Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x4c16-minmax-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c16-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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 void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_1x4c16__wasmusdot(
18 size_t mr,
19 size_t nc,
20 size_t kc,
21 const int8_t* restrict a,
22 size_t a_stride,
23 const void* restrict w,
24 float* restrict c,
25 size_t cm_stride,
26 size_t cn_stride,
27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
28 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 1);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(int8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 16 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 float* c0 = c;
42
43 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
44 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
45 XNN_FORCE_REALIZATION(vmin);
46 XNN_FORCE_REALIZATION(vmax);
47
48 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
49 do {
50 v128_t vksum0 = wasm_v128_load32_zero(w);
51 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
52 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
53 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
54 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
55 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point0);
56 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point0);
57 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point0);
58 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point0);
59 w = (const int32_t*) w + 4;
60
61 size_t k = kc;
62 do {
63 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
64 a0 += 16;
65
66 const v128_t vb0 = wasm_v128_load(w);
67
68 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
69 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
70
71 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
72 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
73
74 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
75 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
76
77 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
78
79 w = (const int8_t*) w + 64;
80 k -= 16 * sizeof(int8_t);
81 } while (k != 0);
82
83 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));
84 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));
85
86 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
87
88 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
89
90 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
91
92 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
93
94 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
95 w = (const float*) w + 4;
96 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
97
98 const v128_t vbias0123 = wasm_v128_load(w);
99 w = (const float*) w + 4;
100 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
101
102 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
103
104 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
105
106 if XNN_LIKELY(nc >= 4) {
107 wasm_v128_store(c0, vacc0x0123);
108
109 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
110
111 c0 = (float*) ((uintptr_t) c0 + cn_stride);
112
113 nc -= 4;
114 } else {
115 if (nc & 2) {
116 wasm_v128_store64_lane(c0, vacc0x0123, 0);
117 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
118 c0 += 2;
119 }
120 if (nc & 1) {
121 wasm_v128_store32_lane(c0, vacc0x0123, 0);
122 }
123 nc = 0;
124 }
125 } while (nc != 0);
126 }