Debian Codegraph @

src/qd8-f32-qc8w-igemm/gen/qd8-f32-qc8w-igemm-1x16c4-minmax-wasmusdot-u2.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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
18 void xnn_qd8_f32_qc8w_igemm_minmax_ukernel_1x16c4__wasmusdot_u2(
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, 4 * 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 const v128_t vsign_mask = wasm_u8x16_const_splat(UINT8_C(0x80));
54 XNN_FORCE_REALIZATION(vsign_mask);
55 do {
56 v128_t vksum0123 = wasm_v128_load((const int32_t*) w);
57 v128_t vksum4567 = wasm_v128_load((const int32_t*) w + 4);
58 v128_t vksum89AB = wasm_v128_load((const int32_t*) w + 8);
59 v128_t vksumCDEF = wasm_v128_load((const int32_t*) w + 12);
60 const v128_t vinput_zero_point = wasm_i32x4_splat((int32_t) quantization_params->zero_point + 128);
61 v128_t vacc0x0x0123 = wasm_i32x4_mul(vksum0123, vinput_zero_point);
62 v128_t vacc0x0x4567 = wasm_i32x4_mul(vksum4567, vinput_zero_point);
63 v128_t vacc0x0x89AB = wasm_i32x4_mul(vksum89AB, vinput_zero_point);
64 v128_t vacc0x0xCDEF = wasm_i32x4_mul(vksumCDEF, vinput_zero_point);
65 w = (const int32_t*) w + 16;
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72 } else {
73 a0 = zero_data;
74 }
75 a += 1;
76
77 size_t k = kc;
78 while (k >= 8 * sizeof(int8_t)) {
79 v128_t va0x0x0123 = wasm_v128_load32_splat(a0);
80 v128_t va0x1x0123 = wasm_v128_load32_splat(a0 + 4);
81 a0 += 8;
82
83 va0x0x0123 = wasm_v128_xor(va0x0x0123, vsign_mask);
84 va0x1x0123 = wasm_v128_xor(va0x1x0123, vsign_mask);
85
86 const v128_t vb0x0123 = wasm_v128_load((const int8_t*) w);
87 const v128_t vb0x4567 = wasm_v128_load((const int8_t*) w + 16);
88 const v128_t vb0x89AB = wasm_v128_load((const int8_t*) w + 32);
89 const v128_t vb0xCDEF = wasm_v128_load((const int8_t*) w + 48);
90 const v128_t vb1x0123 = wasm_v128_load((const int8_t*) w + 64);
91 const v128_t vb1x4567 = wasm_v128_load((const int8_t*) w + 80);
92 const v128_t vb1x89AB = wasm_v128_load((const int8_t*) w + 96);
93 const v128_t vb1xCDEF = wasm_v128_load((const int8_t*) w + 112);
94
95 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x0123, va0x0x0123, vacc0x0x0123);
96 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x4567, va0x0x0123, vacc0x0x4567);
97 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0x89AB, va0x0x0123, vacc0x0x89AB);
98 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0xCDEF, va0x0x0123, vacc0x0xCDEF);
99 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x0123, va0x1x0123, vacc0x0x0123);
100 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x4567, va0x1x0123, vacc0x0x4567);
101 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1x89AB, va0x1x0123, vacc0x0x89AB);
102 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb1xCDEF, va0x1x0123, vacc0x0xCDEF);
103
104 w = (const int8_t*) w + 128;
105 k -= 8 * sizeof(int8_t);
106 }
107
108 while (k != 0) {
109 v128_t va0x0123 = wasm_v128_load32_splat(a0);
110 a0 += 4;
111
112 va0x0123 = wasm_v128_xor(va0x0123, vsign_mask);
113
114 const v128_t vb0123 = wasm_v128_load((const int8_t*) w);
115 const v128_t vb4567 = wasm_v128_load((const int8_t*) w + 16);
116 const v128_t vb89AB = wasm_v128_load((const int8_t*) w + 32);
117 const v128_t vbCDEF = wasm_v128_load((const int8_t*) w + 48);
118
119 vacc0x0x0123 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb0123, va0x0123, vacc0x0x0123);
120 vacc0x0x4567 = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb4567, va0x0123, vacc0x0x4567);
