512.shuffle512_mask[5] = 5; in xnn_init_f32_qs8_cvt_avx512_params()
6049 params->avx512.shuffle512_mask[6] = 9; in xnn_init_f32_qs8_cvt_avx512_params()
6050 params->avx512.shuffle512_mask[7] = 13; in xnn_init_f32_qs8_cvt_avx512_params()
6051 params->avx512.shuffle512_mask[8] = 2; in xnn_init_f32_qs8_cvt_avx512_params()
6052 params->avx512.shuffle512_mask[9] = 6; in xnn_init_f32_qs8_cvt_avx512_params()
6053 params->avx512.shuffle512_mask[10] = 10; in xnn_init_f32_qs8_cvt_avx512_params()
6054 params->avx512.shuffle512_mask[11] = 14; in xnn_init_f32_qs8_cvt_avx512_params()
6055 params->avx512.shuffle512_mask[12] = 3; in xnn_init_f32_qs8_cvt_avx512_params()
6056 params->avx512.shuffle512_mask[13] = 7; in xnn_init_f32_qs8_cvt_avx512_params()
6057 params->avx512.shuffle512_mask[14] = 11; in xnn_init_f32_qs8_cvt_avx512_params()
6058 params->avx512.shuffle512_mask[15] = 15; in xnn_init_f32_qs8_cvt_avx512_params()
6059 params->avx512.shuffle256_mask[0] = 0; in xnn_init_f32_qs8_cvt_avx512_params()
6060 params->avx512.shuffle256_mask[1] = 4; in xnn_init_f32_qs8_cvt_avx512_params()
6061 params->avx512.shuffle256_mask[2] = 2; in xnn_init_f32_qs8_cvt_avx512_params()
6062 params->avx512.shuffle256_mask[3] = 6; in xnn_init_f32_qs8_cvt_avx512_params()
6063 params->avx512.shuffle256_mask[4] = 1; in xnn_init_f32_qs8_cvt_avx512_params()
6064 params->avx512.shuffle256_mask[5] = 5; in xnn_init_f32_qs8_cvt_avx512_params()
6065 params->avx512.shuffle256_mask[6] = 3; in xnn_init_f32_qs8_cvt_avx512_params()
6066 params->avx512.shuffle256_mask[7] = 7; in xnn_init_f32_qs8_cvt_avx512_params()
6067 return sizeof(params->avx512); in xnn_init_f32_qs8_cvt_avx512_params()
6073 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_wasmsimd_cvt_params()
6080 params->wasmsimd_cvt.scale[i] = scale; in xnn_init_f32_qs8_cvt_wasmsimd_cvt_params()
6083 params->wasmsimd_cvt.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qs8_cvt_wasmsimd_cvt_params()
6086 params->wasmsimd_cvt.output_min[i] = output_min; in xnn_init_f32_qs8_cvt_wasmsimd_cvt_params()
6087 params->wasmsimd_cvt.output_max[i] = output_max; in xnn_init_f32_qs8_cvt_wasmsimd_cvt_params()
6089 return sizeof(params->wasmsimd_cvt); in xnn_init_f32_qs8_cvt_wasmsimd_cvt_params()
6093 union xnn_f32_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qs8_cvt_wasmsimd_magic_params()
6103 params->wasmsimd_magic.scale[i] = scale; in xnn_init_f32_qs8_cvt_wasmsimd_magic_params()
6104 params->wasmsimd_magic.magic_bias[i] = 12582912.0f; in xnn_init_f32_qs8_cvt_wasmsimd_magic_params()
6105 params->wasmsimd_magic.magic_min[i] = magic_min; in xnn_init_f32_qs8_cvt_wasmsimd_magic_params()
6106 params->wasmsimd_magic.magic_bias_less_zero_point[i] = magic_bias_less_zero_point; in xnn_init_f32_qs8_cvt_wasmsimd_magic_params()
6109 params->wasmsimd_magic.output_max[i] = output_max; in xnn_init_f32_qs8_cvt_wasmsimd_magic_params()
6111 return sizeof(params->wasmsimd_magic); in xnn_init_f32_qs8_cvt_wasmsimd_magic_params()
6116 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_scalar_fmagic_params()
6122 params->scalar_fmagic.scale = scale; in xnn_init_f32_qu8_cvt_scalar_fmagic_params()
6123params->scalar_fmagic.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) outpu… in xnn_init_f32_qu8_cvt_scalar_fmagic_params()
6124params->scalar_fmagic.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) outpu… in xnn_init_f32_qu8_cvt_scalar_fmagic_params()
