Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-1x16c4-minmax-wasmusdot-u2.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_u2(
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 vacc0x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point0);
58 v128_t vacc0x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point0);
59 v128_t vacc0x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point0);
60 v128_t vacc0x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point0);
61 w = (const int32_t*) w + 16;
62
63 size_t k = kc;
64 while (k >= 8 * sizeof(int8_t)) {
65 v128_t va0x0x0123 = wasm_v128_load32_splat(a0);
66 v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4);
67 a0 += 8;
68
69 va0x0x0123 = wasm_v128_xor(va0x0x0123, vsign_mask);
70 va0x1x0123 = wasm_v128_xor(va0x1x0123, vsign_mask);
71
72 const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w);
73 const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16);
74 const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32);
75 const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48);
76 const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64);
77 const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80);
78 const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96);
79 const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112);
80
81 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123);
82 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567);
83 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB);
84 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF);
85 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x0x0123);
86 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x0x4567);
87 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x0x89AB);
88 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x0xCDEF);
89
90 w = (const int8_t*) w + 128;
91 k -= 8 * sizeof(int8_t);
92 }
93
94 while (k != 0) {
95 v128_t va0x0123 = wasm_v128_load32_splat(a0);
96 a0 += 4;
97
98 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
99
100 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
101 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
102 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
103 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
104
105 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123);
106 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567);
107 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB);
108 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF);
109
110 w = (const int8_t*) w + 64;
111 k -= 4 * sizeof(int8_t);
112 }
113
114 vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123);
115 vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567);
116 vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB);
117 vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF);
118
119 const v128_t vinput_scale0 = wasm_v128_load32_splat(&quantization_params[0].inv_scale);
120
121 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vinput_scale0);
122 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vinput_scale0);
123 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vinput_scale0);
124 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vinput_scale0);
125
126 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
127 w = (const float*) w + 4;
128 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
129 w = (const float*) w + 4;
130 const v128_t vfilter_output_scale89AB = wasm_v128_load(w);
131 w = (const float*) w + 4;
132 const v128_t vfilter_output_scaleCDEF = wasm_v128_load(w);
133 w = (const float*) w + 4;
134 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vfilter_output_scale0123);
135 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vfilter_output_scale4567);
136 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vfilter_output_scale89AB);
137 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vfilter_output_scaleCDEF);
138
139 const v128_t vbias0123 = wasm_v128_load(w);
140 w = (const float*) w + 4;
141 const v128_t vbias4567 = wasm_v128_load(w);
142 w = (const float*) w + 4;
143 const v128_t vbias89AB = wasm_v128_load(w);
144 w = (const float*) w + 4;
145 const v128_t vbiasCDEF = wasm_v128_load(w);
146 w = (const float*) w + 4;
147 vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vbias0123);
148 vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vbias4567);
149 vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vbias89AB);
150 vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vbiasCDEF);
151
152 vacc0x0x0123 = wasm_f32x4_pmax(vacc0x0x0123, vmin);
153 vacc0x0x4567 = wasm_f32x4_pmax(vacc0x0x4567, vmin);
154 vacc0x0x89AB = wasm_f32x4_pmax(vacc0x0x89AB, vmin);
155 vacc0x0xCDEF = wasm_f32x4_pmax(vacc0x0xCDEF, vmin);
156
157 vacc0x0x0123 = wasm_f32x4_pmin(vacc0x0x0123, vmax);
158 vacc0x0x4567 = wasm_f32x4_pmin(vacc0x0x4567, vmax);
159 vacc0x0x89AB = wasm_f32x4_pmin(vacc0x0x89AB, vmax);
160 vacc0x0xCDEF = wasm_f32x4_pmin(vacc0x0xCDEF, vmax);
161
162 if XNN_LIKELY(nc >= 16) {
163 wasm_v128_store(c0, vacc0x0x0123);
164 wasm_v128_store(c0 + 4, vacc0x0x4567);
165 wasm_v128_store(c0 + 8, vacc0x0x89AB);
166 wasm_v128_store(c0 + 12, vacc0x0xCDEF);
167
168 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
169
170 c0 = (float*) ((uintptr_t) c0 + cn_stride);
171
172 nc -= 16;
173 } else {
174 if (nc & 8) {
175 wasm_v128_store(c0, vacc0x0x0123);
176 vacc0x0x0123 = vacc0x0x89AB;
177 c0 += 4;
178 wasm_v128_store(c0, vacc0x0x4567);
179 vacc0x0x4567 = vacc0x0xCDEF;
180 c0 += 4;
181 }
182 if (nc & 4) {
183 wasm_v128_store(c0, vacc0x0x0123);
184 vacc0x0x0123 = vacc0x0x4567;
185 c0 += 4;
186 }
187 if (nc & 2) {
188 wasm_v128_store64_lane(c0, vacc0x0x0123, 0);
189 vacc0x0x0123 = wasm_v64x2_shuffle(vacc0x0x0123, vacc0x0x0123, 1, 1);
190 c0 += 2;
191 }
192 if (nc & 1) {
193 wasm_v128_store32_lane(c0, vacc0x0x0123, 0);
194 }
195 nc = 0;
196 }
197 } while (nc != 0);
198 }