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_x24(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_x24(
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 >= 24 * sizeof(uint8_t); n -= 24 * 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 x += 24;
40
41 vx0123 = _mm_add_epi32(vx0123, vminus_zero_point);
42 vx4567 = _mm_add_epi32(vx4567, vminus_zero_point);
43 vx89AB = _mm_add_epi32(vx89AB, vminus_zero_point);
44 vxCDEF = _mm_add_epi32(vxCDEF, vminus_zero_point);
45 vxGHIJ = _mm_add_epi32(vxGHIJ, vminus_zero_point);
46 vxKLMN = _mm_add_epi32(vxKLMN, vminus_zero_point);
47
48 const __m256i vx01234567 = _mm256_insertf128_si256(_mm256_castsi128_si256(vx0123), vx4567, 1);
49 const __m256i vx89ABCDEF = _mm256_insertf128_si256(_mm256_castsi128_si256(vx89AB), vxCDEF, 1);
50 const __m256i vxGHIJKLMN = _mm256_insertf128_si256(_mm256_castsi128_si256(vxGHIJ), vxKLMN, 1);
51
52 __m256 vy01234567 = _mm256_cvtepi32_ps(vx01234567);
53 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF);
54 __m256 vyGHIJKLMN = _mm256_cvtepi32_ps(vxGHIJKLMN);
55
56 vy01234567 = _mm256_mul_ps(vy01234567, vscale);
57 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale);
58 vyGHIJKLMN = _mm256_mul_ps(vyGHIJKLMN, vscale);
59
60 _mm256_storeu_ps(y, vy01234567);
61 _mm256_storeu_ps(y + 8, vy89ABCDEF);
62 _mm256_storeu_ps(y + 16, vyGHIJKLMN);
63 y += 24;
64 }
65 for (; n >= 4 * sizeof(uint8_t); n -= 4 * sizeof(uint8_t)) {
66 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
67 vx = _mm_add_epi32(vx, vminus_zero_point);
68 x += 4;
69
70 __m128 vy = _mm_cvtepi32_ps(vx);
71 vy = _mm_mul_ps(vy, _mm256_castps256_ps128(vscale));
72
73 _mm_storeu_ps(y, vy);
74 y += 4;
75 }
76 if XNN_UNLIKELY(n != 0) {
77 assert(n >= 1 * sizeof(uint8_t));
78 assert(n <= 3 * sizeof(uint8_t));
79
80 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
81 vx = _mm_add_epi32(vx, vminus_zero_point);
82
83 __m128 vy = _mm_cvtepi32_ps(vx);
84 vy = _mm_mul_ps(vy, _mm256_castps256_ps128(vscale));
85
86 if (n & (2 * sizeof(uint8_t))) {
87 _mm_storel_pi((__m64*) y, vy);
88 vy = _mm_movehl_ps(vy, vy);
89 y += 2;
90 }
91 if (n & (1 * sizeof(uint8_t))) {
92 _mm_store_ss(y, vy);
93 }
94 }
95 }
96