/
32x1-minmax-wasmsimd-x86.c
264 lines (253 loc) · 10.9 KB
/
32x1-minmax-wasmsimd-x86.c
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
// Auto-generated file. Do not edit!
// Template: src/f32-spmm/wasmsimd.c.in
// Generator: tools/xngen
//
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <assert.h>
#include <wasm_simd128.h>
#include <xnnpack/spmm.h>
void xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86(
size_t mc,
size_t nc,
const float*restrict input,
const float*restrict weights,
const int32_t*restrict widx_dmap,
const uint32_t*restrict nidx_nnzmap,
float*restrict output,
size_t output_stride,
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
assert(mc != 0);
assert(mc % sizeof(float) == 0);
assert(nc != 0);
const v128_t vmin = wasm_v128_load32_splat(¶ms->scalar.min);
const v128_t vmax = wasm_v128_load32_splat(¶ms->scalar.max);
size_t output_decrement = output_stride * nc - 32 * sizeof(float);
while XNN_LIKELY(mc >= 32 * sizeof(float)) {
const float*restrict w = weights;
const int32_t* dmap = widx_dmap;
const uint32_t* nnzmap = nidx_nnzmap;
size_t n = nc;
do {
uint32_t nnz = *nnzmap++;
v128_t vacc0123 = wasm_v128_load32_splat(w); w += 1;
v128_t vacc4567 = vacc0123;
v128_t vacc89AB = vacc0123;
v128_t vaccCDEF = vacc0123;
v128_t vaccGHIJ = vacc0123;
v128_t vaccKLMN = vacc0123;
v128_t vaccOPQR = vacc0123;
v128_t vaccSTUV = vacc0123;
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
const v128_t vi0123 = wasm_v128_load(input);
const v128_t vi4567 = wasm_v128_load(input + 4);
const v128_t vi89AB = wasm_v128_load(input + 8);
const v128_t viCDEF = wasm_v128_load(input + 12);
const v128_t viGHIJ = wasm_v128_load(input + 16);
const v128_t viKLMN = wasm_v128_load(input + 20);
const v128_t viOPQR = wasm_v128_load(input + 24);
const v128_t viSTUV = wasm_v128_load(input + 28);
input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
const v128_t vw = wasm_v128_load32_splat(w); w += 1;
vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw));
vacc89AB = wasm_f32x4_add(vacc89AB, wasm_f32x4_mul(vi89AB, vw));
vaccCDEF = wasm_f32x4_add(vaccCDEF, wasm_f32x4_mul(viCDEF, vw));
vaccGHIJ = wasm_f32x4_add(vaccGHIJ, wasm_f32x4_mul(viGHIJ, vw));
vaccKLMN = wasm_f32x4_add(vaccKLMN, wasm_f32x4_mul(viKLMN, vw));
vaccOPQR = wasm_f32x4_add(vaccOPQR, wasm_f32x4_mul(viOPQR, vw));
vaccSTUV = wasm_f32x4_add(vaccSTUV, wasm_f32x4_mul(viSTUV, vw));
} while (--nnz != 0);
}
v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
v128_t vout4567 = wasm_f32x4_pmin(vmax, vacc4567);
v128_t vout89AB = wasm_f32x4_pmin(vmax, vacc89AB);
v128_t voutCDEF = wasm_f32x4_pmin(vmax, vaccCDEF);
v128_t voutGHIJ = wasm_f32x4_pmin(vmax, vaccGHIJ);
v128_t voutKLMN = wasm_f32x4_pmin(vmax, vaccKLMN);
v128_t voutOPQR = wasm_f32x4_pmin(vmax, vaccOPQR);
v128_t voutSTUV = wasm_f32x4_pmin(vmax, vaccSTUV);
vout0123 = wasm_f32x4_pmax(vmin, vout0123);
vout4567 = wasm_f32x4_pmax(vmin, vout4567);
vout89AB = wasm_f32x4_pmax(vmin, vout89AB);
voutCDEF = wasm_f32x4_pmax(vmin, voutCDEF);
voutGHIJ = wasm_f32x4_pmax(vmin, voutGHIJ);
