Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x16c4-minmax-wasmusdot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx16c4-wasmdot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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_1x16c4__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, 4 * 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 XNN_FORCE_REALIZATION(vsign_mask);
50
51 do {
52 v128_t vksum0123 = wasm_v128_load((const int32_t*) w);
53 v128_t vksum4567 = wasm_v128_load((const int32_t*) w + 4);
54 v128_t vksum89AB = wasm_v128_load((const int32_t*) w + 8);
55 v128_t vksumCDEF = wasm_v128_load((const int32_t*) w + 12);
56 const v128_t vinput_zero_point0 = wasm_i32x4_splat((int32_t) quantization_params[0].zero_point + 128);
57 v128_t vacc0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
58 v128_t vacc0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
59 v128_t vacc0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point0);
60 v128_t vacc0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point0);
61 w = (const int32_t*) w + 16;
62
63 size_t k = kc;
64
65 while (k != 0) {
66 v128_t va0x0123 = wasm_v128_load32_splat(a0);
67 a0 += 4;
68
69 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
70
71 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
72 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
73 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
74 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
75
76 vacc0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0123);
77 vacc0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x4567);
78 vacc0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x89AB);
79 vacc0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0xCDEF);
80
81 w = (const int8_t*) w + 64;
82 k -= 4 * sizeof(int8_t);
83 }
84
85 vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
86 vacc0x4567 = wasm_f32x4_convert_i32x4(vacc0x4567);
87 vacc0x89AB = wasm_f32x4_convert_i32x4(vacc0x89AB);
88 vacc0xCDEF = wasm_f32x4_convert_i32x4(vacc0xCDEF);
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 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vinput_scale0);
94 vacc0x89AB = wasm_f32x4_mul(vacc0x89AB, vinput_scale0);
95 vacc0xCDEF = wasm_f32x4_mul(vacc0xCDEF, vinput_scale0);
96
97 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
98 w = (const float*) w + 4;
99 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
100 w = (const float*) w + 4;
101 const v128_t vfilter_output_scale89AB = wasm_v128_load(w);
102 w = (const float*) w + 4;
103 const v128_t vfilter_output_scaleCDEF = wasm_v128_load(w);
104 w = (const float*) w + 4;
105 vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vfilter_output_scale0123);
106 vacc0x4567 = wasm_f32x4_mul(vacc0x4567, vfilter_output_scale4567);
107 vacc0x89AB = wasm_f32x4_mul(vacc0x89AB, vfilter_output_scale89AB);
108 vacc0xCDEF = wasm_f32x4_mul(vacc0xCDEF, vfilter_output_scaleCDEF);
109
110 const v128_t vbias0123 = wasm_v128_load(w);
111 w = (const float*) w + 4;
112 const v128_t vbias4567 = wasm_v128_load(w);
113 w = (const float*) w + 4;
114 const v128_t vbias89AB = wasm_v128_load(w);
115 w = (const float*) w + 4;
116 const v128_t vbiasCDEF = wasm_v128_load(w);
117 w = (const float*) w + 4;
118 vacc0x0123 = wasm_f32x4_add(vacc0x0123, vbias0123);
119 vacc0x4567 = wasm_f32x4_add(vacc0x4567, vbias4567);
120 vacc0x89AB = wasm_f32x4_add(vacc0x89AB, vbias89AB);
121 vacc0xCDEF = wasm_f32x4_add(vacc0xCDEF, vbiasCDEF);
122
123 vacc0x0123 = wasm_f32x4_pmax(vacc0x0123, vmin);
124 vacc0x4567 = wasm_f32x4_pmax(vacc0x4567, vmin);
125 vacc0x89AB = wasm_f32x4_pmax(vacc0x89AB, vmin);
126 vacc0xCDEF = wasm_f32x4_pmax(vacc0xCDEF, vmin);
127
128 vacc0x0123 = wasm_f32x4_pmin(vacc0x0123, vmax);
129 vacc0x4567 = wasm_f32x4_pmin(vacc0x4567, vmax);
130 vacc0x89AB = wasm_f32x4_pmin(vacc0x89AB, vmax);
131 vacc0xCDEF = wasm_f32x4_pmin(vacc0xCDEF, vmax);
132
133 if XNN_LIKELY(nc >= 16) {
134 wasm_v128_store(c0, vacc0x0123);
135 wasm_v128_store(c0 + 4, vacc0x4567);
136 wasm_v128_store(c0 + 8, vacc0x89AB);
137 wasm_v128_store(c0 + 12, vacc0xCDEF);
138
139 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
140
141 c0 = (float*) ((uintptr_t) c0 + cn_stride);
142
143 nc -= 16;
144 } else {
145 if (nc & 8) {
146 wasm_v128_store(c0, vacc0x0123);
147 vacc0x0123 = vacc0x89AB;
148 c0 += 4;
149 wasm_v128_store(c0, vacc0x4567);
150 vacc0x4567 = vacc0xCDEF;
151 c0 += 4;
152 }
153 if (nc & 4) {
154 wasm_v128_store(c0, vacc0x0123);
155 vacc0x0123 = vacc0x4567;
156 c0 += 4;
157 }
158 if (nc & 2) {
159 wasm_v128_store64_lane(c0, vacc0x0123, 0);
160 vacc0x0123 = wasm_v64x2_shuffle(vacc0x0123, vacc0x0123, 1, 1);
161 c0 += 2;
162 }
163 if (nc & 1) {
164 wasm_v128_store32_lane(c0, vacc0x0123, 0);
165 }
166 nc = 0;
167 }
168 } while (nc != 0);
169 }