xref: /aosp_15_r20/external/XNNPACK/src/xx-fill/sse2-x64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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 <emmintrin.h>
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/fill.h>
11*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/unaligned.h>
12*4bdc9457SAndroid Build Coastguard Worker 
13*4bdc9457SAndroid Build Coastguard Worker 
xnn_xx_fill_ukernel__sse2_x64(size_t rows,size_t channels,void * output,size_t output_stride,const uint32_t fill_pattern)14*4bdc9457SAndroid Build Coastguard Worker void xnn_xx_fill_ukernel__sse2_x64(
15*4bdc9457SAndroid Build Coastguard Worker     size_t rows,
16*4bdc9457SAndroid Build Coastguard Worker     size_t channels,
17*4bdc9457SAndroid Build Coastguard Worker     void* output,
18*4bdc9457SAndroid Build Coastguard Worker     size_t output_stride,
19*4bdc9457SAndroid Build Coastguard Worker     const uint32_t fill_pattern)
20*4bdc9457SAndroid Build Coastguard Worker {
21*4bdc9457SAndroid Build Coastguard Worker   assert(rows != 0);
22*4bdc9457SAndroid Build Coastguard Worker   assert(channels != 0);
23*4bdc9457SAndroid Build Coastguard Worker 
24*4bdc9457SAndroid Build Coastguard Worker   const size_t output_increment = output_stride - channels;
25*4bdc9457SAndroid Build Coastguard Worker 
26*4bdc9457SAndroid Build Coastguard Worker   const __m128i vfill = _mm_shuffle_epi32(_mm_cvtsi32_si128(fill_pattern), _MM_SHUFFLE(0, 0, 0, 0));
27*4bdc9457SAndroid Build Coastguard Worker   do {
28*4bdc9457SAndroid Build Coastguard Worker     size_t c = channels;
29*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 64 * sizeof(uint8_t); c -= 64 * sizeof(uint8_t)) {
30*4bdc9457SAndroid Build Coastguard Worker       _mm_storeu_si128((__m128i*) output, vfill);
31*4bdc9457SAndroid Build Coastguard Worker       _mm_storeu_si128((__m128i*) output + 1, vfill);
32*4bdc9457SAndroid Build Coastguard Worker       _mm_storeu_si128((__m128i*) output + 2, vfill);
33*4bdc9457SAndroid Build Coastguard Worker       _mm_storeu_si128((__m128i*) output + 3, vfill);
34*4bdc9457SAndroid Build Coastguard Worker       output = ((uint8_t*) output + 64);
35*4bdc9457SAndroid Build Coastguard Worker     }
36*4bdc9457SAndroid Build Coastguard Worker     for (; c >= 16 * sizeof(uint8_t); c -= 16 * sizeof(uint8_t)) {
37*4bdc9457SAndroid Build Coastguard Worker       _mm_storeu_si128((__m128i*) output, vfill);
38*4bdc9457SAndroid Build Coastguard Worker       output = ((uint8_t*) output + 16);
39*4bdc9457SAndroid Build Coastguard Worker     }
40*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNLIKELY(c != 0) {
41*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(c & (8 * sizeof(uint8_t))) {
42*4bdc9457SAndroid Build Coastguard Worker         _mm_storel_epi64(output, vfill);
43*4bdc9457SAndroid Build Coastguard Worker         output = ((uint8_t*) output + 8);
44*4bdc9457SAndroid Build Coastguard Worker       }
45*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(c & (4 * sizeof(uint8_t))) {
46*4bdc9457SAndroid Build Coastguard Worker         unaligned_store_u32(output, fill_pattern);
47*4bdc9457SAndroid Build Coastguard Worker         output = ((uint8_t*) output + 4);
48*4bdc9457SAndroid Build Coastguard Worker       }
49*4bdc9457SAndroid Build Coastguard Worker       uint32_t vfill_subpattern = fill_pattern;
50*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(c & (2 * sizeof(uint8_t))) {
51*4bdc9457SAndroid Build Coastguard Worker         unaligned_store_u16(output, (uint16_t) vfill_subpattern);
52*4bdc9457SAndroid Build Coastguard Worker         vfill_subpattern >>= 16;
53*4bdc9457SAndroid Build Coastguard Worker         output = ((uint8_t*) output + 2);
54*4bdc9457SAndroid Build Coastguard Worker       }
55*4bdc9457SAndroid Build Coastguard Worker       if XNN_LIKELY(c & (1 * sizeof(uint8_t))) {
56*4bdc9457SAndroid Build Coastguard Worker         *((uint8_t*) output) = (uint8_t) vfill_subpattern;
57*4bdc9457SAndroid Build Coastguard Worker         output = ((uint8_t*) output + 1);
58*4bdc9457SAndroid Build Coastguard Worker       }
59*4bdc9457SAndroid Build Coastguard Worker     }
60*4bdc9457SAndroid Build Coastguard Worker     output = (void*) ((uintptr_t) output + output_increment);
61*4bdc9457SAndroid Build Coastguard Worker   } while (--rows != 0);
62*4bdc9457SAndroid Build Coastguard Worker }
63