121 vacc0x0x89AB = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vb89AB, va0x0123, vacc0x0x89AB);
122 vacc0x0xCDEF = wasm_i32x4_relaxed_dot_i8x16_i7x16_add(vbCDEF, va0x0123, vacc0x0xCDEF);
123
124 w = (const int8_t*) w + 64;
125 k -= 4 * sizeof(int8_t);
126 }
127
128 p -= 1 * sizeof(void*);
129 } while (p != 0);
130
131 vacc0x0x0123 = wasm_f32x4_convert_i32x4(vacc0x0x0123);
132 vacc0x0x4567 = wasm_f32x4_convert_i32x4(vacc0x0x4567);
133 vacc0x0x89AB = wasm_f32x4_convert_i32x4(vacc0x0x89AB);
134 vacc0x0xCDEF = wasm_f32x4_convert_i32x4(vacc0x0xCDEF);
135
136 const v128_t vinput_scale = wasm_v128_load32_splat(&quantization_params->inv_scale);
137
138 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vinput_scale);
139 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vinput_scale);
140 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vinput_scale);
141 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vinput_scale);
142
143 const v128_t vfilter_output_scale0123 = wasm_v128_load(w);
144 w = (const float*) w + 4;
145 const v128_t vfilter_output_scale4567 = wasm_v128_load(w);
146 w = (const float*) w + 4;
147 const v128_t vfilter_output_scale89AB = wasm_v128_load(w);
148 w = (const float*) w + 4;
149 const v128_t vfilter_output_scaleCDEF = wasm_v128_load(w);
150 w = (const float*) w + 4;
151 vacc0x0x0123 = wasm_f32x4_mul(vacc0x0x0123, vfilter_output_scale0123);
152 vacc0x0x4567 = wasm_f32x4_mul(vacc0x0x4567, vfilter_output_scale4567);
153 vacc0x0x89AB = wasm_f32x4_mul(vacc0x0x89AB, vfilter_output_scale89AB);
154 vacc0x0xCDEF = wasm_f32x4_mul(vacc0x0xCDEF, vfilter_output_scaleCDEF);
155
156 const v128_t vbias0123 = wasm_v128_load(w);
157 w = (const float*) w + 4;
158 const v128_t vbias4567 = wasm_v128_load(w);
159 w = (const float*) w + 4;
160 const v128_t vbias89AB = wasm_v128_load(w);
161 w = (const float*) w + 4;
162 const v128_t vbiasCDEF = wasm_v128_load(w);
163 w = (const float*) w + 4;
164 vacc0x0x0123 = wasm_f32x4_add(vacc0x0x0123, vbias0123);
165 vacc0x0x4567 = wasm_f32x4_add(vacc0x0x4567, vbias4567);
166 vacc0x0x89AB = wasm_f32x4_add(vacc0x0x89AB, vbias89AB);
167 vacc0x0xCDEF = wasm_f32x4_add(vacc0x0xCDEF, vbiasCDEF);
168
169 vacc0x0x0123 = wasm_f32x4_pmax(vacc0x0x0123, vmin);
170 vacc0x0x4567 = wasm_f32x4_pmax(vacc0x0x4567, vmin);
171 vacc0x0x89AB = wasm_f32x4_pmax(vacc0x0x89AB, vmin);
172 vacc0x0xCDEF = wasm_f32x4_pmax(vacc0x0xCDEF, vmin);
173
174 vacc0x0x0123 = wasm_f32x4_pmin(vacc0x0x0123, vmax);
175 vacc0x0x4567 = wasm_f32x4_pmin(vacc0x0x4567, vmax);
176 vacc0x0x89AB = wasm_f32x4_pmin(vacc0x0x89AB, vmax);
177 vacc0x0xCDEF = wasm_f32x4_pmin(vacc0x0xCDEF, vmax);
178
179 if XNN_LIKELY(nc >= 16) {
180 wasm_v128_store(c0, vacc0x0x0123);
181 wasm_v128_store(c0 + 4, vacc0x0x4567);
182 wasm_v128_store(c0 + 8, vacc0x0x89AB);
183 wasm_v128_store(c0 + 12, vacc0x0xCDEF);
184
185 c0 = (float*) ((uintptr_t) c0 + cn_stride);
186
187 a = (const int8_t**restrict) ((uintptr_t) a - ks);
188
189 nc -= 16;
190 } else {
191 if (nc & 8) {
192 wasm_v128_store(c0, vacc0x0x0123);
193 vacc0x0x0123 = vacc0x0x89AB;
194 c0 += 4;
195 wasm_v128_store(c0, vacc0x0x4567);
196 vacc0x0x4567 = vacc0x0xCDEF;
197 c0 += 4;
198 }
199 if (nc & 4) {
200 wasm_v128_store(c0, vacc0x0x0123);
201 vacc0x0x0123 = vacc0x0x4567;
202 c0 += 4;
203 }
204 if (nc & 2) {
205 wasm_v128_store64_lane(c0, vacc0x0x0123, 0);
206 vacc0x0x0123 = wasm_v64x2_shuffle(vacc0x0x0123, vacc0x0x0123, 1, 1);
207 c0 += 2;
208 }
209 if (nc & 1) {
210 wasm_v128_store32_lane(c0, vacc0x0x0123, 0);
211 }
212 nc = 0;
213 }
214 } while (nc != 0);
215 }