6125 params->scalar_fmagic.magic_bias = 12582912.0f; in xnn_init_f32_qu8_cvt_scalar_fmagic_params()
6126params->scalar_fmagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_poi… in xnn_init_f32_qu8_cvt_scalar_fmagic_params()
6127 return sizeof(params->scalar_fmagic); in xnn_init_f32_qu8_cvt_scalar_fmagic_params()
6131 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_scalar_imagic_params()
6139 params->scalar_imagic.scale = scale; in xnn_init_f32_qu8_cvt_scalar_imagic_params()
6140 params->scalar_imagic.magic_bias = 12582912.0f; in xnn_init_f32_qu8_cvt_scalar_imagic_params()
6141params->scalar_imagic.magic_min = (int32_t) float_as_uint32(12582912.0f + output_min_less_zero_poi… in xnn_init_f32_qu8_cvt_scalar_imagic_params()
6142params->scalar_imagic.magic_max = (int32_t) float_as_uint32(12582912.0f + output_max_less_zero_poi… in xnn_init_f32_qu8_cvt_scalar_imagic_params()
6143params->scalar_imagic.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_poi… in xnn_init_f32_qu8_cvt_scalar_imagic_params()
6144 return sizeof(params->scalar_imagic); in xnn_init_f32_qu8_cvt_scalar_imagic_params()
6148 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_scalar_lrintf_params()
6154 params->scalar_lrintf.scale = scale; in xnn_init_f32_qu8_cvt_scalar_lrintf_params()
6155params->scalar_lrintf.output_min_less_zero_point = (float) ((int32_t) output_min - (int32_t) outpu… in xnn_init_f32_qu8_cvt_scalar_lrintf_params()
6156params->scalar_lrintf.output_max_less_zero_point = (float) ((int32_t) output_max - (int32_t) outpu… in xnn_init_f32_qu8_cvt_scalar_lrintf_params()
6157 params->scalar_lrintf.output_zero_point = (int32_t) output_zero_point; in xnn_init_f32_qu8_cvt_scalar_lrintf_params()
6158 return sizeof(params->scalar_lrintf); in xnn_init_f32_qu8_cvt_scalar_lrintf_params()
6163 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_neon_params()
6169 params->neon.scale = scale; in xnn_init_f32_qu8_cvt_neon_params()
6170 params->neon.magic_bias = 12582912.0f; in xnn_init_f32_qu8_cvt_neon_params()
6171 params->neon.magic_bias_less_zero_point = INT32_C(0x4B400000) - (int32_t) output_zero_point; in xnn_init_f32_qu8_cvt_neon_params()
6172 params->neon.output_min = output_min; in xnn_init_f32_qu8_cvt_neon_params()
6173 params->neon.output_max = output_max; in xnn_init_f32_qu8_cvt_neon_params()
6174 return sizeof(params->neon); in xnn_init_f32_qu8_cvt_neon_params()
6178 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_neonv8_params()
6184 params->neonv8.scale = scale; in xnn_init_f32_qu8_cvt_neonv8_params()
6185 params->neonv8.output_zero_point = (int16_t) output_zero_point; in xnn_init_f32_qu8_cvt_neonv8_params()
6186 params->neonv8.output_min = output_min; in xnn_init_f32_qu8_cvt_neonv8_params()
6187 params->neonv8.output_max = output_max; in xnn_init_f32_qu8_cvt_neonv8_params()
6188 return sizeof(params->neonv8); in xnn_init_f32_qu8_cvt_neonv8_params()
6194 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_sse2_params()
6202 params->sse2.scale[i] = scale; in xnn_init_f32_qu8_cvt_sse2_params()
6203 params->sse2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qu8_cvt_sse2_params()
6206 params->sse2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qu8_cvt_sse2_params()
6209 params->sse2.output_min[i] = output_min; in xnn_init_f32_qu8_cvt_sse2_params()
6211 return sizeof(params->sse2); in xnn_init_f32_qu8_cvt_sse2_params()
