Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x8c8-minmax-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c8-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_igemm_minmax_ukernel_3x8c8__wasmsdot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 float* c0 = c;
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
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_point = wasm_v128_load32_splat(&quantization_params->zero_point);
63 v128_t vksum0123 = wasm_v128_load(w); w = (const int32_t*) w + 4;
64 v128_t vsum0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
65 v128_t vacc0x01 = wasm_u64x2_extend_low_u32x4(vsum0x0123);
66 v128_t vacc0x23 = wasm_u64x2_extend_high_u32x4(vsum0x0123);
67 v128_t vsum1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
68 v128_t vacc1x01 = wasm_u64x2_extend_low_u32x4(vsum1x0123);
69 v128_t vacc1x23 = wasm_u64x2_extend_high_u32x4(vsum1x0123);
70 v128_t vsum2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
71 v128_t vacc2x01 = wasm_u64x2_extend_low_u32x4(vsum2x0123);
72 v128_t vacc2x23 = wasm_u64x2_extend_high_u32x4(vsum2x0123);
73 v128_t vksum4567 = wasm_v128_load(w); w = (const int32_t*) w + 4;
74 v128_t vsum0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
75 v128_t vacc0x45 = wasm_u64x2_extend_low_u32x4(vsum0x4567);
76 v128_t vacc0x67 = wasm_u64x2_extend_high_u32x4(vsum0x4567);
77 v128_t vsum1x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
78 v128_t vacc1x45 = wasm_u64x2_extend_low_u32x4(vsum1x4567);
79 v128_t vacc1x67 = wasm_u64x2_extend_high_u32x4(vsum1x4567);
80 v128_t vsum2x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
81 v128_t vacc2x45 = wasm_u64x2_extend_low_u32x4(vsum2x4567);
82 v128_t vacc2x67 = wasm_u64x2_extend_high_u32x4(vsum2x4567);
83
84 size_t p = ks;
85 do {
86 const int8_t* restrict a0 = a[0];
87 if XNN_UNPREDICTABLE(a0 != zero) {
88 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
89 } else {
90 a0 = zero_data;
91 }
92 const int8_t* restrict a1 = a[1];
93 if XNN_UNPREDICTABLE(a1 != zero) {
94 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
95 } else {
96 a1 = zero_data;
97 }
98 const int8_t* restrict a2 = a[2];
99 if XNN_UNPREDICTABLE(a2 != zero) {
100 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
101 } else {
102 a2 = zero_data;
103 }
104 a += 3;
105
106 size_t k = kc;
107 while (k != 0) {
108 const v128_t va0x01 = wasm_v128_load64_splat(a0);
109 a0 += 8;
110 const v128_t va1x01 = wasm_v128_load64_splat(a1);
111 a1 += 8;
112 const v128_t va2x01 = wasm_v128_load64_splat(a2);
113 a2 += 8;
114
115 const v128_t vb01x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
116 const v128_t vb23x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
117 const v128_t vb45x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
118 const v128_t vb67x01 = wasm_v128_load(w); w = (const int8_t*) w + 16;
119
120 vacc0x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va0x01, vacc0x01);
121 vacc0x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va0x01, vacc0x23);
122 vacc0x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va0x01, vacc0x45);
123 vacc0x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va0x01, vacc0x67);
124 vacc1x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va1x01, vacc1x01);
125 vacc1x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va1x01, vacc1x23);
126 vacc1x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va1x01, vacc1x45);
127 vacc1x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va1x01, vacc1x67);
128 vacc2x01 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb01x01, va2x01, vacc2x01);
129 vacc2x23 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb23x01, va2x01, vacc2x23);
130 vacc2x45 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb45x01, va2x01, vacc2x45);
131 vacc2x67 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb67x01, va2x01, vacc2x67);
132
133 k -= 8 * sizeof(int8_t);
134 }
135
136 p -= 3 * sizeof(void*);
137 } while (p != 0);
138
139 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x01, vacc0x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x01, vacc0x23, 1, 3, 5, 7));
140 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x45, vacc0x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc0x45, vacc0x67, 1, 3, 5, 7));
141 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x01, vacc1x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x01, vacc1x23, 1, 3, 5, 7));