voutKLMN = wasm_f32x4_pmax(vmin, voutKLMN);
voutOPQR = wasm_f32x4_pmax(vmin, voutOPQR);
voutSTUV = wasm_f32x4_pmax(vmin, voutSTUV);
wasm_v128_store(output, vout0123);
wasm_v128_store(output + 4, vout4567);
wasm_v128_store(output + 8, vout89AB);
wasm_v128_store(output + 12, voutCDEF);
wasm_v128_store(output + 16, voutGHIJ);
wasm_v128_store(output + 20, voutKLMN);
wasm_v128_store(output + 24, voutOPQR);
wasm_v128_store(output + 28, voutSTUV);
output = (float*restrict) ((uintptr_t) output + output_stride);
} while (--n != 0);
output = (float*restrict) ((uintptr_t) output - output_decrement);
input += 32;
mc -= 32 * sizeof(float);
}
if XNN_UNLIKELY(mc != 0) {
output_decrement += 16 * sizeof(float);
if (mc & (16 * sizeof(float))) {
const float*restrict w = weights;
const int32_t* dmap = widx_dmap;
const uint32_t* nnzmap = nidx_nnzmap;
size_t n = nc;
do {
uint32_t nnz = *nnzmap++;
v128_t vacc0123 = wasm_v128_load32_splat(w); w += 1;
v128_t vacc4567 = vacc0123;
v128_t vacc89AB = vacc0123;
v128_t vaccCDEF = vacc0123;
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
const v128_t vi0123 = wasm_v128_load(input);
const v128_t vi4567 = wasm_v128_load(input + 4);
const v128_t vi89AB = wasm_v128_load(input + 8);
const v128_t viCDEF = wasm_v128_load(input + 12);
input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
const v128_t vw = wasm_v128_load32_splat(w); w += 1;
vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw));
vacc89AB = wasm_f32x4_add(vacc89AB, wasm_f32x4_mul(vi89AB, vw));
vaccCDEF = wasm_f32x4_add(vaccCDEF, wasm_f32x4_mul(viCDEF, vw));
} while (--nnz != 0);
}
v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
v128_t vout4567 = wasm_f32x4_pmin(vmax, vacc4567);
v128_t vout89AB = wasm_f32x4_pmin(vmax, vacc89AB);
v128_t voutCDEF = wasm_f32x4_pmin(vmax, vaccCDEF);
vout0123 = wasm_f32x4_pmax(vmin, vout0123);
vout4567 = wasm_f32x4_pmax(vmin, vout4567);
vout89AB = wasm_f32x4_pmax(vmin, vout89AB);
voutCDEF = wasm_f32x4_pmax(vmin, voutCDEF);
wasm_v128_store(output, vout0123);
wasm_v128_store(output + 4, vout4567);
wasm_v128_store(output + 8, vout89AB);
wasm_v128_store(output + 12, voutCDEF);
output = (float*restrict) ((uintptr_t) output + output_stride);
} while (--n != 0);
output = (float*restrict) ((uintptr_t) output - output_decrement);
input += 16;
}
output_decrement += 8 * sizeof(float);
if (mc & (8 * sizeof(float))) {
const float*restrict w = weights;
const int32_t* dmap = widx_dmap;
const uint32_t* nnzmap = nidx_nnzmap;
size_t n = nc;
do {
uint32_t nnz = *nnzmap++;
v128_t vacc0123 = wasm_v128_load32_splat(w); w += 1;
v128_t vacc4567 = vacc0123;
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
const v128_t vi0123 = wasm_v128_load(input);
const v128_t vi4567 = wasm_v128_load(input + 4);
input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
const v128_t vw = wasm_v128_load32_splat(w); w += 1;
vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
vacc4567 = wasm_f32x4_add(vacc4567, wasm_f32x4_mul(vi4567, vw));
} while (--nnz != 0);
}
v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
