Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-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_qc4w_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 struct xnn_f32_qc4w_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 const v128_t vmask = wasm_i8x16_const_splat(0xF0);
68 do {
69 const v128_t vksum0123 = wasm_v128_load(w);
70 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
71 const v128_t vinput_zero_point1 = wasm_v128_load32_splat(&quantization_params[1].zero_point);
72 const v128_t vinput_zero_point2 = wasm_v128_load32_splat(&quantization_params[2].zero_point);
73 const v128_t vinput_zero_point3 = wasm_v128_load32_splat(&quantization_params[3].zero_point);
74 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
75 const v128_t vinit1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point1);
76 const v128_t vinit2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point2);
77 const v128_t vinit3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point3);
78
79 const v128_t vzero = wasm_i32x4_const_splat(0);
80 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
81 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
82 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
83 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
84 v128_t vacc1x0 = wasm_v32x4_shuffle(vinit1x0123, vzero, 0, 5, 6, 7);
85 v128_t vacc1x1 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 1, 6, 7);
86 v128_t vacc1x2 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 2, 7);
87 v128_t vacc1x3 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 6, 3);
88 v128_t vacc2x0 = wasm_v32x4_shuffle(vinit2x0123, vzero, 0, 5, 6, 7);
89 v128_t vacc2x1 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 1, 6, 7);
90 v128_t vacc2x2 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 2, 7);
91 v128_t vacc2x3 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 6, 3);
92 v128_t vacc3x0 = wasm_v32x4_shuffle(vinit3x0123, vzero, 0, 5, 6, 7);
93 v128_t vacc3x1 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 1, 6, 7);
94 v128_t vacc3x2 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 5, 2, 7);
95 v128_t vacc3x3 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 5, 6, 3);
96 w = (const int32_t*) w + 4;
97
98 size_t k = kc;
99 while (k >= 16 * sizeof(int8_t)) {
100 v128_t vxa0;
101 v128_t vxa1;
102 v128_t vxa2;
103 v128_t vxa3;
104 vxa0 = wasm_i16x8_load8x8(a0);
105 a0 += 8;
106 vxa1 = wasm_i16x8_load8x8(a1);
107 a1 += 8;
108 vxa2 = wasm_i16x8_load8x8(a2);
109 a2 += 8;
110 vxa3 = wasm_i16x8_load8x8(a3);
111 a3 += 8;
112 const v128_t vb0 = wasm_v128_load(w);
113 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 16);
114 const v128_t vbm0 = wasm_i8x16_shl(vb0, 4);
115 const v128_t vbm2 = wasm_i8x16_shl(vb2, 4);
116 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vbm0);
117 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vbm0);
118 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vbm2);
119 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vbm2);
120 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
121 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
122 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
123 vacc3x0 = wasm_i32x4_add(vacc3x0, wasm_i32x4_dot_i16x8(vxa3, vxb0));
124 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
125 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
126 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
127 vacc3x1 = wasm_i32x4_add(vacc3x1, wasm_i32x4_dot_i16x8(vxa3, vxb1));
128 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
129 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
130 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
131 vacc3x2 = wasm_i32x4_add(vacc3x2, wasm_i32x4_dot_i16x8(vxa3, vxb2));
132 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
133 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
134 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
135 vacc3x3 = wasm_i32x4_add(vacc3x3, wasm_i32x4_dot_i16x8(vxa3, vxb3));
136 vxa0 = wasm_i16x8_load8x8(a0);
137 a0 += 8;
138 vxa1 = wasm_i16x8_load8x8(a1);
139 a1 += 8;
140 vxa2 = wasm_i16x8_load8x8(a2);
141 a2 += 8;
142 vxa3 = wasm_i16x8_load8x8(a3);
143 a3 += 8;
144 const v128_t vbm4 = wasm_v128_and(vb0, vmask);
145 const v128_t vbm6 = wasm_v128_and(vb2, vmask);
146 const v128_t vxb4 = wasm_i16x8_extend_low_i8x16(vbm4);
147 const v128_t vxb5 = wasm_i16x8_extend_high_i8x16(vbm4);
148 const v128_t vxb6 = wasm_i16x8_extend_low_i8x16(vbm6);
149 const v128_t vxb7 = wasm_i16x8_extend_high_i8x16(vbm6);
150 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb4));
151 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb4));
152 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb4));
153 vacc3x0 = wasm_i32x4_add(vacc3x0, wasm_i32x4_dot_i16x8(vxa3, vxb4));
154 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb5));
155 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb5));
156 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb5));
157 vacc3x1 = wasm_i32x4_add(vacc3x1, wasm_i32x4_dot_i16x8(vxa3, vxb5));
158 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb6));
159 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb6));
160 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb6));
161 vacc3x2 = wasm_i32x4_add(vacc3x2, wasm_i32x4_dot_i16x8(vxa3, vxb6));
162 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb7));
163 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb7));
164 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb7));
165 vacc3x3 = wasm_i32x4_add(vacc3x3, wasm_i32x4_dot_i16x8(vxa3, vxb7));
166 w = (const int8_t*) w + 32;
167 k -= 16 * sizeof(int8_t);
168 }
169
170 while (k >= 8 * sizeof(int8_t)) {
171 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
172 a0 += 8;
173 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
174 a1 += 8;
175 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