142 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x45, vacc1x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc1x45, vacc1x67, 1, 3, 5, 7));
143 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x01, vacc2x23, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x01, vacc2x23, 1, 3, 5, 7));
144 v128_t vacc2x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x45, vacc2x67, 0, 2, 4, 6), wasm_v32x4_shuffle(vacc2x45, vacc2x67, 1, 3, 5, 7));
145
146 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
147 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
148 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
149 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
150 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
151 vacc2x4567 = wasm_f32x4_convert_i32x4(vacc2x4567);
152
153 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
154
155 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
156 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale);
157 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
158 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale);
159 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
160 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vinput_scale);
161
162 const v128_t vfilter_output_scale0123 = wasm_v128_load(w); w = (const float*) w + 4;
163 const v128_t vfilter_output_scale4567 = wasm_v128_load(w); w = (const float*) w + 4;
164
165 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
166 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
167 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
168 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
169 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
170 vacc2x4567 = wasm_f32x4_mul(vacc2x4567, vfilter_output_scale4567);
171
172 const v128_t vbias0123 = wasm_v128_load(w); w = (const float*) w + 4;
173 const v128_t vbias4567 = wasm_v128_load(w); w = (const float*) w + 4;
174 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
175 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
176 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
177 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
178 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
179 vacc2x4567 = wasm_f32x4_add(vacc2x4567, vbias4567);
180
181 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
182 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
183 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
184 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
185 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
186 vacc2x4567 = wasm_f32x4_pmax(vacc2x4567, vmin);
187
188 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
189 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
190 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
191 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
192 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
193 vacc2x4567 = wasm_f32x4_pmin(vacc2x4567, vmax);
194
195 if XNN_LIKELY(nc >= 8) {
196 wasm_v128_store(c2, vacc2x0123);
197 wasm_v128_store(c2 + 4, vacc2x4567);
198 wasm_v128_store(c1, vacc1x0123);
199 wasm_v128_store(c1 + 4, vacc1x4567);
200 wasm_v128_store(c0, vacc0x0123);
201 wasm_v128_store(c0 + 4, vacc0x4567);
202
203 c0 = (float*) ((uintptr_t) c0 + cn_stride);
204 c1 = (float*) ((uintptr_t) c1 + cn_stride);
205 c2 = (float*) ((uintptr_t) c2 + cn_stride);
206
207 a = (const int8_t**restrict) ((uintptr_t) a - ks);
208
209 nc -= 8;
210 } else {
211 if (nc & 4) {
212 wasm_v128_store(c2, vacc2x0123);
213 vacc2x0123 = vacc2x4567;
214 c2 += 4;
215 wasm_v128_store(c1, vacc1x0123);
216 vacc1x0123 = vacc1x4567;
217 c1 += 4;
218 wasm_v128_store(c0, vacc0x0123);
219 vacc0x0123 = vacc0x4567;
220 c0 += 4;
221 }
222 if (nc & 2) {
223 wasm_v128_store64_lane(c2, vacc2x0123, 0);
224 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
225 c2 += 2;
226 wasm_v128_store64_lane(c1, vacc1x0123, 0);
227 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
228 c1 += 2;
229 wasm_v128_store64_lane(c0, vacc0x0123, 0);
230 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
231 c0 += 2;
232 }
233 if (nc & 1) {
234 wasm_v128_store32_lane(c2, vacc2x0123, 0);
235 wasm_v128_store32_lane(c1, vacc1x0123, 0);
236 wasm_v128_store32_lane(c0, vacc0x0123, 0);
237 }
238 nc = 0;
239 }
240 } while (nc != 0);
241 }