Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-4x4c8-minmax-wasmsimd-dot16x2-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-wasmsimd-dot16x2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 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_4x4c8__wasmsimd_dot16x2_ld64(
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 vksum0123 = wasm_v128_load(w);
69 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
70 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
71 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
72 const v128_t vinput_zero_point3 = wasm_v128_load32_splat(&quantization_params[3].zero_point);
73 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
74 const v128_t vinit1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
75 const v128_t vinit2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
76 const v128_t vinit3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
77
78 const v128_t vzero = wasm_i32x4_const_splat(0);
79 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
80 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
81 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
82 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
83 v128_t vacc1x0 = wasm_v32x4_shuffle(vinit1x0123, vzero, 0, 5, 6, 7);
84 v128_t vacc1x1 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 1, 6, 7);
85 v128_t vacc1x2 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 2, 7);
86 v128_t vacc1x3 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 6, 3);
87 v128_t vacc2x0 = wasm_v32x4_shuffle(vinit2x0123, vzero, 0, 5, 6, 7);
88 v128_t vacc2x1 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 1, 6, 7);
89 v128_t vacc2x2 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 2, 7);
90 v128_t vacc2x3 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 6, 3);
91 v128_t vacc3x0 = wasm_v32x4_shuffle(vinit3x0123, vzero, 0, 5, 6, 7);
92 v128_t vacc3x1 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 1, 6, 7);
93 v128_t vacc3x2 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 5, 2, 7);
94 v128_t vacc3x3 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 5, 6, 3);
95 w = (const int32_t*) w + 4;
96
97 size_t k = kc;
98
99 while (k >= 8 * sizeof(int8_t)) {
100 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
101 a0 += 8;
102 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
103 a1 += 8;
104 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
105 a2 += 8;
106 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
107 a3 += 8;
108
109 const v128_t vxb0 = wasm_i16x8_load8x8(w);
110
111 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
112 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
113 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
114 vacc3x0 = wasm_i32x4_add(vacc3x0, wasm_i32x4_dot_i16x8(vxa3, vxb0));
115 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
116
117 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
118 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
119 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
120 vacc3x1 = wasm_i32x4_add(vacc3x1, wasm_i32x4_dot_i16x8(vxa3, vxb1));
121 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
122
123 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
124 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
125 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
126 vacc3x2 = wasm_i32x4_add(vacc3x2, wasm_i32x4_dot_i16x8(vxa3, vxb2));
127 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
128
129 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
130 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
131 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
132 vacc3x3 = wasm_i32x4_add(vacc3x3, wasm_i32x4_dot_i16x8(vxa3, vxb3));
133
134 w = (const int8_t*) w + 32;
135 k -= 8 * sizeof(int8_t);
136 };
137
138 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));
139 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));
140 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));
141 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));
142 const v128_t vacc2x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x0, vacc2x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x0, vacc2x2, 2, 6, 3, 7));
143 const v128_t vacc2x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x1, vacc2x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x1, vacc2x3, 2, 6, 3, 7));
144 const v128_t vacc3x02 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x0, vacc3x2, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x0, vacc3x2, 2, 6, 3, 7));
145 const v128_t vacc3x13 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x1, vacc3x3, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x1, vacc3x3, 2, 6, 3, 7));
146
147 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
148 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
149 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
150 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
151
152 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
153 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
154 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
155 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
156
157 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
158 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
159 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
160 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
161
162 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
163 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
164 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
165 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
166
167 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
168 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
169 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
170 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
171 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
172 w = (const float*) w + 4;
173
174 const v128_t vbias0123 = wasm_v128_load(w);
175 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
176 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
177 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
178 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
179 w = (const float*) w + 4;
180
181 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
182 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
183 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
184 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
185
186 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
187 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
188 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
189 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
190
191 if XNN_LIKELY(nc >= 4) {
192 wasm_v128_store(c0, vacc0x0123);
193 wasm_v128_store(c1, vacc1x0123);
194 wasm_v128_store(c2, vacc2x0123);
195 wasm_v128_store(c3, vacc3x0123);
196
197 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
198 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
199 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
200 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
201
202 c0 = (float*) ((uintptr_t) c0 + cn_stride);
203 c1 = (float*) ((uintptr_t) c1 + cn_stride);
204 c2 = (float*) ((uintptr_t) c2 + cn_stride);
205 c3 = (float*) ((uintptr_t) c3 + cn_stride);
206
207 nc -= 4;
208 } else {
209 if (nc & 2) {
210 wasm_v128_store64_lane(c0, vacc0x0123, 0);
211 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
212 c0 += 2;
213 wasm_v128_store64_lane(c1, vacc1x0123, 0);
214 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
215 c1 += 2;
216 wasm_v128_store64_lane(c2, vacc2x0123, 0);
217 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
218 c2 += 2;
219 wasm_v128_store64_lane(c3, vacc3x0123, 0);
220 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
221 c3 += 2;
222 }
223 if (nc & 1) {
224 wasm_v128_store32_lane(c0, vacc0x0123, 0);
225 wasm_v128_store32_lane(c1, vacc1x0123, 0);
226 wasm_v128_store32_lane(c2, vacc2x0123, 0);
227 wasm_v128_store32_lane(c3, vacc3x0123, 0);
228 }
229 nc = 0;
230 }
231 } while (nc != 0);
232 }