Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x4c8-minmax-wasmsimd-dot16x2-ld128.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x4c8__wasmsimd_dot16x2_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const int8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 float* c0 = c;
43
44 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
45 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
46 XNN_FORCE_REALIZATION(vmin);
47 XNN_FORCE_REALIZATION(vmax);
48
49 do {
50 const v128_t vksum0123 = wasm_v128_load(w);
51 const v128_t vinput_zero_point0 = wasm_v128_load32_splat(&quantization_params[0].zero_point);
52 const v128_t vinit0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
53
54 const v128_t vzero = wasm_i32x4_const_splat(0);
55 v128_t vacc0x0 = wasm_v32x4_shuffle(vinit0x0123, vzero, 0, 5, 6, 7);
56 v128_t vacc0x1 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 1, 6, 7);
57 v128_t vacc0x2 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 2, 7);
58 v128_t vacc0x3 = wasm_v32x4_shuffle(vinit0x0123, vzero, 4, 5, 6, 3);
59 w = (const int32_t*) w + 4;
60
61 size_t k = kc;
62
63 while (k >= 8 * sizeof(int8_t)) {
64 const v128_t vxa0 = wasm_i16x8_load8x8(a0);
65 a0 += 8;
66
67 const v128_t vb01 = wasm_v128_load(w);
68 const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
69 const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
70
71 vacc0x0 = wasm_i32x4_add(vacc0x0, wasm_i32x4_dot_i16x8(vxa0, vxb0));
72 vacc0x1 = wasm_i32x4_add(vacc0x1, wasm_i32x4_dot_i16x8(vxa0, vxb1));
73 const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
74 const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
75 const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
76
77 vacc0x2 = wasm_i32x4_add(vacc0x2, wasm_i32x4_dot_i16x8(vxa0, vxb2));
78 vacc0x3 = wasm_i32x4_add(vacc0x3, wasm_i32x4_dot_i16x8(vxa0, vxb3));
79
80 w = (const int8_t*) w + 32;
81 k -= 8 * sizeof(int8_t);
82 };
83
84 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));
85 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));
86
87 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
88
89 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
90
91 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
92
93 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
94
95 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
96 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
97 w = (const float*) w + 4;
98
99 const v128_t vbias0123 = wasm_v128_load(w);
100 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
101 w = (const float*) w + 4;
102
103 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
104
105 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
106
107 if XNN_LIKELY(nc >= 4) {
108 wasm_v128_store(c0, vacc0x0123);
109
110 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
111
112 c0 = (float*) ((uintptr_t) c0 + cn_stride);
113
114 nc -= 4;
115 } else {
116 if (nc & 2) {
117 wasm_v128_store64_lane(c0, vacc0x0123, 0);
118 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
119 c0 += 2;
120 }
121 if (nc & 1) {
122 wasm_v128_store32_lane(c0, vacc0x0123, 0);
123 }
124 nc = 0;
125 }
126 } while (nc != 0);
127 }