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_ld128(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_ld128(
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 vb01 = wasm_v128_load(w);
98 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb01), vb_zero_point);
99 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb01), vb_zero_point);
100
101 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
102 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
103 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
104 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
105 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
106 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
107 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
108 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
109
110 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
111 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
112 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
113 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
114 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
115 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
116 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
117 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
118 const v128_t vb23 = wasm_v128_load((const uint8_t*) w + 16);
119 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_extend_low_u8x16(vb23), vb_zero_point);
120 const v128_t vxb3 = wasm_i16x8_sub(wasm_u16x8_extend_high_u8x16(vb23), vb_zero_point);
121
122 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
123 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
124 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
125 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
126 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
127 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
128 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
129 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
130
131 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
132 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
133 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
134 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
135 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
136 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
137 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
138 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 3, 3, 3, 3), vxb3));
139
140 w = (const void*) ((const uint8_t*) w + 32);
141 k -= 8 * sizeof(uint8_t);
142 }
143 if (k != 0) {
144 const v128_t vxa0 = wasm_u16x8_load8x8(a0);
145 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
146 const v128_t vxa1 = wasm_u16x8_load8x8(a1);
147 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
148 const v128_t vxa2 = wasm_u16x8_load8x8(a2);
149 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
150 const v128_t vxa3 = wasm_u16x8_load8x8(a3);
151 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
152
153 const v128_t vxb0 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
154 w = (const void*) ((const uint8_t*) w + 8);
155
156 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
157 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
158 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
159 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
160 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
161 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
162 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
163 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
164
165 if (k > 2 * sizeof(uint8_t)) {
166 const v128_t vxb1 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
167 w = (const void*) ((const uint8_t*) w + 8);
168
169 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
170 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
171 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
172 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
173 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
174 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
175 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
176 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
177
178 if (k > 4 * sizeof(uint8_t)) {
179 const v128_t vxb2 = wasm_i16x8_sub(wasm_u16x8_load8x8(w), vb_zero_point);
180 w = (const void*) ((const uint8_t*) w + 8);
181
182 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
183 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
184 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
185 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
186 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
187 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
188 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
189 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
190 }
191 }
192 }
193 p -= 4 * sizeof(void*);
194 } while (p != 0);
195
196 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
197 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
198 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
199 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
200
201 const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
202 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
203 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
204 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vscale);
205 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vscale);
206
207 const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
208 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
209 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
210 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vmagic_bias);
211 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vmagic_bias);
212
213 const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
214 vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
215 vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
216 vacc2x0123 = wasm_i32x4_max(vacc2x0123, vmagic_min);
217 vacc3x0123 = wasm_i32x4_max(vacc3x0123, vmagic_min);
218
219 const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
220 vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
221 vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
222 vacc2x0123 = wasm_i32x4_sub(vacc2x0123, vmagic_bias_less_output_zero_point);
223 vacc3x0123 = wasm_i32x4_sub(vacc3x0123, vmagic_bias_less_output_zero_point);
224
225 v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
226 v128_t vacc23x0123 = wasm_i16x8_narrow_i32x4(vacc2x0123, vacc3x0123);
227
228 v128_t vout = wasm_u8x16_narrow_i16x8(vacc01x0123, vacc23x0123);
229
230 const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
231 vout = wasm_u8x16_min(vout, voutput_max);
232
233 if (nc >= 4) {
234 *((float*) c3) = (float) wasm_f32x4_extract_lane(vout, 3);
235 *((float*) c2) = (float) wasm_f32x4_extract_lane(vout, 2);
236 *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
237 *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
238
239 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
240 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
241 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
242 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
243
244 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
245
246 nc -= 4;
247 } else {
248 uint32_t vout3 = wasm_i32x4_extract_lane(vout, 3);
249 uint32_t vout2 = wasm_i32x4_extract_lane(vout, 2);
250 uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
251 uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
252 if (nc & 2) {
253 *((uint16_t*) c3) = (uint16_t) vout3;
254 vout3 >>= 16;
255 c3 += 2;
256 *((uint16_t*) c2) = (uint16_t) vout2;
257 vout2 >>= 16;
258 c2 += 2;
259 *((uint16_t*) c1) = (uint16_t) vout1;
260 vout1 >>= 16;
261 c1 += 2;
262 *((uint16_t*) c0) = (uint16_t) vout0;
263 vout0 >>= 16;
264 c0 += 2;
265 }
266 if (nc & 1) {
267 *c3 = (uint8_t) vout3;
268 *c2 = (uint8_t) vout2;
269 *c1 = (uint8_t) vout1;
270 *c0 = (uint8_t) vout0;
271 }
272
273 nc = 0;
274 }
275 } while (nc != 0);
276 }
277