1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-wasmsimd-dot16x2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <wasm_simd128.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2__wasmsimd_dot16x2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (4 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 2);
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr != 4) {
55 c3 = c2;
56 }
57
58 do {
59 v128_t vacc0x0123 = wasm_v128_load(w);
60 v128_t vacc1x0123 = vacc0x0123;
61 v128_t vacc2x0123 = vacc0x0123;
62 v128_t vacc3x0123 = vacc0x0123;
63 w = (const void*) ((const int32_t*) w + 4);
64
65 size_t p = ks;
66 do {
67 const uint8_t* restrict a0 = a[0];
68 if XNN_UNPREDICTABLE(a0 != zero) {
69 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
70 }
71 const uint8_t* restrict a1 = a[1];
72 if XNN_UNPREDICTABLE(a1 != zero) {
73 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
74 }
75 const uint8_t* restrict a2 = a[2];
76 if XNN_UNPREDICTABLE(a2 != zero) {
77 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
78 }
79 const uint8_t* restrict a3 = a[3];
80 if XNN_UNPREDICTABLE(a3 != zero) {
81 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
82 }
83 a += 4;
84
85 size_t k = kc;
86 const v128_t vb_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.kernel_zero_point);
87 while (k >= 8 * sizeof(uint8_t)) {
88 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
89 a0 += 8;
90 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
91 a1 += 8;
92 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
93 a2 += 8;
94 const v128_t vxa3 = wasm_u16x8_load8x8(a3);
95 a3 += 8;
96
97 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
98
99 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
100 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
101 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
102 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
103 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
104 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
105 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
106 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
107 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 8), vb_zero_point);
108
109 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
110 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
111 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
112 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
113 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
114 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
115 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
116 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
117 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 16), vb_zero_point);
118
119 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
120 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
121 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
122 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
123 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
124 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
125 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
126 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
127 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_load8x8((const uint8_t*) w + 24), vb_zero_point);
128
129 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
130 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
131 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
132 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
133 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
134 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
135 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
136 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 3, 3, 3, 3), vxb3));
137
138 w = (const void*) ((const uint8_t*) w + 32);
139 k -= 8 * sizeof(uint8_t);
140 }
141 if (k != 0) {
142 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
143 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
144 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
145 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
146 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
147 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
148 const v128_t vxa3 = wasm_u16x8_load8x8(a3);
149 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
150
151 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
152 w = (const void*) ((const uint8_t*) w + 8);
153
154 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
155 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
156 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
157 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
158 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
159 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
160 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
161 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
162
163 if (k > 2 * sizeof(uint8_t)) {
164 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
165 w = (const void*) ((const uint8_t*) w + 8);
166
167 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
168 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
169 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
170 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
171 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
172 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
173 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
174 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
175
176 if (k > 4 * sizeof(uint8_t)) {
177 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
178 w = (const void*) ((const uint8_t*) w + 8);
179
180 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
181 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
182 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
183 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
184 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
185 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
186 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
187 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
188 }
189 }
190 }
191 p -= 4 * sizeof(void*);
192 } while (p != 0);
193
194 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
195 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
196 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
197 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
198
199 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
200 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
201 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
202 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
203 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale);
204
205 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
206 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
207 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
208 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
209 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
210
211 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
212 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
213 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
214 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
215 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
216
217 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
218 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
219 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
220 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
221 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
222
223 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
224 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
225
226 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
227
228 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
229 vout = wasm_u8x16_min(vout, voutput_max);
230
231 if (nc >= 4) {
232 *((float*) c3) = (float) wasm_f32x4_extract_lane(vout, 3);
233 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
234 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
235 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
236
237 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
238 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
239 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
240 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
241
242 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
243
244 nc -= 4;
245 } else {
246 uint32_t vout3 = wasm_i32x4_extract_lane(vout, 3);
247 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
248 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
249 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
250 if (nc & 2) {
251 *((uint16_t*) c3) = (uint16_t) vout3;
252 vout3 >>= 16;
253 c3 += 2;
254 *((uint16_t*) c2) = (uint16_t) vout2;
255 vout2 >>= 16;
256 c2 += 2;
257 *((uint16_t*) c1) = (uint16_t) vout1;
258 vout1 >>= 16;
259 c1 += 2;
260 *((uint16_t*) c0) = (uint16_t) vout0;
261 vout0 >>= 16;
262 c0 += 2;
263 }
264 if (nc & 1) {
265 *c3 = (uint8_t) vout3;
266 *c2 = (uint8_t) vout2;
267 *c1 = (uint8_t) vout1;
268 *c0 = (uint8_t) vout0;
269 }
270
271 nc = 0;
272 }
273 } while (nc != 0);
274 }
275