Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-4x4c2s4-minmax-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2s4-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/igemm.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_4x4c2s4__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 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 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 kc = round_up_po2(kc, 8 * sizeof(int8_t));
61 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
62 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
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
68 do {
69 const v128_t vksum0123 = wasm_v128_load(w);
70 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
71 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
72 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
73 v128_t vacc3x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
74 w = (const void*) ((const int32_t*) w + 4);
75
76 size_t p = ks;
77 do {
78 const int8_t* restrict a0 = a[0];
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
81 } else {
82 a0 = zero_data;
83 }
84 const int8_t* restrict a1 = a[1];
85 if XNN_UNPREDICTABLE(a1 != zero) {
86 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
87 } else {
88 a1 = zero_data;
89 }
90 const int8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93 } else {
94 a2 = zero_data;
95 }
96 const int8_t* restrict a3 = a[3];
97 if XNN_UNPREDICTABLE(a3 != zero) {
98 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
99 } else {
100 a3 = zero_data;
101 }
102 a += 4;
103
104 size_t k = kc;
105 do {
106 v128_t vxa0 = wasm_i16x8_load8x8(a0);
107 a0 += 8;
108 v128_t vxa1 = wasm_i16x8_load8x8(a1);
109 a1 += 8;
110 v128_t vxa2 = wasm_i16x8_load8x8(a2);
111 a2 += 8;
112 v128_t vxa3 = wasm_i16x8_load8x8(a3);
113 a3 += 8;
114
115 const v128_t vb01 = wasm_v128_load(w);
116 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
117 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
118
119 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
120 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
121 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
122 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
123 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
124 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
125 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb0));
126 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
127
128 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
129 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
130 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
131 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
132 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
133 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
134 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb1));
135 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
136 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
137 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
138 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
139
140 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
141 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
142 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
143 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
144 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
145 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
146 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb2));
147 vxa3 = wasm_v32x4_shuffle(vxa3, vxa3, 1, 2, 3, 4);
148
149 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
150 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
151 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
152 vacc3x0123 = wasm_i32x4_add(vacc3x0123, wasm_i32x4_dot_i16x8(vxa3, vxb3));
153
154 w = (const int8_t*) w + 32;
155 k -= 8 * sizeof(int8_t);
156 } while (k != 0);
157 p -= 4 * sizeof(void*);
158 } while (p != 0);
159
160 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
161 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
162 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
163 vacc3x0123 = wasm_f32x4_convert_i32x4(vacc3x0123);
164
165 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
166 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
167 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
168 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vinput_scale);
169
170 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
171 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
172 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
173 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
174 vacc3x0123 = wasm_f32x4_mul(vacc3x0123, vfilter_output_scale0123);
175 w = (const float*) w + 4;
176
177 const v128_t vbias0123 = wasm_v128_load(w);
178 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
179 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
180 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
181 vacc3x0123 = wasm_f32x4_add(vacc3x0123, vbias0123);
182 w = (const float*) w + 4;
183
184 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
185 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
186 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
187 vacc3x0123 = wasm_f32x4_pmax(vacc3x0123, vmin);
188
189 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
190 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
191 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
192 vacc3x0123 = wasm_f32x4_pmin(vacc3x0123, vmax);
193
194 if XNN_LIKELY(nc >= 4) {
195 wasm_v128_store(c3, vacc3x0123);
196 wasm_v128_store(c2, vacc2x0123);
197 wasm_v128_store(c1, vacc1x0123);
198 wasm_v128_store(c0, vacc0x0123);
199
200 a = (const int8_t**restrict) ((uintptr_t) a - ks);
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(c3, vacc3x0123, 0);
211 vacc3x0123 = wasm_v64x2_shuffle(vacc3x0123, vacc3x0123, 1, 1);
212 c3 += 2;
213 wasm_v128_store64_lane(c2, vacc2x0123, 0);
214 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
215 c2 += 2;
216 wasm_v128_store64_lane(c1, vacc1x0123, 0);
217 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
218 c1 += 2;
219 wasm_v128_store64_lane(c0, vacc0x0123, 0);
220 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
221 c0 += 2;
222 }
223 if (nc & 1) {
224 wasm_v128_store32_lane(c3, vacc3x0123, 0);
225 wasm_v128_store32_lane(c2, vacc2x0123, 0);
226 wasm_v128_store32_lane(c1, vacc1x0123, 0);
227 wasm_v128_store32_lane(c0, vacc0x0123, 0);
228 }
229 nc = 0;
230 }
231 } while (nc != 0);
232 }