6215 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_avx_params()
6223 params->avx.scale[i] = scale; in xnn_init_f32_qu8_cvt_avx_params()
6224 params->avx.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qu8_cvt_avx_params()
6227 params->avx.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qu8_cvt_avx_params()
6230 params->avx.output_min[i] = output_min; in xnn_init_f32_qu8_cvt_avx_params()
6233 params->avx.mask_table[i] = -1; in xnn_init_f32_qu8_cvt_avx_params()
6236 params->avx.mask_table[i] = 0; in xnn_init_f32_qu8_cvt_avx_params()
6238 return sizeof(params->avx); in xnn_init_f32_qu8_cvt_avx_params()
6242 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_avx2_params()
6250 params->avx2.scale[i] = scale; in xnn_init_f32_qu8_cvt_avx2_params()
6251 params->avx2.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qu8_cvt_avx2_params()
6254 params->avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qu8_cvt_avx2_params()
6256 params->avx2.shuffle_mask[0] = 0; in xnn_init_f32_qu8_cvt_avx2_params()
6257 params->avx2.shuffle_mask[1] = 4; in xnn_init_f32_qu8_cvt_avx2_params()
6258 params->avx2.shuffle_mask[2] = 1; in xnn_init_f32_qu8_cvt_avx2_params()
6259 params->avx2.shuffle_mask[3] = 5; in xnn_init_f32_qu8_cvt_avx2_params()
6260 params->avx2.shuffle_mask[4] = 2; in xnn_init_f32_qu8_cvt_avx2_params()
6261 params->avx2.shuffle_mask[5] = 6; in xnn_init_f32_qu8_cvt_avx2_params()
6262 params->avx2.shuffle_mask[6] = 3; in xnn_init_f32_qu8_cvt_avx2_params()
6263 params->avx2.shuffle_mask[7] = 7; in xnn_init_f32_qu8_cvt_avx2_params()
6265 params->avx2.output_min[i] = output_min; in xnn_init_f32_qu8_cvt_avx2_params()
6268 params->avx2.mask_table[i] = -1; in xnn_init_f32_qu8_cvt_avx2_params()
6271 params->avx2.mask_table[i] = 0; in xnn_init_f32_qu8_cvt_avx2_params()
6273 return sizeof(params->avx2); in xnn_init_f32_qu8_cvt_avx2_params()
6277 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_avx512_params()
6285 params->avx512.scale[i] = scale; in xnn_init_f32_qu8_cvt_avx512_params()
6286 params->avx512.output_max_less_zero_point[i] = output_max_less_zero_point; in xnn_init_f32_qu8_cvt_avx512_params()
6289 params->avx512.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qu8_cvt_avx512_params()
6292 params->avx512.output_min[i] = output_min; in xnn_init_f32_qu8_cvt_avx512_params()
6294 params->avx512.shuffle512_mask[0] = 0; in xnn_init_f32_qu8_cvt_avx512_params()
6295 params->avx512.shuffle512_mask[1] = 4; in xnn_init_f32_qu8_cvt_avx512_params()
6296 params->avx512.shuffle512_mask[2] = 8; in xnn_init_f32_qu8_cvt_avx512_params()
6297 params->avx512.shuffle512_mask[3] = 12; in xnn_init_f32_qu8_cvt_avx512_params()
6298 params->avx512.shuffle512_mask[4] = 1; in xnn_init_f32_qu8_cvt_avx512_params()
6299 params->avx512.shuffle512_mask[5] = 5; in xnn_init_f32_qu8_cvt_avx512_params()
6300 params->avx512.shuffle512_mask[6] = 9; in xnn_init_f32_qu8_cvt_avx512_params()
6301 params->avx512.shuffle512_mask[7] = 13; in xnn_init_f32_qu8_cvt_avx512_params()
6302 params->avx512.shuffle512_mask[8] = 2; in xnn_init_f32_qu8_cvt_avx512_params()
6303 params->avx512.shuffle512_mask[9] = 6; in xnn_init_f32_qu8_cvt_avx512_params()
6304 params->avx512.shuffle512_mask[10] = 10; in xnn_init_f32_qu8_cvt_avx512_params()
6305 params->avx512.shuffle512_mask[11] = 14; in xnn_init_f32_qu8_cvt_avx512_params()
6306 params->avx512.shuffle512_mask[12] = 3; in xnn_init_f32_qu8_cvt_avx512_params()
