Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-2x4c8-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_2x4c8__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 <= 2);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (2 * 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
53 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
54 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
55 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
56 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
57 XNN_FORCE_REALIZATION(vmin);
58 XNN_FORCE_REALIZATION(vmax);
59 do {
60 const v128_t vksum0123 = wasm_v128_load(w);
61 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
62 const v128_t vinit1x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
63
64 const v128_t vzero = wasm_i32x4_const_splat(0);
65 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
66 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
67 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
68 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
69 v128_t vacc1x0 = wasm_v32x4_shuffle(vinit1x0123, vzero, 0, 5, 6, 7);
70 v128_t vacc1x1 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 1, 6, 7);
71 v128_t vacc1x2 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 2, 7);
72 v128_t vacc1x3 = wasm_v32x4_shuffle(vinit1x0123, vzero, 4, 5, 6, 3);
73 w = (const void*) ((const int32_t*) w + 4);
74
75 size_t p = ks;
76 do {
77 const int8_t* restrict a0 = a[0];
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
80 } else {
81 a0 = zero_data;
82 }
83 const int8_t* restrict a1 = a[1];
84 if XNN_UNPREDICTABLE(a1 != zero) {
85 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
86 } else {
87 a1 = zero_data;
88 }
89 a += 2;
90
91 size_t k = kc;
92 do {
93 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
94 a0 += 8;
95 const v128_t vxa1 = wasm_i16x8_load8x8(a1);
96 a1 += 8;
97
98 const v128_t vb01 = wasm_v128_load(w);
99 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
100 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
101
102 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
103 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
104 vacc1x0 = wasm_i32x4_add(vacc1x0, wasm_i32x4_dot_i16x8(vxa1, vxb0));
105 vacc1x1 = wasm_i32x4_add(vacc1x1, wasm_i32x4_dot_i16x8(vxa1, vxb1));
106 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
107 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
108 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
109
110 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
111 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
112 vacc1x2 = wasm_i32x4_add(vacc1x2, wasm_i32x4_dot_i16x8(vxa1, vxb2));
113 vacc1x3 = wasm_i32x4_add(vacc1x3, wasm_i32x4_dot_i16x8(vxa1, vxb3));
114
115 w = (const void*) ((const int8_t*) w + 32);
116 k -= 8 * sizeof(int8_t);
117 } while (k != 0);
118 p -= 2 * sizeof(void*);
119 } while (p != 0);
120
121 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));
122 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));
123 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));
124 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));
125
126 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
127 v128_t vacc1x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc1x02, vacc1x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc1x02, vacc1x13, 2, 6, 3, 7));
128
129 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
130 vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
131
132 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
133 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vinput_scale);
134
135 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
136 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
137 vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vfilter_output_scale0123);
138 w = (const float*) w + 4;
139
140 const v128_t vbias0123 = wasm_v128_load(w);
141 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
142 vacc1x0123 = wasm_f32x4_add(vacc1x0123, vbias0123);
143 w = (const float*) w + 4;
144
145 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
146 vacc1x0123 = wasm_f32x4_pmax(vacc1x0123, vmin);
147
148 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
149 vacc1x0123 = wasm_f32x4_pmin(vacc1x0123, vmax);
150
151 if XNN_LIKELY(nc >= 4) {
152 wasm_v128_store(c1, vacc1x0123);
153 wasm_v128_store(c0, vacc0x0123);
154
155 a = (const int8_t**restrict) ((uintptr_t) a - ks);
156
157 c0 = (float*) ((uintptr_t) c0 + cn_stride);
158 c1 = (float*) ((uintptr_t) c1 + cn_stride);
159
160 nc -= 4;
161 } else {
162 if (nc & 2) {
163 wasm_v128_store64_lane(c1, vacc1x0123, 0);
164 vacc1x0123 = wasm_v64x2_shuffle(vacc1x0123, vacc1x0123, 1, 1);
165 c1 += 2;
166 wasm_v128_store64_lane(c0, vacc0x0123, 0);
167 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
168 c0 += 2;
169 }
170 if (nc & 1) {
171 wasm_v128_store32_lane(c1, vacc1x0123, 0);
172 wasm_v128_store32_lane(c0, vacc0x0123, 0);
173 }
174 nc = 0;
175 }
176 } while (nc != 0);
177 }