Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x8c16-minmax-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c16-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 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_2x8c16__wasmsdot(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
28 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 2);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 16 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 float* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr != 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48
49 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
50 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
51 XNN_FORCE_REALIZATION(vmin);
52 XNN_FORCE_REALIZATION(vmax);
53
54 do {
55 v128_t vksum0 = wasm_v128_load32_zero(w);
56 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
57 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
58 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
59 v128_t vksum4 = wasm_v128_load32_zero((const int32_t*) w + 4);
60 v128_t vksum5 = wasm_v128_load32_zero((const int32_t*) w + 5);
61 v128_t vksum6 = wasm_v128_load32_zero((const int32_t*) w + 6);
62 v128_t vksum7 = wasm_v128_load32_zero((const int32_t*) w + 7);
63 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
64 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
65 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point0);
66 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point0);
67 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point0);
68 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point0);
69 v128_t vacc0x4 = wasm_i32x4_mul(vksum4, vinput_zero_point0);
70 v128_t vacc0x5 = wasm_i32x4_mul(vksum5, vinput_zero_point0);
71 v128_t vacc0x6 = wasm_i32x4_mul(vksum6, vinput_zero_point0);
72 v128_t vacc0x7 = wasm_i32x4_mul(vksum7, vinput_zero_point0);
73 v128_t vacc1x0 = wasm_i32x4_mul(vksum0, vinput_zero_point1);
74 v128_t vacc1x1 = wasm_i32x4_mul(vksum1, vinput_zero_point1);
75 v128_t vacc1x2 = wasm_i32x4_mul(vksum2, vinput_zero_point1);
76 v128_t vacc1x3 = wasm_i32x4_mul(vksum3, vinput_zero_point1);
77 v128_t vacc1x4 = wasm_i32x4_mul(vksum4, vinput_zero_point1);
78 v128_t vacc1x5 = wasm_i32x4_mul(vksum5, vinput_zero_point1);
79 v128_t vacc1x6 = wasm_i32x4_mul(vksum6, vinput_zero_point1);
80 v128_t vacc1x7 = wasm_i32x4_mul(vksum7, vinput_zero_point1);
81 w = (const int32_t*) w + 8;
82
83 size_t k = kc;
84 do {
85 const v128_t va0 = wasm_v128_load(a0);
86 a0 += 16;
87 const v128_t va1 = wasm_v128_load(a1);
88 a1 += 16;
89
90 const v128_t vb0 = wasm_v128_load(w);
91
92 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
93 vacc1x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va1, vacc1x0);
94 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
95
96 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
97 vacc1x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va1, vacc1x1);
98 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
99
100 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
101 vacc1x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va1, vacc1x2);
102 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
103
104 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
105 vacc1x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va1, vacc1x3);
106 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
107
108 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
109 vacc1x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va1, vacc1x4);
110 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
111
112 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
113 vacc1x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va1, vacc1x5);
114 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
115
116 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
117 vacc1x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va1, vacc1x6);
118 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
119
120 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
121 vacc1x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va1, vacc1x7);
122
123 w = (const int8_t*) w + 128;
124 k -= 16 * sizeof(int8_t);
125 } while (k != 0);
126
127 const v128_t vacc0x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x0, vacc0x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x0, vacc0x2, 2, 6, 3, 7));
128 const v128_t vacc0x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x1, vacc0x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x1, vacc0x3, 2, 6, 3, 7));
129 const v128_t vacc0x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x4, vacc0x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x4, vacc0x6, 2, 6, 3, 7));
130 const v128_t vacc0x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x5, vacc0x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x5, vacc0x7, 2, 6, 3, 7));
131 const v128_t vacc1x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x0, vacc1x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x0, vacc1x2, 2, 6, 3, 7));
132 const v128_t vacc1x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x1, vacc1x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x1, vacc1x3, 2, 6, 3, 7));
133 const v128_t vacc1x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x4, vacc1x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x4, vacc1x6, 2, 6, 3, 7));
134 const v128_t vacc1x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x5, vacc1x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x5, vacc1x7, 2, 6, 3, 7));
135
136 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
137 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
138 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
139 v128_t vacc1x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x46, vacc1x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x46, vacc1x57, 2, 6, 3, 7));
140
141 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
142 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
143 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
144 vacc1x4567 = wasm_f32x4_convert_i32x4(vacc1x4567);
145
146 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
147 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
148
149 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
150 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
151 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
152 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vinput_scale1);
153
154 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
155 w = (const float*) w + 4;
156 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
157 w = (const float*) w + 4;
158 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
159 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
160 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
161 vacc1x4567 = wasm_f32x4_mul(vacc1x4567, vfilter_output_scale4567);
162
163 const v128_t vbias0123 = wasm_v128_load(w);
164 w = (const float*) w + 4;
165 const v128_t vbias4567 = wasm_v128_load(w);
166 w = (const float*) w + 4;
167 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
168 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
169 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
170 vacc1x4567 = wasm_f32x4_add(vacc1x4567, vbias4567);
171
172 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
173 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
174 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
175 vacc1x4567 = wasm_f32x4_pmax(vacc1x4567, vmin);
176
177 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
178 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
179 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
180 vacc1x4567 = wasm_f32x4_pmin(vacc1x4567, vmax);
181
182 if XNN_LIKELY(nc >= 8) {
183 wasm_v128_store(c0, vacc0x0123);
184 wasm_v128_store(c0 + 4, vacc0x4567);
185 wasm_v128_store(c1, vacc1x0123);
186 wasm_v128_store(c1 + 4, vacc1x4567);
187
188 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
189 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
190
191 c0 = (float*) ((uintptr_t) c0 + cn_stride);
192 c1 = (float*) ((uintptr_t) c1 + cn_stride);
193
194 nc -= 8;
195 } else {
196 if (nc & 4) {
197 wasm_v128_store(c0, vacc0x0123);
198 vacc0x0123 = vacc0x4567;
199 c0 += 4;
200 wasm_v128_store(c1, vacc1x0123);
201 vacc1x0123 = vacc1x4567;
202 c1 += 4;
203 }
204 if (nc & 2) {
205 wasm_v128_store64_lane(c0, vacc0x0123, 0);
206 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
207 c0 += 2;
208 wasm_v128_store64_lane(c1, vacc1x0123, 0);
209 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
210 c1 += 2;
211 }
212 if (nc & 1) {
213 wasm_v128_store32_lane(c0, vacc0x0123, 0);
214 wasm_v128_store32_lane(c1, vacc1x0123, 0);
215 }
216 nc = 0;
217 }
218 } while (nc != 0);
219 }