1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-f32-vcvt/avx.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/unaligned.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_qu8_f32_vcvt_ukernel__avx_x32(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__avx_x32(
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 __m128i vminus_zero_point = _mm_load_si128((const __m128i*) params->avx.minus_zero_point);
31 const __m256 vscale = _mm256_load_ps(params->avx.scale);
32 for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
33 __m128i vx0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
34 __m128i vx4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 4)));
35 __m128i vx89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 8)));
36 __m128i vxCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 12)));
37 __m128i vxGHIJ = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 16)));
38 __m128i vxKLMN = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 20)));
39 __m128i vxOPQR = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 24)));
40 __m128i vxSTUV = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 28)));
41 x += 32;
42
43 vx0123 = _mm_add_epi32(vx0123, vminus_zero_point);
44 vx4567 = _mm_add_epi32(vx4567, vminus_zero_point);
45 vx89AB = _mm_add_epi32(vx89AB, vminus_zero_point);
46 vxCDEF = _mm_add_epi32(vxCDEF, vminus_zero_point);
47 vxGHIJ = _mm_add_epi32(vxGHIJ, vminus_zero_point);
48 vxKLMN = _mm_add_epi32(vxKLMN, vminus_zero_point);
49 vxOPQR = _mm_add_epi32(vxOPQR, vminus_zero_point);
50 vxSTUV = _mm_add_epi32(vxSTUV, vminus_zero_point);
51
52 const __m256i vx01234567 = _mm256_insertf128_si256(_mm256_castsi128_si256(vx0123), vx4567, 1);
53 const __m256i vx89ABCDEF = _mm256_insertf128_si256(_mm256_castsi128_si256(vx89AB), vxCDEF, 1);
54 const __m256i vxGHIJKLMN = _mm256_insertf128_si256(_mm256_castsi128_si256(vxGHIJ), vxKLMN, 1);
55 const __m256i vxOPQRSTUV = _mm256_insertf128_si256(_mm256_castsi128_si256(vxOPQR), vxSTUV, 1);
56
57 __m256 vy01234567 = _mm256_cvtepi32_ps(vx01234567);
58 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF);
59 __m256 vyGHIJKLMN = _mm256_cvtepi32_ps(vxGHIJKLMN);
60 __m256 vyOPQRSTUV = _mm256_cvtepi32_ps(vxOPQRSTUV);
61
62 vy01234567 = _mm256_mul_ps(vy01234567, vscale);
63 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale);
64 vyGHIJKLMN = _mm256_mul_ps(vyGHIJKLMN, vscale);
65 vyOPQRSTUV = _mm256_mul_ps(vyOPQRSTUV, vscale);
66
67 _mm256_storeu_ps(y, vy01234567);
68 _mm256_storeu_ps(y + 8, vy89ABCDEF);
69 _mm256_storeu_ps(y + 16, vyGHIJKLMN);
70 _mm256_storeu_ps(y + 24, vyOPQRSTUV);
71 y += 32;
72 }
73 for (; n >= 4 * sizeof(uint8_t); n -= 4 * sizeof(uint8_t)) {
74 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
75 vx = _mm_add_epi32(vx, vminus_zero_point);
76 x += 4;
77
78 __m128 vy = _mm_cvtepi32_ps(vx);
79 vy = _mm_mul_ps(vy, _mm256_castps256_ps128(vscale));
80
81 _mm_storeu_ps(y, vy);
82 y += 4;
83 }
84 if XNN_UNLIKELY(n != 0) {
85 assert(n >= 1 * sizeof(uint8_t));
86 assert(n <= 3 * sizeof(uint8_t));
87
88 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
89 vx = _mm_add_epi32(vx, vminus_zero_point);
90
91 __m128 vy = _mm_cvtepi32_ps(vx);
92 vy = _mm_mul_ps(vy, _mm256_castps256_ps128(vscale));
93
94 if (n & (2 * sizeof(uint8_t))) {
95 _mm_storel_pi((__m64*) y, vy);
96 vy = _mm_movehl_ps(vy, vy);
97 y += 2;
98 }
99 if (n & (1 * sizeof(uint8_t))) {
100 _mm_store_ss(y, vy);
101 }
102 }
103 }
104