Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x4c16-minmax-wasmsdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x4c16__wasmsdot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 16 * sizeof(int8_t));
46 float* c0 = c;
47
48 const v128_t vmin = wasm_v128_load32_splat(&params->scalar.min);
49 const v128_t vmax = wasm_v128_load32_splat(&params->scalar.max);
50 XNN_FORCE_REALIZATION(vmin);
51 XNN_FORCE_REALIZATION(vmax);
52
53 do {
54 v128_t vksum0 = wasm_v128_load32_zero(w);
55 v128_t vksum1 = wasm_v128_load32_zero((const int32_t*) w + 1);
56 v128_t vksum2 = wasm_v128_load32_zero((const int32_t*) w + 2);
57 v128_t vksum3 = wasm_v128_load32_zero((const int32_t*) w + 3);
58 const v128_t vinput_zero_point = wasm_v128_load32_splat(&quantization_params->zero_point);
59 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point);
60 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point);
61 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point);
62 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point);
63 w = (const int32_t*) w + 4;
64
65 size_t p = ks;
66 do {
67 const int8_t* restrict a0 = a[0];
68 if XNN_UNPREDICTABLE(a0 != zero) {
69 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
70 } else {
71 a0 = zero_data;
72 }
73 a += 1;
74
75 size_t k = kc;
76 do {
77 const v128_t va0 = wasm_v128_load(a0);
78 a0 += 16;
79
80 const v128_t vb0 = wasm_v128_load(w);
81 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
82 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
83 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
84 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
85 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
86 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
87 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
88
89 w = (const int8_t*) w + 64;
90 k -= 16 * sizeof(int8_t);
91 } while (k != 0);
92 p -= 1 * sizeof(void*);
93 } while (p != 0);
94
95 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));
96 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));
97
98 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
99
100 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
101
102 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
103
104 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale);
105
106 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
107 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
108 w = (const float*) w + 4;
109
110 const v128_t vbias0123 = wasm_v128_load(w);
111 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
112 w = (const float*) w + 4;
113
114 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
115
116 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
117
118 if XNN_LIKELY(nc >= 4) {
119 wasm_v128_store(c0, vacc0x0123);
120
121 c0 = (float*) ((uintptr_t) c0 + cn_stride);
122
123 a = (const int8_t**restrict) ((uintptr_t) a - ks);
124
125 nc -= 4;
126 } else {
127 if (nc & 2) {
128 wasm_v128_store64_lane(c0, vacc0x0123, 0);
129 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
130 c0 += 2;
131 }
132 if (nc & 1) {
133 wasm_v128_store32_lane(c0, vacc0x0123, 0);
134 }
135 nc = 0;
136 }
137 } while (nc != 0);
138 }