Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x8c8-minmax-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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_4x8c8__wasmsdot(
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, 8 * 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 do {
68 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
69 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
70 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
71 const v128_t vinput_zero_point3 = wasm_v128_load32_splat(&quantization_params[3].zero_point);
72 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
73 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
74 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
75 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
76 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
77 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
78 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
79 v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
80 v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123);
81 v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123);
82 v128_t vsum3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
83 v128_t vacc3x01 = wasm_u64x2_extend_low_u32x4(vsum3x0123);
84 v128_t vacc3x23 = wasm_u64x2_extend_high_u32x4(vsum3x0123);
85 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
86 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
87 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
88 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
89 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1);
90 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
91 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
92 v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2);
93 v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567);
94 v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567);
95 v128_t vsum3x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point3);
96 v128_t vacc3x45 = wasm_u64x2_extend_low_u32x4(vsum3x4567);
97 v128_t vacc3x67 = wasm_u64x2_extend_high_u32x4(vsum3x4567);
98
99 size_t k = kc;
100 while (k != 0) {
101 const v128_t va0x01 = wasm_v128_load64_splat(a0);
102 a0 += 8;
103 const v128_t va1x01 = wasm_v128_load64_splat(a1);
104 a1 += 8;
105 const v128_t va2x01 = wasm_v128_load64_splat(a2);
106 a2 += 8;
107 const v128_t va3x01 = wasm_v128_load64_splat(a3);
108 a3 += 8;
109
110 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
111 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
112 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
113 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
114
115 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
116 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
117 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
118 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
119 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
120 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
121 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
122 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
123 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
124 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
125 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
126 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
127 vacc3x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va3x01, vacc3x01);
128 vacc3x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va3x01, vacc3x23);
129 vacc3x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va3x01, vacc3x45);
130 vacc3x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va3x01, vacc3x67);
131
132 k -= 8 * sizeof(int8_t);
133 }
134
135
136 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
137 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
138 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
139 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
140 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
141 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
142 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x01, vacc3x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x01, vacc3x23, 1, 3, 5, 7));
143 v128_t vacc3x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x45, vacc3x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc3x45, vacc3x67, 1, 3, 5, 7));
144
145 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
146 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
147 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
148 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
149 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
150 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
151 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
152 vacc3x4567 = wasm_f32x4_convert_i32x4(vacc3x4567);
153
154 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
155 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
156 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
157 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
158
159 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
160 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
161 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
162 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
163 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
164 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
165 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
166 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vinput_scale3);
167
168 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
169 w = (const float*) w + 4;
170 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
171 w = (const float*) w + 4;
172 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
173 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
174 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
175 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
176 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
177 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
178 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
179 vacc3x4567 = wasm_f32x4_mul(vacc3x4567, vfilter_output_scale4567);
180
181 const v128_t vbias0123 = wasm_v128_load(w);
182 w = (const float*) w + 4;
183 const v128_t vbias4567 = wasm_v128_load(w);
184 w = (const float*) w + 4;
185 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
186 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
187 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
188 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
189 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
190 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
191 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
192 vacc3x4567 = wasm_f32x4_add(vacc3x4567, vbias4567);
193
194 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
195 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
196 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
197 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
198 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
199 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
200 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
201 vacc3x4567 = wasm_f32x4_pmax(vacc3x4567, vmin);
202
203 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
204 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
205 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
206 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
207 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
208 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
209 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
210 vacc3x4567 = wasm_f32x4_pmin(vacc3x4567, vmax);
211
212 if XNN_LIKELY(nc >= 8) {
213 wasm_v128_store(c0, vacc0x0123);
214 wasm_v128_store(c0 + 4, vacc0x4567);
215 wasm_v128_store(c1, vacc1x0123);
216 wasm_v128_store(c1 + 4, vacc1x4567);
217 wasm_v128_store(c2, vacc2x0123);
218 wasm_v128_store(c2 + 4, vacc2x4567);
219 wasm_v128_store(c3, vacc3x0123);
220 wasm_v128_store(c3 + 4, vacc3x4567);
221
222 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
223 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
224 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
225 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
226
227 c0 = (float*) ((uintptr_t) c0 + cn_stride);
228 c1 = (float*) ((uintptr_t) c1 + cn_stride);
229 c2 = (float*) ((uintptr_t) c2 + cn_stride);
230 c3 = (float*) ((uintptr_t) c3 + cn_stride);
231
232 nc -= 8;
233 } else {
234 if (nc & 4) {
235 wasm_v128_store(c0, vacc0x0123);
236 vacc0x0123 = vacc0x4567;
237 c0 += 4;
238 wasm_v128_store(c1, vacc1x0123);
239 vacc1x0123 = vacc1x4567;
240 c1 += 4;
241 wasm_v128_store(c2, vacc2x0123);
242 vacc2x0123 = vacc2x4567;
243 c2 += 4;
244 wasm_v128_store(c3, vacc3x0123);
245 vacc3x0123 = vacc3x4567;
246 c3 += 4;
247 }
248 if (nc & 2) {
249 wasm_v128_store64_lane(c0, vacc0x0123, 0);
250 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
251 c0 += 2;
252 wasm_v128_store64_lane(c1, vacc1x0123, 0);
253 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
254 c1 += 2;
255 wasm_v128_store64_lane(c2, vacc2x0123, 0);
256 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
257 c2 += 2;
258 wasm_v128_store64_lane(c3, vacc3x0123, 0);
259 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
260 c3 += 2;
261 }
262 if (nc & 1) {
263 wasm_v128_store32_lane(c0, vacc0x0123, 0);
264 wasm_v128_store32_lane(c1, vacc1x0123, 0);
265 wasm_v128_store32_lane(c2, vacc2x0123, 0);
266 wasm_v128_store32_lane(c3, vacc3x0123, 0);
267 }
268 nc = 0;
269 }
270 } while (nc != 0);
271 }