6307 params->avx512.shuffle512_mask[13] = 7; in xnn_init_f32_qu8_cvt_avx512_params()
6308 params->avx512.shuffle512_mask[14] = 11; in xnn_init_f32_qu8_cvt_avx512_params()
6309 params->avx512.shuffle512_mask[15] = 15; in xnn_init_f32_qu8_cvt_avx512_params()
6310 params->avx512.shuffle256_mask[0] = 0; in xnn_init_f32_qu8_cvt_avx512_params()
6311 params->avx512.shuffle256_mask[1] = 4; in xnn_init_f32_qu8_cvt_avx512_params()
6312 params->avx512.shuffle256_mask[2] = 2; in xnn_init_f32_qu8_cvt_avx512_params()
6313 params->avx512.shuffle256_mask[3] = 6; in xnn_init_f32_qu8_cvt_avx512_params()
6314 params->avx512.shuffle256_mask[4] = 1; in xnn_init_f32_qu8_cvt_avx512_params()
6315 params->avx512.shuffle256_mask[5] = 5; in xnn_init_f32_qu8_cvt_avx512_params()
6316 params->avx512.shuffle256_mask[6] = 3; in xnn_init_f32_qu8_cvt_avx512_params()
6317 params->avx512.shuffle256_mask[7] = 7; in xnn_init_f32_qu8_cvt_avx512_params()
6318 return sizeof(params->avx512); in xnn_init_f32_qu8_cvt_avx512_params()
6324 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_wasmsimd_cvt_params()
6331 params->wasmsimd_cvt.scale[i] = scale; in xnn_init_f32_qu8_cvt_wasmsimd_cvt_params()
6334 params->wasmsimd_cvt.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_f32_qu8_cvt_wasmsimd_cvt_params()
6337 params->wasmsimd_cvt.output_min[i] = output_min; in xnn_init_f32_qu8_cvt_wasmsimd_cvt_params()
6338 params->wasmsimd_cvt.output_max[i] = output_max; in xnn_init_f32_qu8_cvt_wasmsimd_cvt_params()
6340 return sizeof(params->wasmsimd_cvt); in xnn_init_f32_qu8_cvt_wasmsimd_cvt_params()
6344 union xnn_f32_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_f32_qu8_cvt_wasmsimd_magic_params()
6354 params->wasmsimd_magic.scale[i] = scale; in xnn_init_f32_qu8_cvt_wasmsimd_magic_params()
6355 params->wasmsimd_magic.magic_bias[i] = 12582912.0f; in xnn_init_f32_qu8_cvt_wasmsimd_magic_params()
6356 params->wasmsimd_magic.magic_min[i] = magic_min; in xnn_init_f32_qu8_cvt_wasmsimd_magic_params()
6357 params->wasmsimd_magic.magic_bias_less_zero_point[i] = magic_bias_less_zero_point; in xnn_init_f32_qu8_cvt_wasmsimd_magic_params()
6360 params->wasmsimd_magic.output_max[i] = output_max; in xnn_init_f32_qu8_cvt_wasmsimd_magic_params()
6362 return sizeof(params->wasmsimd_magic); in xnn_init_f32_qu8_cvt_wasmsimd_magic_params()
6367 union xnn_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_cvt_scalar_params()
6378params->scalar.bias = ((int32_t) output_zero_point << 8) - (int32_t) multiplier * (int32_t) input_… in xnn_init_qs8_cvt_scalar_params()
6379 params->scalar.multiplier = (int32_t) multiplier; in xnn_init_qs8_cvt_scalar_params()
6380 return sizeof(params->scalar); in xnn_init_qs8_cvt_scalar_params()
6385 union xnn_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_cvt_armsimd32_params()
6397params->armsimd32.minus_input_zero_point = (uint32_t) minus_input_zero_point * UINT32_C(0x00010001… in xnn_init_qs8_cvt_armsimd32_params()
6398 params->armsimd32.multiplier = (int32_t) multiplier; in xnn_init_qs8_cvt_armsimd32_params()
6399 params->armsimd32.bias = ((int32_t) output_zero_point << 1) + INT32_C(1); in xnn_init_qs8_cvt_armsimd32_params()
6400 return sizeof(params->armsimd32); in xnn_init_qs8_cvt_armsimd32_params()
6406 union xnn_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_cvt_neon_params()
6417 params->neon.input_zero_point = (int16_t) input_zero_point; in xnn_init_qs8_cvt_neon_params()