v128_t vout4567 = wasm_f32x4_pmin(vmax, vacc4567);
vout0123 = wasm_f32x4_pmax(vmin, vout0123);
vout4567 = wasm_f32x4_pmax(vmin, vout4567);
wasm_v128_store(output, vout0123);
wasm_v128_store(output + 4, vout4567);
output = (float*restrict) ((uintptr_t) output + output_stride);
} while (--n != 0);
output = (float*restrict) ((uintptr_t) output - output_decrement);
input += 8;
}
output_decrement += 4 * sizeof(float);
if (mc & (4 * sizeof(float))) {
const float*restrict w = weights;
const int32_t* dmap = widx_dmap;
const uint32_t* nnzmap = nidx_nnzmap;
size_t n = nc;
do {
uint32_t nnz = *nnzmap++;
v128_t vacc0123 = wasm_v128_load32_splat(w); w += 1;
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
const v128_t vi0123 = wasm_v128_load(input);
input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
const v128_t vw = wasm_v128_load32_splat(w); w += 1;
vacc0123 = wasm_f32x4_add(vacc0123, wasm_f32x4_mul(vi0123, vw));
} while (--nnz != 0);
}
v128_t vout0123 = wasm_f32x4_pmin(vmax, vacc0123);
vout0123 = wasm_f32x4_pmax(vmin, vout0123);
wasm_v128_store(output, vout0123);
output = (float*restrict) ((uintptr_t) output + output_stride);
} while (--n != 0);
output = (float*restrict) ((uintptr_t) output - output_decrement);
input += 4;
}
output_decrement += 2 * sizeof(float);
if (mc & (2 * sizeof(float))) {
const float*restrict w = weights;
const int32_t* dmap = widx_dmap;
const uint32_t* nnzmap = nidx_nnzmap;
size_t n = nc;
do {
uint32_t nnz = *nnzmap++;
v128_t vacc01 = wasm_v128_load32_splat(w); w += 1;
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
const v128_t vi01 = wasm_v128_load64_splat(input);
input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
const v128_t vw = wasm_v128_load32_splat(w); w += 1;
vacc01 = wasm_f32x4_add(vacc01, wasm_f32x4_mul(vi01, vw));
} while (--nnz != 0);
}
v128_t vout01 = wasm_f32x4_pmin(vmax, vacc01);
vout01 = wasm_f32x4_pmax(vmin, vout01);
*((double*) output) = wasm_f64x2_extract_lane(vout01, 0);
output = (float*restrict) ((uintptr_t) output + output_stride);
} while (--n != 0);
output = (float*restrict) ((uintptr_t) output - output_decrement);
input += 2;
}
output_decrement += 1 * sizeof(float);
if (mc & (1 * sizeof(float))) {
const float*restrict w = weights;
const int32_t* dmap = widx_dmap;
const uint32_t* nnzmap = nidx_nnzmap;
size_t n = nc;
do {
uint32_t nnz = *nnzmap++;
v128_t vacc0 = wasm_v128_load32_splat(w); w += 1;
if XNN_LIKELY(nnz != 0) {
do {
const intptr_t diff = *dmap++;
const v128_t vi0 = wasm_v128_load32_splat(input);
input = (const float*restrict) ((uintptr_t) input + (uintptr_t) diff);
const v128_t vw = wasm_v128_load32_splat(w); w += 1;
vacc0 = wasm_f32x4_add(vacc0, wasm_f32x4_mul(vi0, vw));
} while (--nnz != 0);
}
v128_t vout0 = wasm_f32x4_pmin(vmax, vacc0);
vout0 = wasm_f32x4_pmax(vmin, vout0);
*output = wasm_f32x4_extract_lane(vout0, 0);
output = (float*restrict) ((uintptr_t) output + output_stride);
} while (--n != 0);
output = (float*restrict) ((uintptr_t) output - output_decrement);
input += 1;
}
}
}