Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x4c8-minmax-wasmsimd-dot16x2-ld64.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_1x4c8__wasmsimd_dot16x2_ld64(
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 <= 1);
37 assert(nc != 0);
38 assert(kc != 0);
39 assert(ks != 0);
40 assert(ks % (1 * 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
49 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
50 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
51 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
52 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
53 XNN_FORCE_REALIZATION(vmin);
54 XNN_FORCE_REALIZATION(vmax);
55 do {
56 const v128_t vksum0123 = wasm_v128_load(w);
57 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
58
59 const v128_t vzero = wasm_i32x4_const_splat(0);
60 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
61 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
62 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
63 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
64 w = (const void*) ((const int32_t*) w + 4);
65
66 size_t p = ks;
67 do {
68 const int8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
71 } else {
72 a0 = zero_data;
73 }
74 a += 1;
75
76 size_t k = kc;
77 do {
78 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
79 a0 += 8;
80
81 const v128_t vxb0 = wasm_i16x8_load8x8(w);
82
83 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
84 const v128_t vxb1 = wasm_i16x8_load8x8((const int8_t*) w + 8);
85
86 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
87 const v128_t vxb2 = wasm_i16x8_load8x8((const int8_t*) w + 16);
88
89 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
90 const v128_t vxb3 = wasm_i16x8_load8x8((const int8_t*) w + 24);
91
92 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
93
94 w = (const void*) ((const int8_t*) w + 32);
95 k -= 8 * sizeof(int8_t);
96 } while (k != 0);
97 p -= 1 * sizeof(void*);
98 } while (p != 0);
99
100 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));
101 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));
102
103 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
104
105 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
106
107 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
108
109 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
110 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
111 w = (const float*) w + 4;
112
113 const v128_t vbias0123 = wasm_v128_load(w);
114 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
115 w = (const float*) w + 4;
116
117 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
118
119 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
120
121 if XNN_LIKELY(nc >= 4) {
122 wasm_v128_store(c0, vacc0x0123);
123
124 a = (const int8_t**restrict) ((uintptr_t) a - ks);
125
126 c0 = (float*) ((uintptr_t) c0 + cn_stride);
127
128 nc -= 4;
129 } else {
130 if (nc & 2) {
131 wasm_v128_store64_lane(c0, vacc0x0123, 0);
132 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
133 c0 += 2;
134 }
135 if (nc & 1) {
136 wasm_v128_store32_lane(c0, vacc0x0123, 0);
137 }
138 nc = 0;
139 }
140 } while (nc != 0);
141 }