6418 params->neon.multiplier = (int16_t) multiplier; in xnn_init_qs8_cvt_neon_params()
6419 params->neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qs8_cvt_neon_params()
6420 return sizeof(params->neon); in xnn_init_qs8_cvt_neon_params()
6426 union xnn_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_cvt_sse2_params()
6439 params->sse2.multiplier[i] = (int16_t) multiplier; in xnn_init_qs8_cvt_sse2_params()
6442 params->sse2.bias[i] = bias; in xnn_init_qs8_cvt_sse2_params()
6444 return sizeof(params->sse2); in xnn_init_qs8_cvt_sse2_params()
6448 union xnn_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_cvt_ssse3_params()
6460 params->ssse3.input_zero_point[i] = (int16_t) input_zero_point; in xnn_init_qs8_cvt_ssse3_params()
6461 params->ssse3.multiplier[i] = (int16_t) multiplier; in xnn_init_qs8_cvt_ssse3_params()
6462 params->ssse3.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_cvt_ssse3_params()
6464 return sizeof(params->ssse3); in xnn_init_qs8_cvt_ssse3_params()
6468 union xnn_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_cvt_avx2_params()
6480 params->avx2.input_zero_point[i] = (int16_t) input_zero_point; in xnn_init_qs8_cvt_avx2_params()
6481 params->avx2.multiplier[i] = (int16_t) multiplier; in xnn_init_qs8_cvt_avx2_params()
6482 params->avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_cvt_avx2_params()
6484 return sizeof(params->avx2); in xnn_init_qs8_cvt_avx2_params()
6490 union xnn_qs8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_cvt_wasmsimd_params()
6502 params->wasmsimd.input_zero_point[i] = (int16_t) input_zero_point; in xnn_init_qs8_cvt_wasmsimd_params()
6503 params->wasmsimd.multiplier[i] = (int16_t) multiplier; in xnn_init_qs8_cvt_wasmsimd_params()
6504 params->wasmsimd.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qs8_cvt_wasmsimd_params()
6506 return sizeof(params->wasmsimd); in xnn_init_qs8_cvt_wasmsimd_params()
6511 union xnn_qs8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_f32_cvt_scalar_params()
6515 params->scalar.zero_point = (int32_t) zero_point; in xnn_init_qs8_f32_cvt_scalar_params()
6516 params->scalar.scale = scale; in xnn_init_qs8_f32_cvt_scalar_params()
6517 return sizeof(params->scalar); in xnn_init_qs8_f32_cvt_scalar_params()
6522 union xnn_qs8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_f32_cvt_neon_params()
6526 params->neon.minus_zero_point[0] = -(int16_t) zero_point; in xnn_init_qs8_f32_cvt_neon_params()
6527 params->neon.minus_zero_point[1] = -(int16_t) zero_point; in xnn_init_qs8_f32_cvt_neon_params()
6528 params->neon.scale = scale; in xnn_init_qs8_f32_cvt_neon_params()
6529 return sizeof(params->neon); in xnn_init_qs8_f32_cvt_neon_params()
6535 union xnn_qs8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_f32_cvt_sse2_params()
6540 params->sse2.sign_mask[i] = UINT8_C(0x80); in xnn_init_qs8_f32_cvt_sse2_params()
6543 params->sse2.magic_exp[i] = UINT16_C(0x4B00); in xnn_init_qs8_f32_cvt_sse2_params()
6547 params->sse2.magic_bias[i] = magic_bias; in xnn_init_qs8_f32_cvt_sse2_params()
6548 params->sse2.scale[i] = scale; in xnn_init_qs8_f32_cvt_sse2_params()
6550 return sizeof(params->sse2); in xnn_init_qs8_f32_cvt_sse2_params()
6554 union xnn_qs8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_f32_cvt_sse4_params()
6559 params->sse4.minus_zero_point[i] = -(int32_t) zero_point; in xnn_init_qs8_f32_cvt_sse4_params()
