Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x8c16-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_1x8c16__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 v128_t vksum4 = wasm_v128_load32_zero((const int32_t*) w + 4);
55 v128_t vksum5 = wasm_v128_load32_zero((const int32_t*) w + 5);
56 v128_t vksum6 = wasm_v128_load32_zero((const int32_t*) w + 6);
57 v128_t vksum7 = wasm_v128_load32_zero((const int32_t*) w + 7);
58 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
59 v128_t vacc0x0 = wasm_i32x4_mul(vksum0, vinput_zero_point0);
60 v128_t vacc0x1 = wasm_i32x4_mul(vksum1, vinput_zero_point0);
61 v128_t vacc0x2 = wasm_i32x4_mul(vksum2, vinput_zero_point0);
62 v128_t vacc0x3 = wasm_i32x4_mul(vksum3, vinput_zero_point0);
63 v128_t vacc0x4 = wasm_i32x4_mul(vksum4, vinput_zero_point0);
64 v128_t vacc0x5 = wasm_i32x4_mul(vksum5, vinput_zero_point0);
65 v128_t vacc0x6 = wasm_i32x4_mul(vksum6, vinput_zero_point0);
66 v128_t vacc0x7 = wasm_i32x4_mul(vksum7, vinput_zero_point0);
67 w = (const int32_t*) w + 8;
68
69 size_t k = kc;
70 do {
71 const v128_t va0 = wasm_v128_xor(wasm_v128_load(a0), vsign_mask);
72 a0 += 16;
73
74 const v128_t vb0 = wasm_v128_load(w);
75
76 vacc0x0 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0, va0, vacc0x0);
77 const v128_t vb1 = wasm_v128_load((const int8_t*) w + 16);
78
79 vacc0x1 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1, va0, vacc0x1);
80 const v128_t vb2 = wasm_v128_load((const int8_t*) w + 32);
81
82 vacc0x2 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb2, va0, vacc0x2);
83 const v128_t vb3 = wasm_v128_load((const int8_t*) w + 48);
84
85 vacc0x3 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb3, va0, vacc0x3);
86 const v128_t vb4 = wasm_v128_load((const int8_t*) w + 64);
87
88 vacc0x4 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4, va0, vacc0x4);
89 const v128_t vb5 = wasm_v128_load((const int8_t*) w + 80);
90
91 vacc0x5 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb5, va0, vacc0x5);
92 const v128_t vb6 = wasm_v128_load((const int8_t*) w + 96);
93
94 vacc0x6 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb6, va0, vacc0x6);
95 const v128_t vb7 = wasm_v128_load((const int8_t*) w + 112);
96
97 vacc0x7 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb7, va0, vacc0x7);
98
99 w = (const int8_t*) w + 128;
100 k -= 16 * sizeof(int8_t);
101 } while (k != 0);
102
103 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));
104 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));
105 const v128_t vacc0x46 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x4, vacc0x6, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x4, vacc0x6, 2, 6, 3, 7));
106 const v128_t vacc0x57 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x5, vacc0x7, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x5, vacc0x7, 2, 6, 3, 7));
107
108 v128_t vacc0x0123 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x02, vacc0x13, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x02, vacc0x13, 2, 6, 3, 7));
109 v128_t vacc0x4567 = wasm_i32x4_add(wasm_v32x4_shuffle(vacc0x46, vacc0x57, 0, 4, 1, 5), wasm_v32x4_shuffle(vacc0x46, vacc0x57, 2, 6, 3, 7));
110
111 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
112 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
113
114 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
115
116 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vinput_scale0);
117 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
118
119 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
120 w = (const float*) w + 4;
121 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
122 w = (const float*) w + 4;
123 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
124 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
125
126 const v128_t vbias0123 = wasm_v128_load(w);
127 w = (const float*) w + 4;
128 const v128_t vbias4567 = wasm_v128_load(w);
129 w = (const float*) w + 4;
130 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
131 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
132
133 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
134 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
135
136 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
137 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
138
139 if XNN_LIKELY(nc >= 8) {
140 wasm_v128_store(c0, vacc0x0123);
141 wasm_v128_store(c0 + 4, vacc0x4567);
142
143 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
144
145 c0 = (float*) ((uintptr_t) c0 + cn_stride);
146
147 nc -= 8;
148 } else {
149 if (nc & 4) {
150 wasm_v128_store(c0, vacc0x0123);
151 vacc0x0123 = vacc0x4567;
152 c0 += 4;
153 }
154 if (nc & 2) {
155 wasm_v128_store64_lane(c0, vacc0x0123, 0);
156 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
157 c0 += 2;
158 }
159 if (nc & 1) {
160 wasm_v128_store32_lane(c0, vacc0x0123, 0);
161 }
162 nc = 0;
163 }
164 } while (nc != 0);
165 }