Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x4c2-minmax-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x4c2__wasmsimd_dot16x2_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61
62 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
63 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
64 XNN_FORCE_REALIZATION(vmin);
65 XNN_FORCE_REALIZATION(vmax);
66
67
68 do {
69 const v128_t vksum0123 = wasm_v128_load(w);
70 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
71 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
72 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
73 const v128_t vinput_zero_point3 = wasm_v128_load32_splat(&quantization_params[3].zero_point);
74 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
75 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
76 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
77 v128_t vacc3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
78 w = (const int32_t*) w + 4;
79
80 size_t k = kc;
81 while (k >= 8 * sizeof(int8_t)) {
82 const v128_t vxa0 = wasm_i16x8_load8x8((const v128_t*) a0);
83 a0 += 8;
84 const v128_t vxa1 = wasm_i16x8_load8x8((const v128_t*) a1);
85 a1 += 8;
86 const v128_t vxa2 = wasm_i16x8_load8x8((const v128_t*) a2);
87 a2 += 8;
88 const v128_t vxa3 = wasm_i16x8_load8x8((const v128_t*) a3);
89 a3 += 8;
90
91 const v128_t vb01 = wasm_v128_load(w);
92 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
93 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
94
95 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
96 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
97 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
98 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
99 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
100 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
101 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
102 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
103
104 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
105 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
106 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
107 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
108 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
109 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
110 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
111 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
112 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
113 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
114 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
115
116 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
117 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
118 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
119 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
120 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
121 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
122 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
123 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
124
125 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
126 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 3, 3, 3, 3), vxb3));
127 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
128 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 3, 3, 3, 3), vxb3));
129 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
130 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 3, 3, 3, 3), vxb3));
131 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
132 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 3, 3, 3, 3), vxb3));
133
134 w = (const int8_t*) w + 32;
135 k -= 8 * sizeof(int8_t);
136 }
137 if (k != 0) {
138 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
139 a0 = (const int8_t*) ((uintptr_t) a0 + k);
140 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
141 a1 = (const int8_t*) ((uintptr_t) a1 + k);
142 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
143 a2 = (const int8_t*) ((uintptr_t) a2 + k);
144 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
145 a3 = (const int8_t*) ((uintptr_t) a3 + k);
146
147 const v128_t vxb0 = wasm_i16x8_load8x8(w);
148 w = (const int8_t*) w + 8;
149
150 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
151 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 0, 0, 0, 0), vxb0));
152 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
153 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 0, 0, 0, 0), vxb0));
154 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
155 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 0, 0, 0, 0), vxb0));
156 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
157 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 0, 0, 0, 0), vxb0));
158
159 if (k > 2 * sizeof(int8_t)) {
160 const v128_t vxb1 = wasm_i16x8_load8x8(w);
161 w = (const int8_t*) w + 8;
162
163 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
164 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 1, 1, 1, 1), vxb1));
165 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
166 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 1, 1, 1, 1), vxb1));
167 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
168 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 1, 1, 1, 1), vxb1));
169 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
170 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 1, 1, 1, 1), vxb1));
171
172 if (k > 4 * sizeof(int8_t)) {
173 const v128_t vxb2 = wasm_i16x8_load8x8(w);
174 w = (const int8_t*) w + 8;
175
176 vacc0x0123 = wasm_i32x4_add(vacc0x0123,
177 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa0, vxa0, 2, 2, 2, 2), vxb2));
178 vacc1x0123 = wasm_i32x4_add(vacc1x0123,
179 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa1, vxa1, 2, 2, 2, 2), vxb2));
180 vacc2x0123 = wasm_i32x4_add(vacc2x0123,
181 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa2, vxa2, 2, 2, 2, 2), vxb2));
182 vacc3x0123 = wasm_i32x4_add(vacc3x0123,
183 wasm_i32x4_dot_i16x8(wasm_v32x4_shuffle(vxa3, vxa3, 2, 2, 2, 2), vxb2));
184 }
185 }
186 }
187
188 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
189 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
190 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
191 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
192
193 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
194 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
195 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
196 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
197
198 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
199 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
200 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
201 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
202
203 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
204 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
205 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
206 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
207 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
208 w = (const float*) w + 4;
209
210 const v128_t vbias0123 = wasm_v128_load(w);
211 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
212 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
213 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
214 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
215 w = (const float*) w + 4;
216
217 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
218 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
219 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
220 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
221
222 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
223 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
224 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
225 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
226
227 if XNN_LIKELY(nc >= 4) {
228 wasm_v128_store(c0, vacc0x0123);
229 wasm_v128_store(c1, vacc1x0123);
230 wasm_v128_store(c2, vacc2x0123);
231 wasm_v128_store(c3, vacc3x0123);
232
233 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
234 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
235 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
236 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
237
238 c0 = (float*) ((uintptr_t) c0 + cn_stride);
239 c1 = (float*) ((uintptr_t) c1 + cn_stride);
240 c2 = (float*) ((uintptr_t) c2 + cn_stride);
241 c3 = (float*) ((uintptr_t) c3 + cn_stride);
242
243 nc -= 4;
244 } else {
245 if (nc & 2) {
246 wasm_v128_store64_lane(c0, vacc0x0123, 0);
247 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
248 c0 += 2;
249 wasm_v128_store64_lane(c1, vacc1x0123, 0);
250 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
251 c1 += 2;
252 wasm_v128_store64_lane(c2, vacc2x0123, 0);
253 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
254 c2 += 2;
255 wasm_v128_store64_lane(c3, vacc3x0123, 0);
256 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
257 c3 += 2;
258 }
259 if (nc & 1) {
260 wasm_v128_store32_lane(c0, vacc0x0123, 0);
261 wasm_v128_store32_lane(c1, vacc1x0123, 0);
262 wasm_v128_store32_lane(c2, vacc2x0123, 0);
263 wasm_v128_store32_lane(c3, vacc3x0123, 0);
264 }
265 nc = 0;
266 }
267 } while (nc != 0);
268 }