6560 params->sse4.scale[i] = scale; in xnn_init_qs8_f32_cvt_sse4_params()
6562 return sizeof(params->sse4); in xnn_init_qs8_f32_cvt_sse4_params()
6566 union xnn_qs8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_f32_cvt_avx_params()
6571 params->avx.minus_zero_point[i] = -(int32_t) zero_point; in xnn_init_qs8_f32_cvt_avx_params()
6572 params->avx.scale[i] = scale; in xnn_init_qs8_f32_cvt_avx_params()
6574 return sizeof(params->avx); in xnn_init_qs8_f32_cvt_avx_params()
6578 union xnn_qs8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_f32_cvt_avx512_params()
6583 params->avx512.minus_zero_point[i] = -(int32_t) zero_point; in xnn_init_qs8_f32_cvt_avx512_params()
6584 params->avx512.scale[i] = scale; in xnn_init_qs8_f32_cvt_avx512_params()
6586 return sizeof(params->avx512); in xnn_init_qs8_f32_cvt_avx512_params()
6592 union xnn_qs8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qs8_f32_cvt_wasmsimd_params()
6597 params->wasmsimd.minus_zero_point[i] = -(int16_t) zero_point; in xnn_init_qs8_f32_cvt_wasmsimd_params()
6600 params->wasmsimd.scale[i] = scale; in xnn_init_qs8_f32_cvt_wasmsimd_params()
6602 return sizeof(params->wasmsimd); in xnn_init_qs8_f32_cvt_wasmsimd_params()
6607 union xnn_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_cvt_scalar_params()
6618params->scalar.bias = ((int32_t) output_zero_point << 8) - (int32_t) multiplier * (int32_t) input_… in xnn_init_qu8_cvt_scalar_params()
6619 params->scalar.multiplier = (int32_t) multiplier; in xnn_init_qu8_cvt_scalar_params()
6620 return sizeof(params->scalar); in xnn_init_qu8_cvt_scalar_params()
6625 union xnn_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_cvt_armsimd32_params()
6637params->armsimd32.minus_input_zero_point = (uint32_t) minus_input_zero_point * UINT32_C(0x00010001… in xnn_init_qu8_cvt_armsimd32_params()
6638 params->armsimd32.multiplier = (int32_t) multiplier; in xnn_init_qu8_cvt_armsimd32_params()
6639 params->armsimd32.bias = ((int32_t) output_zero_point << 1) + INT32_C(1); in xnn_init_qu8_cvt_armsimd32_params()
6640 return sizeof(params->armsimd32); in xnn_init_qu8_cvt_armsimd32_params()
6646 union xnn_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_cvt_neon_params()
6657 params->neon.input_zero_point = (uint16_t) input_zero_point; in xnn_init_qu8_cvt_neon_params()
6658 params->neon.multiplier = (int16_t) multiplier; in xnn_init_qu8_cvt_neon_params()
6659 params->neon.output_zero_point = (int16_t) output_zero_point; in xnn_init_qu8_cvt_neon_params()
6660 return sizeof(params->neon); in xnn_init_qu8_cvt_neon_params()
6666 union xnn_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_cvt_sse2_params()
6679 params->sse2.multiplier[i] = (uint16_t) multiplier; in xnn_init_qu8_cvt_sse2_params()
6682 params->sse2.bias[i] = bias; in xnn_init_qu8_cvt_sse2_params()
6684 return sizeof(params->sse2); in xnn_init_qu8_cvt_sse2_params()
6688 union xnn_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_cvt_ssse3_params()
6700 params->ssse3.input_zero_point[i] = (uint16_t) input_zero_point; in xnn_init_qu8_cvt_ssse3_params()
6701 params->ssse3.multiplier[i] = (int16_t) multiplier; in xnn_init_qu8_cvt_ssse3_params()
6702 params->ssse3.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qu8_cvt_ssse3_params()
6704 return sizeof(params->ssse3); in xnn_init_qu8_cvt_ssse3_params()
6708 union xnn_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_cvt_avx2_params()
