Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x4c8-minmax-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/common.h"
15 #include "xnnpack/gemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x4c8__wasmsimd_dot16x2_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 float* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const int8_t* zero_data,
32 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
33 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
34 {
35 assert(mr != 0);
36 assert(mr <= 4);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (4 * sizeof(void*)) == 0);
41 assert(a_offset % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 8 * sizeof(int8_t));
47 float* c0 = c;
48 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 c1 = c0;
51 }
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 c2 = c1;
55 }
56 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 c3 = c2;
59 }
60
61 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
62 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
63 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
64 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
65 XNN_FORCE_REALIZATION(vmin);
66 XNN_FORCE_REALIZATION(vmax);
67 do {
68 const v128_t vksum0123 = wasm_v128_load(w);
69 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
70 const v128_t vinit1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
71 const v128_t vinit2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
72 const v128_t vinit3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
73
74 const v128_t vzero = wasm_i32x4_const_splat(0);
75 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
76 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
77 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
78 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
79 v128_t vacc1x0 = wasm_v32x4_shuffle(vinit1x0123, vzero, 0, 5, 6, 7);
80 v128_t vacc1x1 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 1, 6, 7);
81 v128_t vacc1x2 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 2, 7);
82 v128_t vacc1x3 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 6, 3);
83 v128_t vacc2x0 = wasm_v32x4_shuffle(vinit2x0123, vzero, 0, 5, 6, 7);
84 v128_t vacc2x1 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 1, 6, 7);
85 v128_t vacc2x2 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 2, 7);
86 v128_t vacc2x3 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 6, 3);
87 v128_t vacc3x0 = wasm_v32x4_shuffle(vinit3x0123, vzero, 0, 5, 6, 7);
88 v128_t vacc3x1 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 1, 6, 7);
89 v128_t vacc3x2 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 5, 2, 7);
90 v128_t vacc3x3 = wasm_v32x4_shuffle(vinit3x0123, vzero, 4, 5, 6, 3);
91 w = (const void*) ((const int32_t*) w + 4);
92
93 size_t p = ks;
94 do {
95 const int8_t* restrict a0 = a[0];
96 if XNN_UNPREDICTABLE(a0 != zero) {
97 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
98 } else {
99 a0 = zero_data;
100 }
101 const int8_t* restrict a1 = a[1];
102 if XNN_UNPREDICTABLE(a1 != zero) {
103 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
104 } else {
105 a1 = zero_data;
106 }
107 const int8_t* restrict a2 = a[2];
108 if XNN_UNPREDICTABLE(a2 != zero) {
109 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
110 } else {
111 a2 = zero_data;
112 }
113 const int8_t* restrict a3 = a[3];
114 if XNN_UNPREDICTABLE(a3 != zero) {
115 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
116 } else {
117 a3 = zero_data;
118 }
119 a += 4;
120
121 size_t k = kc;
122 do {
123 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
124 a0 += 8;
125 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
126 a1 += 8;
127 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
128 a2 += 8;
129 const v128_t vxa3 = wasm_i16x8_load8x8(a3);
130 a3 += 8;
131
132 const v128_t vb01 = wasm_v128_load(w);
133 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
134 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
135
136 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
137 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
138 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
139 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
140 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
141 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
142 vacc3x0 = wasm_i32x4_add(vacc3x0, wasm_i32x4_dot_i16x8(vxa3, vxb0));
143 vacc3x1 = wasm_i32x4_add(vacc3x1, wasm_i32x4_dot_i16x8(vxa3, vxb1));
144 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
145 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
146 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
147
148 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
149 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
150 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
151 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
152 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
153 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
154 vacc3x2 = wasm_i32x4_add(vacc3x2, wasm_i32x4_dot_i16x8(vxa3, vxb2));
155 vacc3x3 = wasm_i32x4_add(vacc3x3, wasm_i32x4_dot_i16x8(vxa3, vxb3));
156
157 w = (const void*) ((const int8_t*) w + 32);
158 k -= 8 * sizeof(int8_t);
159 } while (k != 0);
160 p -= 4 * sizeof(void*);
161 } while (p != 0);
162
163 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));
164 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));
165 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));
166 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));
167 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));
168 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));
169 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));
170 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));
171
172 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
173 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
174 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
175 v128_t vacc3x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc3x02, vacc3x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc3x02, vacc3x13, 2, 6, 3, 7));
176
177 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
178 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
179 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
180 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
181
182 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
183 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
184 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
185 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale);
186
187 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
188 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
189 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
190 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
191 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
192 w = (const float*) w + 4;
193
194 const v128_t vbias0123 = wasm_v128_load(w);
195 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
196 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
197 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
198 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
199 w = (const float*) w + 4;
200
201 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
202 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
203 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
204 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
205
206 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
207 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
208 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
209 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
210
211 if XNN_LIKELY(nc >= 4) {
212 wasm_v128_store(c3, vacc3x0123);
213 wasm_v128_store(c2, vacc2x0123);
214 wasm_v128_store(c1, vacc1x0123);
215 wasm_v128_store(c0, vacc0x0123);
216
217 a = (const int8_t**restrict) ((uintptr_t) a - ks);
218
219 c0 = (float*) ((uintptr_t) c0 + cn_stride);
220 c1 = (float*) ((uintptr_t) c1 + cn_stride);
221 c2 = (float*) ((uintptr_t) c2 + cn_stride);
222 c3 = (float*) ((uintptr_t) c3 + cn_stride);
223
224 nc -= 4;
225 } else {
226 if (nc & 2) {
227 wasm_v128_store64_lane(c3, vacc3x0123, 0);
228 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
229 c3 += 2;
230 wasm_v128_store64_lane(c2, vacc2x0123, 0);
231 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
232 c2 += 2;
233 wasm_v128_store64_lane(c1, vacc1x0123, 0);
234 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
235 c1 += 2;
236 wasm_v128_store64_lane(c0, vacc0x0123, 0);
237 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
238 c0 += 2;
239 }
240 if (nc & 1) {
241 wasm_v128_store32_lane(c3, vacc3x0123, 0);
242 wasm_v128_store32_lane(c2, vacc2x0123, 0);
243 wasm_v128_store32_lane(c1, vacc1x0123, 0);
244 wasm_v128_store32_lane(c0, vacc0x0123, 0);
245 }
246 nc = 0;
247 }
248 } while (nc != 0);
249 }