Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x4c8-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_3x4c8__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 <= 3);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (3 * 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
57 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
58 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
59 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
60 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
61 XNN_FORCE_REALIZATION(vmin);
62 XNN_FORCE_REALIZATION(vmax);
63 do {
64 const v128_t vksum0123 = wasm_v128_load(w);
65 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
66 const v128_t vinit1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
67 const v128_t vinit2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
68
69 const v128_t vzero = wasm_i32x4_const_splat(0);
70 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
71 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
72 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
73 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
74 v128_t vacc1x0 = wasm_v32x4_shuffle(vinit1x0123, vzero, 0, 5, 6, 7);
75 v128_t vacc1x1 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 1, 6, 7);
76 v128_t vacc1x2 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 2, 7);
77 v128_t vacc1x3 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 6, 3);
78 v128_t vacc2x0 = wasm_v32x4_shuffle(vinit2x0123, vzero, 0, 5, 6, 7);
79 v128_t vacc2x1 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 1, 6, 7);
80 v128_t vacc2x2 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 2, 7);
81 v128_t vacc2x3 = wasm_v32x4_shuffle(vinit2x0123, vzero, 4, 5, 6, 3);
82 w = (const void*) ((const int32_t*) w + 4);
83
84 size_t p = ks;
85 do {
86 const int8_t* restrict a0 = a[0];
87 if XNN_UNPREDICTABLE(a0 != zero) {
88 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
89 } else {
90 a0 = zero_data;
91 }
92 const int8_t* restrict a1 = a[1];
93 if XNN_UNPREDICTABLE(a1 != zero) {
94 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
95 } else {
96 a1 = zero_data;
97 }
98 const int8_t* restrict a2 = a[2];
99 if XNN_UNPREDICTABLE(a2 != zero) {
100 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
101 } else {
102 a2 = zero_data;
103 }
104 a += 3;
105
106 size_t k = kc;
107 do {
108 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
109 a0 += 8;
110 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
111 a1 += 8;
112 const v128_t vxa2 = wasm_i16x8_load8x8(a2);
113 a2 += 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 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
120 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
121 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
122 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
123 vacc2x0 = wasm_i32x4_add(vacc2x0, wasm_i32x4_dot_i16x8(vxa2, vxb0));
124 vacc2x1 = wasm_i32x4_add(vacc2x1, wasm_i32x4_dot_i16x8(vxa2, vxb1));
125 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
126 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
127 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
128
129 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
130 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
131 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
132 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
133 vacc2x2 = wasm_i32x4_add(vacc2x2, wasm_i32x4_dot_i16x8(vxa2, vxb2));
134 vacc2x3 = wasm_i32x4_add(vacc2x3, wasm_i32x4_dot_i16x8(vxa2, vxb3));
135
136 w = (const void*) ((const int8_t*) w + 32);
137 k -= 8 * sizeof(int8_t);
138 } while (k != 0);
139 p -= 3 * sizeof(void*);
140 } while (p != 0);
141
142 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));
143 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));
144 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));
145 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));
146 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));
147 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));
148
149 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
150 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
151 v128_t vacc2x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc2x02, vacc2x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc2x02, vacc2x13, 2, 6, 3, 7));
152
153 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
154 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
155 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
156
157 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
158 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
159 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
160
161 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
162 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
163 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
164 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
165 w = (const float*) w + 4;
166
167 const v128_t vbias0123 = wasm_v128_load(w);
168 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
169 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
170 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
171 w = (const float*) w + 4;
172
173 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
174 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
175 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
176
177 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
178 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
179 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
180
181 if XNN_LIKELY(nc >= 4) {
182 wasm_v128_store(c2, vacc2x0123);
183 wasm_v128_store(c1, vacc1x0123);
184 wasm_v128_store(c0, vacc0x0123);
185
186 a = (const int8_t**restrict) ((uintptr_t) a - ks);
187
188 c0 = (float*) ((uintptr_t) c0 + cn_stride);
189 c1 = (float*) ((uintptr_t) c1 + cn_stride);
190 c2 = (float*) ((uintptr_t) c2 + cn_stride);
191
192 nc -= 4;
193 } else {
194 if (nc & 2) {
195 wasm_v128_store64_lane(c2, vacc2x0123, 0);
196 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
197 c2 += 2;
198 wasm_v128_store64_lane(c1, vacc1x0123, 0);
199 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
200 c1 += 2;
201 wasm_v128_store64_lane(c0, vacc0x0123, 0);
202 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
203 c0 += 2;
204 }
205 if (nc & 1) {
206 wasm_v128_store32_lane(c2, vacc2x0123, 0);
207 wasm_v128_store32_lane(c1, vacc1x0123, 0);
208 wasm_v128_store32_lane(c0, vacc0x0123, 0);
209 }
210 nc = 0;
211 }
212 } while (nc != 0);
213 }