176 a2 += 8;
177 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
178 a3 += 8;
179
180 const v128_t vb0 = wasm_v128_load(w);
181 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 16);
182 const v128_t vbm0 = wasm_i8x16_shl(vb0, 4);
183 const v128_t vbm2 = wasm_i8x16_shl(vb2, 4);
184 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vbm0);
185 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vbm0);
186 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vbm2);
187 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vbm2);
188
189 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
190 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
191 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
192 vacc3x0 = wasm_i32x4_add(vacc3x0, wasm_i32x4_dot_i16x8(vxa3, vxb0));
193
194 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
195 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
196 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
197 vacc3x1 = wasm_i32x4_add(vacc3x1, wasm_i32x4_dot_i16x8(vxa3, vxb1));
198
199 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
200 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
201 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
202 vacc3x2 = wasm_i32x4_add(vacc3x2, wasm_i32x4_dot_i16x8(vxa3, vxb2));
203
204 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
205 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
206 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
207 vacc3x3 = wasm_i32x4_add(vacc3x3, wasm_i32x4_dot_i16x8(vxa3, vxb3));
208
209 w = (const int8_t*) w + 32; // only low nibble used
210 k -= 8 * sizeof(int8_t);
211 };
212
213 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));
214 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));
215 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));
216 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));
217 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));
218 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));
219 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));
220 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));
221
222 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
223 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
224 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
225 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
226
227 vacc0x0123 = wasm_i32x4_shr(vacc0x0123, 4);
228 vacc1x0123 = wasm_i32x4_shr(vacc1x0123, 4);
229 vacc2x0123 = wasm_i32x4_shr(vacc2x0123, 4);
230 vacc3x0123 = wasm_i32x4_shr(vacc3x0123, 4);
231 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
232 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
233 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
234 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
235
236 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
237 const v128_t vinput_scale1 = wasm_v128_load32_splat(&quantization_params[1].inv_scale);
238 const v128_t vinput_scale2 = wasm_v128_load32_splat(&quantization_params[2].inv_scale);
239 const v128_t vinput_scale3 = wasm_v128_load32_splat(&quantization_params[3].inv_scale);
240
241 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
242 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale1);
243 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale2);
244 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale3);
245
246 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
247 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
248 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
249 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
250 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
251 w = (const float*) w + 4;
252
253 const v128_t vbias0123 = wasm_v128_load(w);
254 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
255 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
256 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
257 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
258 w = (const float*) w + 4;
259
260 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
261 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
262 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
263 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
264
265 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
266 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
267 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
268 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
269
270 if XNN_LIKELY(nc >= 4) {
271 wasm_v128_store(c0, vacc0x0123);
272 wasm_v128_store(c1, vacc1x0123);
273 wasm_v128_store(c2, vacc2x0123);
274 wasm_v128_store(c3, vacc3x0123);
275
276 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
277 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
278 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
279 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
280
281 c0 = (float*) ((uintptr_t) c0 + cn_stride);
282 c1 = (float*) ((uintptr_t) c1 + cn_stride);
283 c2 = (float*) ((uintptr_t) c2 + cn_stride);
284 c3 = (float*) ((uintptr_t) c3 + cn_stride);
285
286 nc -= 4;
287 } else {
288 if (nc & 2) {
289 wasm_v128_store64_lane(c0, vacc0x0123, 0);
290 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
291 c0 += 2;
292 wasm_v128_store64_lane(c1, vacc1x0123, 0);
293 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
294 c1 += 2;
295 wasm_v128_store64_lane(c2, vacc2x0123, 0);
296 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
297 c2 += 2;
298 wasm_v128_store64_lane(c3, vacc3x0123, 0);
299 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
300 c3 += 2;
301 }
302 if (nc & 1) {
303 wasm_v128_store32_lane(c0, vacc0x0123, 0);
304 wasm_v128_store32_lane(c1, vacc1x0123, 0);
305 wasm_v128_store32_lane(c2, vacc2x0123, 0);
306 wasm_v128_store32_lane(c3, vacc3x0123, 0);
307 }
308 nc = 0;
309 }
310 } while (nc != 0);
311 }