1 // Copyright 2019 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 <immintrin.h>
9
10 #include <xnnpack/common.h>
11 #include <xnnpack/vscale.h>
12
13
xnn_f32_vscale_ukernel__avx_x32(size_t n,const float * x,float * y,float c)14 void xnn_f32_vscale_ukernel__avx_x32(
15 size_t n,
16 const float* x,
17 float* y,
18 float c)
19 {
20 assert(n != 0);
21 assert(n % sizeof(float) == 0);
22
23 __m256 vc = _mm256_set1_ps(c);
24 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
25 const __m256 vx0 = _mm256_loadu_ps(x);
26 const __m256 vx1 = _mm256_loadu_ps(x + 8);
27 const __m256 vx2 = _mm256_loadu_ps(x + 16);
28 const __m256 vx3 = _mm256_loadu_ps(x + 24);
29 x += 32;
30
31 const __m256 vy0 = _mm256_mul_ps(vx0, vc);
32 const __m256 vy1 = _mm256_mul_ps(vx1, vc);
33 const __m256 vy2 = _mm256_mul_ps(vx2, vc);
34 const __m256 vy3 = _mm256_mul_ps(vx3, vc);
35
36 _mm256_storeu_ps(y, vy0);
37 _mm256_storeu_ps(y + 8, vy1);
38 _mm256_storeu_ps(y + 16, vy2);
39 _mm256_storeu_ps(y + 24, vy3);
40 y += 32;
41 }
42 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
43 const __m256 vx = _mm256_loadu_ps(x);
44 x += 8;
45
46 const __m256 vy = _mm256_mul_ps(vx, vc);
47
48 _mm256_storeu_ps(y, vy);
49 y += 8;
50 }
51 if XNN_UNLIKELY(n != 0) {
52 do {
53 const __m128 vx = _mm_load_ss(x);
54 x += 1;
55
56 const __m128 vy = _mm_mul_ss(vx, _mm256_castps256_ps128(vc));
57
58 _mm_store_ss(y, vy);
59 y += 1;
60
61 n -= sizeof(float);
62 } while (n != 0);
63 }
64 }
65