Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8c8-minmax-wasmsdot-u2.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_3x8c8__wasmsdot_u2(
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 <= 3);
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
56 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
57 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
58 XNN_FORCE_REALIZATION(vmin);
59 XNN_FORCE_REALIZATION(vmax);
60
61 do {
62 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
63 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
64 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
65 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
66 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
67 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
68 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
69 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
70 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
71 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
72 v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
73 v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123);
74 v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123);
75 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
76 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
77 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
78 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
79 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1);
80 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
81 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
82 v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2);
83 v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567);
84 v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567);
85
86 size_t k = kc;
87 while (k >= 16 * sizeof(int8_t)) {
88 const v128_t va0x01 = wasm_v128_load64_splat(a0);
89 const v128_t va0x23 = wasm_v128_load64_splat((const int8_t*) a0 + 8);
90 a0 += 16;
91 const v128_t va1x01 = wasm_v128_load64_splat(a1);
92 const v128_t va1x23 = wasm_v128_load64_splat((const int8_t*) a1 + 8);
93 a1 += 16;
94 const v128_t va2x01 = wasm_v128_load64_splat(a2);
95 const v128_t va2x23 = wasm_v128_load64_splat((const int8_t*) a2 + 8);
96 a2 += 16;
97
98 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
99 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
100 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
101 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
102 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
103 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
104 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
105 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
106
107 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
108 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
109 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
110 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
111 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
112 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
113 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
114 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
115 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
116 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
117 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
118 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
119 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
120 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
121 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
122 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
123 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
124 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
125 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
126 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
127 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va2x23, vacc2x01);
128 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va2x23, vacc2x23);
129 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va2x23, vacc2x45);
130 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va2x23, vacc2x67);
131
132 k -= 16 * sizeof(int8_t);
133 }
134
135 if (k != 0) {
136 const v128_t va0x01 = wasm_v128_load64_splat(a0);
137 a0 += 8;
138 const v128_t va1x01 = wasm_v128_load64_splat(a1);
139 a1 += 8;
140 const v128_t va2x01 = wasm_v128_load64_splat(a2);
141 a2 += 8;
142
143 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
144 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
145 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
146 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
147
148 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
149 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
150 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
151 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
152 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
153 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
154 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
155 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
156 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va2x01, vacc2x01);
157 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va2x01, vacc2x23);
158 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va2x01, vacc2x45);
159 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va2x01, vacc2x67);
160
161 k -= 8 * sizeof(int8_t);
162 }
163
164 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
165 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
166 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
167 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
168 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
169 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
170
171 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
172 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
173 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
174 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
175 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
176 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
177
178 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
179 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
180 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
181
182 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
183 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
184 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
185 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
186 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
187 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
188
189 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
190 w = (const float*) w + 4;
191 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
192 w = (const float*) w + 4;
193 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
194 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
195 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
196 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
197 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
198 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
199
200 const v128_t vbias0123 = wasm_v128_load(w);
201 w = (const float*) w + 4;
202 const v128_t vbias4567 = wasm_v128_load(w);
203 w = (const float*) w + 4;
204 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
205 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
206 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
207 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
208 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
209 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
210
211 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
212 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
213 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
214 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
215 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
216 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
217
218 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
219 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
220 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
221 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
222 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
223 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
224
225 if XNN_LIKELY(nc >= 8) {
226 wasm_v128_store(c0, vacc0x0123);
227 wasm_v128_store(c0 + 4, vacc0x4567);
228 wasm_v128_store(c1, vacc1x0123);
229 wasm_v128_store(c1 + 4, vacc1x4567);
230 wasm_v128_store(c2, vacc2x0123);
231 wasm_v128_store(c2 + 4, vacc2x4567);
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
237 c0 = (float*) ((uintptr_t) c0 + cn_stride);
238 c1 = (float*) ((uintptr_t) c1 + cn_stride);
239 c2 = (float*) ((uintptr_t) c2 + cn_stride);
240
241 nc -= 8;
242 } else {
243 if (nc & 4) {
244 wasm_v128_store(c0, vacc0x0123);
245 vacc0x0123 = vacc0x4567;
246 c0 += 4;
247 wasm_v128_store(c1, vacc1x0123);
248 vacc1x0123 = vacc1x4567;
249 c1 += 4;
250 wasm_v128_store(c2, vacc2x0123);
251 vacc2x0123 = vacc2x4567;
252 c2 += 4;
253 }
254 if (nc & 2) {
255 wasm_v128_store64_lane(c0, vacc0x0123, 0);
256 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
257 c0 += 2;
258 wasm_v128_store64_lane(c1, vacc1x0123, 0);
259 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
260 c1 += 2;
261 wasm_v128_store64_lane(c2, vacc2x0123, 0);
262 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
263 c2 += 2;
264 }
265 if (nc & 1) {
266 wasm_v128_store32_lane(c0, vacc0x0123, 0);
267 wasm_v128_store32_lane(c1, vacc1x0123, 0);
268 wasm_v128_store32_lane(c2, vacc2x0123, 0);
269 }
270 nc = 0;
271 }
272 } while (nc != 0);
273 }