1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker
6*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
7*4bdc9457SAndroid Build Coastguard Worker
8*4bdc9457SAndroid Build Coastguard Worker #include <wasm_simd128.h>
9*4bdc9457SAndroid Build Coastguard Worker
10*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/fill.h>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker
xnn_xx_fill_ukernel__wasmsimd_x64(size_t rows,size_t channels,void * output,size_t output_stride,const uint32_t fill_pattern)13*4bdc9457SAndroid Build Coastguard Worker void xnn_xx_fill_ukernel__wasmsimd_x64(
14*4bdc9457SAndroid Build Coastguard Worker size_t rows,
15*4bdc9457SAndroid Build Coastguard Worker size_t channels,
16*4bdc9457SAndroid Build Coastguard Worker void* output,
17*4bdc9457SAndroid Build Coastguard Worker size_t output_stride,
18*4bdc9457SAndroid Build Coastguard Worker const uint32_t fill_pattern)
19*4bdc9457SAndroid Build Coastguard Worker {
20*4bdc9457SAndroid Build Coastguard Worker assert(rows != 0);
21*4bdc9457SAndroid Build Coastguard Worker assert(channels != 0);
22*4bdc9457SAndroid Build Coastguard Worker
23*4bdc9457SAndroid Build Coastguard Worker const size_t output_increment = output_stride - channels;
24*4bdc9457SAndroid Build Coastguard Worker
25*4bdc9457SAndroid Build Coastguard Worker const v128_t vfill_pattern = wasm_i32x4_splat(fill_pattern);
26*4bdc9457SAndroid Build Coastguard Worker do {
27*4bdc9457SAndroid Build Coastguard Worker size_t c = channels;
28*4bdc9457SAndroid Build Coastguard Worker for (; c >= 64 * sizeof(uint8_t); c -= 64 * sizeof(uint8_t)) {
29*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(output, vfill_pattern);
30*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store((uint8_t*) output + 16, vfill_pattern);
31*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store((uint8_t*) output + 32, vfill_pattern);
32*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store((uint8_t*) output + 48, vfill_pattern);
33*4bdc9457SAndroid Build Coastguard Worker output = ((uint8_t*) output + 64);
34*4bdc9457SAndroid Build Coastguard Worker }
35*4bdc9457SAndroid Build Coastguard Worker for (; c >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) {
36*4bdc9457SAndroid Build Coastguard Worker wasm_v128_store(output, vfill_pattern);
37*4bdc9457SAndroid Build Coastguard Worker output = ((uint8_t*) output + 16);
38*4bdc9457SAndroid Build Coastguard Worker }
39*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(c != 0) {
40*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(c & (8 * sizeof(uint8_t))) {
41*4bdc9457SAndroid Build Coastguard Worker *((double*) output) = wasm_f64x2_extract_lane(vfill_pattern, 0);
42*4bdc9457SAndroid Build Coastguard Worker output = ((uint8_t*) output + 8);
43*4bdc9457SAndroid Build Coastguard Worker }
44*4bdc9457SAndroid Build Coastguard Worker uint32_t vfill_subpattern = fill_pattern;
45*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(c & (4 * sizeof(uint8_t))) {
46*4bdc9457SAndroid Build Coastguard Worker *((uint32_t*) output) = vfill_subpattern;
47*4bdc9457SAndroid Build Coastguard Worker output = ((uint8_t*) output + 4);
48*4bdc9457SAndroid Build Coastguard Worker }
49*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(c & (2 * sizeof(uint8_t))) {
50*4bdc9457SAndroid Build Coastguard Worker *((uint16_t*) output) = (uint16_t) vfill_subpattern;
51*4bdc9457SAndroid Build Coastguard Worker vfill_subpattern >>= 16;
52*4bdc9457SAndroid Build Coastguard Worker output = ((uint8_t*) output + 2);
53*4bdc9457SAndroid Build Coastguard Worker }
54*4bdc9457SAndroid Build Coastguard Worker if XNN_LIKELY(c & (1 * sizeof(uint8_t))) {
55*4bdc9457SAndroid Build Coastguard Worker *((uint8_t*) output) = (uint8_t) vfill_subpattern;
56*4bdc9457SAndroid Build Coastguard Worker output = ((uint8_t*) output + 1);
57*4bdc9457SAndroid Build Coastguard Worker }
58*4bdc9457SAndroid Build Coastguard Worker }
59*4bdc9457SAndroid Build Coastguard Worker output = (void*) ((uintptr_t) output + output_increment);
60*4bdc9457SAndroid Build Coastguard Worker } while (--rows != 0);
61*4bdc9457SAndroid Build Coastguard Worker }
62