Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-3x4c2s4-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_3x4c2s4__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 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
56 kc = round_up_po2(kc, 8 * sizeof(int8_t));
57 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
58 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
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
64 do {
65 const v128_t vksum0123 = wasm_v128_load(w);
66 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
67 v128_t vacc1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
68 v128_t vacc2x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
69 w = (const void*) ((const int32_t*) w + 4);
70
71 size_t p = ks;
72 do {
73 const int8_t* restrict a0 = a[0];
74 if XNN_UNPREDICTABLE(a0 != zero) {
75 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
76 } else {
77 a0 = zero_data;
78 }
79 const int8_t* restrict a1 = a[1];
80 if XNN_UNPREDICTABLE(a1 != zero) {
81 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
82 } else {
83 a1 = zero_data;
84 }
85 const int8_t* restrict a2 = a[2];
86 if XNN_UNPREDICTABLE(a2 != zero) {
87 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
88 } else {
89 a2 = zero_data;
90 }
91 a += 3;
92
93 size_t k = kc;
94 do {
95 v128_t vxa0 = wasm_i16x8_load8x8(a0);
96 a0 += 8;
97 v128_t vxa1 = wasm_i16x8_load8x8(a1);
98 a1 += 8;
99 v128_t vxa2 = wasm_i16x8_load8x8(a2);
100 a2 += 8;
101
102 const v128_t vb01 = wasm_v128_load(w);
103 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
104 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
105
106 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
107 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
108 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
109 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
110 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb0));
111 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
112
113 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
114 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
115 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
116 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
117 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb1));
118 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
119 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
120 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
121 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
122
123 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
124 vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
125 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
126 vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
127 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb2));
128 vxa2 = wasm_v32x4_shuffle(vxa2, vxa2, 1, 2, 3, 4);
129
130 vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
131 vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
132 vacc2x0123 = wasm_i32x4_add(vacc2x0123, wasm_i32x4_dot_i16x8(vxa2, vxb3));
133
134 w = (const int8_t*) w + 32;
135 k -= 8 * sizeof(int8_t);
136 } while (k != 0);
137 p -= 3 * sizeof(void*);
138 } while (p != 0);
139
140 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
141 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
142 vacc2x0123 = wasm_f32x4_convert_i32x4(vacc2x0123);
143
144 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
145 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
146 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vinput_scale);
147
148 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
149 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
150 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
151 vacc2x0123 = wasm_f32x4_mul(vacc2x0123, vfilter_output_scale0123);
152 w = (const float*) w + 4;
153
154 const v128_t vbias0123 = wasm_v128_load(w);
155 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
156 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
157 vacc2x0123 = wasm_f32x4_add(vacc2x0123, vbias0123);
158 w = (const float*) w + 4;
159
160 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
161 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
162 vacc2x0123 = wasm_f32x4_pmax(vacc2x0123, vmin);
163
164 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
165 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
166 vacc2x0123 = wasm_f32x4_pmin(vacc2x0123, vmax);
167
168 if XNN_LIKELY(nc >= 4) {
169 wasm_v128_store(c2, vacc2x0123);
170 wasm_v128_store(c1, vacc1x0123);
171 wasm_v128_store(c0, vacc0x0123);
172
173 a = (const int8_t**restrict) ((uintptr_t) a - ks);
174
175 c0 = (float*) ((uintptr_t) c0 + cn_stride);
176 c1 = (float*) ((uintptr_t) c1 + cn_stride);
177 c2 = (float*) ((uintptr_t) c2 + cn_stride);
178
179 nc -= 4;
180 } else {
181 if (nc & 2) {
182 wasm_v128_store64_lane(c2, vacc2x0123, 0);
183 vacc2x0123 = wasm_v64x2_shuffle(vacc2x0123, vacc2x0123, 1, 1);
184 c2 += 2;
185 wasm_v128_store64_lane(c1, vacc1x0123, 0);
186 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
187 c1 += 2;
188 wasm_v128_store64_lane(c0, vacc0x0123, 0);
189 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
190 c0 += 2;
191 }
192 if (nc & 1) {
193 wasm_v128_store32_lane(c2, vacc2x0123, 0);
194 wasm_v128_store32_lane(c1, vacc1x0123, 0);
195 wasm_v128_store32_lane(c0, vacc0x0123, 0);
196 }
197 nc = 0;
198 }
199 } while (nc != 0);
200 }