1 // Copyright 2020 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <wasm_simd128.h>
9
10 #include <xnnpack/fill.h>
11
12
xnn_xx_fill_ukernel__wasmsimd_x64(size_t rows,size_t channels,void * output,size_t output_stride,const uint32_t fill_pattern)13 void xnn_xx_fill_ukernel__wasmsimd_x64(
14 size_t rows,
15 size_t channels,
16 void* output,
17 size_t output_stride,
18 const uint32_t fill_pattern)
19 {
20 assert(rows != 0);
21 assert(channels != 0);
22
23 const size_t output_increment = output_stride - channels;
24
25 const v128_t vfill_pattern = wasm_i32x4_splat(fill_pattern);
26 do {
27 size_t c = channels;
28 for (; c >= 64 * sizeof(uint8_t); c -= 64 * sizeof(uint8_t)) {
29 wasm_v128_store(output, vfill_pattern);
30 wasm_v128_store((uint8_t*) output + 16, vfill_pattern);
31 wasm_v128_store((uint8_t*) output + 32, vfill_pattern);
32 wasm_v128_store((uint8_t*) output + 48, vfill_pattern);
33 output = ((uint8_t*) output + 64);
34 }
35 for (; c >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) {
36 wasm_v128_store(output, vfill_pattern);
37 output = ((uint8_t*) output + 16);
38 }
39 if XNN_UNLIKELY(c != 0) {
40 if XNN_LIKELY(c & (8 * sizeof(uint8_t))) {
41 *((double*) output) = wasm_f64x2_extract_lane(vfill_pattern, 0);
42 output = ((uint8_t*) output + 8);
43 }
44 uint32_t vfill_subpattern = fill_pattern;
45 if XNN_LIKELY(c & (4 * sizeof(uint8_t))) {
46 *((uint32_t*) output) = vfill_subpattern;
47 output = ((uint8_t*) output + 4);
48 }
49 if XNN_LIKELY(c & (2 * sizeof(uint8_t))) {
50 *((uint16_t*) output) = (uint16_t) vfill_subpattern;
51 vfill_subpattern >>= 16;
52 output = ((uint8_t*) output + 2);
53 }
54 if XNN_LIKELY(c & (1 * sizeof(uint8_t))) {
55 *((uint8_t*) output) = (uint8_t) vfill_subpattern;
56 output = ((uint8_t*) output + 1);
57 }
58 }
59 output = (void*) ((uintptr_t) output + output_increment);
60 } while (--rows != 0);
61 }
62