Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8c8-minmax-wasmusdot-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__wasmusdot_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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
62 do {
63 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
64 const v128_t vinput_zero_point1 = wasm_i32x4_splat((int32_t) quantization_params[1].zero_point + 128);
65 const v128_t vinput_zero_point2 = wasm_i32x4_splat((int32_t) quantization_params[2].zero_point + 128);
66 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
67 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
68 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
69 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
70 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
71 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
72 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
73 v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
74 v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123);
75 v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123);
76 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
77 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
78 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
79 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
80 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point1);
81 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
82 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
83 v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point2);
84 v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567);
85 v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567);
86
87 size_t k = kc;
88 while (k >= 16 * sizeof(int8_t)) {
89 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
90 const v128_t va0x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a0 + 8), vsign_mask);
91 a0 += 16;
92 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
93 const v128_t va1x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a1 + 8), vsign_mask);
94 a1 += 16;
95 const v128_t va2x01 = wasm_v128_xor(wasm_v128_load64_splat(a2), vsign_mask);
96 const v128_t va2x23 = wasm_v128_xor(wasm_v128_load64_splat((const int8_t*) a2 + 8), vsign_mask);
97 a2 += 16;
98
99 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
100 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
101 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
102 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
103 const v128_t vb01x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
104 const v128_t vb23x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
105 const v128_t vb45x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
106 const v128_t vb67x23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
107
108 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
109 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
110 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
111 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
112 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
113 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
114 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
115 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
116 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
117 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
118 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
119 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
120 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va0x23, vacc0x01);
121 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va0x23, vacc0x23);
122 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va0x23, vacc0x45);
123 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va0x23, vacc0x67);
124 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va1x23, vacc1x01);
125 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va1x23, vacc1x23);
126 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va1x23, vacc1x45);
127 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va1x23, vacc1x67);
128 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x23, va2x23, vacc2x01);
129 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x23, va2x23, vacc2x23);
130 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x23, va2x23, vacc2x45);
131 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x23, va2x23, vacc2x67);
132
133 k -= 16 * sizeof(int8_t);
134 }
135
136 if (k != 0) {
137 const v128_t va0x01 = wasm_v128_xor(wasm_v128_load64_splat(a0), vsign_mask);
138 a0 += 8;
139 const v128_t va1x01 = wasm_v128_xor(wasm_v128_load64_splat(a1), vsign_mask);
140 a1 += 8;
141 const v128_t va2x01 = wasm_v128_xor(wasm_v128_load64_splat(a2), vsign_mask);
142 a2 += 8;
143
144 const v128_t vb01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
145 const v128_t vb23 = wasm_v128_load(w); w = (const int8_t*) w + 16;
146 const v128_t vb45 = wasm_v128_load(w); w = (const int8_t*) w + 16;
147 const v128_t vb67 = wasm_v128_load(w); w = (const int8_t*) w + 16;
148
149 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va0x01, vacc0x01);
150 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va0x01, vacc0x23);
151 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va0x01, vacc0x45);
152 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va0x01, vacc0x67);
153 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va1x01, vacc1x01);
154 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va1x01, vacc1x23);
155 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va1x01, vacc1x45);
156 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va1x01, vacc1x67);
157 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01, va2x01, vacc2x01);
158 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23, va2x01, vacc2x23);
159 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45, va2x01, vacc2x45);
160 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67, va2x01, vacc2x67);
161
162 k -= 8 * sizeof(int8_t);
163 }
164
165 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
166 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
167 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
168 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
169 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
170 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
171
172 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
173 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
174 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
175 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
176 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
177 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
178
179 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
180 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
181 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
182
183 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
184 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
185 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
186 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
187 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
188 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale2);
189
190 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
191 w = (const float*) w + 4;
192 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
193 w = (const float*) w + 4;
194 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
195 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
196 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
197 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
198 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
199 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
200
201 const v128_t vbias0123 = wasm_v128_load(w);
202 w = (const float*) w + 4;
203 const v128_t vbias4567 = wasm_v128_load(w);
204 w = (const float*) w + 4;
205 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
206 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
207 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
208 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
209 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
210 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
211
212 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
213 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
214 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
215 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
216 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
217 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
218
219 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
220 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
221 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
222 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
223 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
224 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
225
226 if XNN_LIKELY(nc >= 8) {
227 wasm_v128_store(c0, vacc0x0123);
228 wasm_v128_store(c0 + 4, vacc0x4567);
229 wasm_v128_store(c1, vacc1x0123);
230 wasm_v128_store(c1 + 4, vacc1x4567);
231 wasm_v128_store(c2, vacc2x0123);
232 wasm_v128_store(c2 + 4, vacc2x4567);
233
234 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
235 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
236 a2 = (const int8_t*) ((uintptr_t) a2 - 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
242 nc -= 8;
243 } else {
244 if (nc & 4) {
245 wasm_v128_store(c0, vacc0x0123);
246 vacc0x0123 = vacc0x4567;
247 c0 += 4;
248 wasm_v128_store(c1, vacc1x0123);
249 vacc1x0123 = vacc1x4567;
250 c1 += 4;
251 wasm_v128_store(c2, vacc2x0123);
252 vacc2x0123 = vacc2x4567;
253 c2 += 4;
254 }
255 if (nc & 2) {
256 wasm_v128_store64_lane(c0, vacc0x0123, 0);
257 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
258 c0 += 2;
259 wasm_v128_store64_lane(c1, vacc1x0123, 0);
260 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
261 c1 += 2;
262 wasm_v128_store64_lane(c2, vacc2x0123, 0);
263 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
264 c2 += 2;
265 }
266 if (nc & 1) {
267 wasm_v128_store32_lane(c0, vacc0x0123, 0);
268 wasm_v128_store32_lane(c1, vacc1x0123, 0);
269 wasm_v128_store32_lane(c2, vacc2x0123, 0);
270 }
271 nc = 0;
272 }
273 } while (nc != 0);
274 }