1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-f32-vcvt/avx2.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_qu8_f32_vcvt_ukernel__avx2_x16(size_t n,const uint8_t * x,float * y,const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_f32_vcvt_ukernel__avx2_x16(
20 size_t n,
21 const uint8_t* x,
22 float* y,
23 const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(uint8_t) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m256i vminus_zero_point = _mm256_load_si256((const __m256i*) params->avx.minus_zero_point);
31 const __m256 vscale = _mm256_load_ps(params->avx.scale);
32 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
33 __m256i vx01234567 = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) x));
34 __m256i vx89ABCDEF = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (x + 8)));
35 x += 16;
36
37 vx01234567 = _mm256_add_epi32(vx01234567, vminus_zero_point);
38 vx89ABCDEF = _mm256_add_epi32(vx89ABCDEF, vminus_zero_point);
39
40 __m256 vy01234567 = _mm256_cvtepi32_ps(vx01234567);
41 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF);
42
43 vy01234567 = _mm256_mul_ps(vy01234567, vscale);
44 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale);
45
46 _mm256_storeu_ps(y, vy01234567);
47 _mm256_storeu_ps(y + 8, vy89ABCDEF);
48 y += 16;
49 }
50 for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
51 __m256i vx = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) x));
52 vx = _mm256_add_epi32(vx, vminus_zero_point);
53 x += 8;
54
55 __m256 vy = _mm256_cvtepi32_ps(vx);
56 vy = _mm256_mul_ps(vy, vscale);
57
58 _mm256_storeu_ps(y, vy);
59 y += 8;
60 }
61 if XNN_UNLIKELY(n != 0) {
62 assert(n >= 1 * sizeof(uint8_t));
63 assert(n <= 7 * sizeof(uint8_t));
64
65 __m256i vx = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) x));
66 vx = _mm256_add_epi32(vx, vminus_zero_point);
67
68 __m256 vy = _mm256_cvtepi32_ps(vx);
69 vy = _mm256_mul_ps(vy, vscale);
70
71 __m128 vy_lo = _mm256_castps256_ps128(vy);
72 if (n & (4 * sizeof(uint8_t))) {
73 _mm_storeu_ps(y, vy_lo);
74 vy_lo = _mm256_extractf128_ps(vy, 1);
75 y += 4;
76 }
77 if (n & (2 * sizeof(uint8_t))) {
78 _mm_storel_pi((__m64*) y, vy_lo);
79 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
80 y += 2;
81 }
82 if (n & (1 * sizeof(uint8_t))) {
83 _mm_store_ss(y, vy_lo);
84 }
85 }
86 }
87