6720 params->avx2.input_zero_point[i] = (uint16_t) input_zero_point; in xnn_init_qu8_cvt_avx2_params()
6721 params->avx2.multiplier[i] = (int16_t) multiplier; in xnn_init_qu8_cvt_avx2_params()
6722 params->avx2.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qu8_cvt_avx2_params()
6724 return sizeof(params->avx2); in xnn_init_qu8_cvt_avx2_params()
6730 union xnn_qu8_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_cvt_wasmsimd_params()
6742 params->wasmsimd.input_zero_point[i] = (uint16_t) input_zero_point; in xnn_init_qu8_cvt_wasmsimd_params()
6743 params->wasmsimd.multiplier[i] = (int16_t) multiplier; in xnn_init_qu8_cvt_wasmsimd_params()
6744 params->wasmsimd.output_zero_point[i] = (int16_t) output_zero_point; in xnn_init_qu8_cvt_wasmsimd_params()
6746 return sizeof(params->wasmsimd); in xnn_init_qu8_cvt_wasmsimd_params()
6751 union xnn_qu8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_f32_cvt_scalar_params()
6755 params->scalar.zero_point = (int32_t) zero_point; in xnn_init_qu8_f32_cvt_scalar_params()
6756 params->scalar.scale = scale; in xnn_init_qu8_f32_cvt_scalar_params()
6757 return sizeof(params->scalar); in xnn_init_qu8_f32_cvt_scalar_params()
6762 union xnn_qu8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_f32_cvt_neon_params()
6766 params->neon.minus_zero_point[0] = -(int16_t) zero_point; in xnn_init_qu8_f32_cvt_neon_params()
6767 params->neon.minus_zero_point[1] = -(int16_t) zero_point; in xnn_init_qu8_f32_cvt_neon_params()
6768 params->neon.scale = scale; in xnn_init_qu8_f32_cvt_neon_params()
6769 return sizeof(params->neon); in xnn_init_qu8_f32_cvt_neon_params()
6775 union xnn_qu8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_f32_cvt_sse2_params()
6780 params->sse2.magic_exp[i] = UINT16_C(0x4B00); in xnn_init_qu8_f32_cvt_sse2_params()
6784 params->sse2.magic_bias[i] = magic_bias; in xnn_init_qu8_f32_cvt_sse2_params()
6785 params->sse2.scale[i] = scale; in xnn_init_qu8_f32_cvt_sse2_params()
6787 return sizeof(params->sse2); in xnn_init_qu8_f32_cvt_sse2_params()
6791 union xnn_qu8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_f32_cvt_sse4_params()
6796 params->sse4.minus_zero_point[i] = -(int32_t) zero_point; in xnn_init_qu8_f32_cvt_sse4_params()
6797 params->sse4.scale[i] = scale; in xnn_init_qu8_f32_cvt_sse4_params()
6799 return sizeof(params->sse4); in xnn_init_qu8_f32_cvt_sse4_params()
6803 union xnn_qu8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_f32_cvt_avx_params()
6808 params->avx.minus_zero_point[i] = -(int32_t) zero_point; in xnn_init_qu8_f32_cvt_avx_params()
6809 params->avx.scale[i] = scale; in xnn_init_qu8_f32_cvt_avx_params()
6811 return sizeof(params->avx); in xnn_init_qu8_f32_cvt_avx_params()
6815 union xnn_qu8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_f32_cvt_avx512_params()
6820 params->avx512.minus_zero_point[i] = -(int32_t) zero_point; in xnn_init_qu8_f32_cvt_avx512_params()
6821 params->avx512.scale[i] = scale; in xnn_init_qu8_f32_cvt_avx512_params()
6823 return sizeof(params->avx512); in xnn_init_qu8_f32_cvt_avx512_params()
6829 union xnn_qu8_f32_cvt_params params[XNN_MIN_ELEMENTS(1)], in xnn_init_qu8_f32_cvt_wasmsimd_params()
6834 params->wasmsimd.minus_zero_point[i] = -(int16_t) zero_point; in xnn_init_qu8_f32_cvt_wasmsimd_params()
6837 params->wasmsimd.scale[i] = scale; in xnn_init_qu8_f32_cvt_wasmsimd_params()
6839 return sizeof(params->wasmsimd); in xnn_init_qu8_f32_cvt_wasmsimd_params()