Debian Codegraph @

src/qu8-f32-vcvt/gen/qu8-f32-vcvt-avx-u32.c

raw
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
19 void xnn_qu8_f32_vcvt_ukernel__avx_u32(
20 size_t batch,
21 const uint8_t* input,
22 float* output,
23 const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const __m128i vzero_point = _mm_set1_epi32(params->scalar.zero_point);
31 const __m256 vscale = _mm256_set1_ps(params->scalar.scale);
32 XNN_FORCE_REALIZATION(vzero_point);
33 XNN_FORCE_REALIZATION(vscale);
34 for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) {
35 __m128i vx0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
36 __m128i vx4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 4)));
37 __m128i vx89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 8)));
38 __m128i vxCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 12)));
39 __m128i vxGHIJ = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 16)));
40 __m128i vxKLMN = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 20)));
41 __m128i vxOPQR = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 24)));
42 __m128i vxSTUV = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 28)));
43 input += 32;
44
45 vx0123 = _mm_sub_epi32(vx0123, vzero_point);
46 vx4567 = _mm_sub_epi32(vx4567, vzero_point);
47 vx89AB = _mm_sub_epi32(vx89AB, vzero_point);
48 vxCDEF = _mm_sub_epi32(vxCDEF, vzero_point);
49 vxGHIJ = _mm_sub_epi32(vxGHIJ, vzero_point);
50 vxKLMN = _mm_sub_epi32(vxKLMN, vzero_point);
51 vxOPQR = _mm_sub_epi32(vxOPQR, vzero_point);
52 vxSTUV = _mm_sub_epi32(vxSTUV, vzero_point);
53
54 const __m256i vx01234567 = _mm256_insertf128_si256(_mm256_castsi128_si256(vx0123), vx4567, 1);
55 const __m256i vx89ABCDEF = _mm256_insertf128_si256(_mm256_castsi128_si256(vx89AB), vxCDEF, 1);
56 const __m256i vxGHIJKLMN = _mm256_insertf128_si256(_mm256_castsi128_si256(vxGHIJ), vxKLMN, 1);
57 const __m256i vxOPQRSTUV = _mm256_insertf128_si256(_mm256_castsi128_si256(vxOPQR), vxSTUV, 1);
58
59 __m256 vy01234567 = _mm256_cvtepi32_ps(vx01234567);
60 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF);
61 __m256 vyGHIJKLMN = _mm256_cvtepi32_ps(vxGHIJKLMN);
62 __m256 vyOPQRSTUV = _mm256_cvtepi32_ps(vxOPQRSTUV);
63
64 vy01234567 = _mm256_mul_ps(vy01234567, vscale);
65 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale);
66 vyGHIJKLMN = _mm256_mul_ps(vyGHIJKLMN, vscale);
67 vyOPQRSTUV = _mm256_mul_ps(vyOPQRSTUV, vscale);
68
69 _mm256_storeu_ps(output, vy01234567);
70 _mm256_storeu_ps(output + 8, vy89ABCDEF);
71 _mm256_storeu_ps(output + 16, vyGHIJKLMN);
72 _mm256_storeu_ps(output + 24, vyOPQRSTUV);
73 output += 32;
74 }
75 for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) {
76 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
77 vx = _mm_sub_epi32(vx, vzero_point);
78 input += 4;
79
80 __m128 vy = _mm_cvtepi32_ps(vx);
81 vy = _mm_mul_ps(vy, _mm256_castps256_ps128(vscale));
82
83 _mm_storeu_ps(output, vy);
84 output += 4;
85 }
86 if XNN_UNLIKELY(batch != 0) {
87 assert(batch >= 1 * sizeof(uint8_t));
88 assert(batch <= 3 * sizeof(uint8_t));
89
90 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
91 vx = _mm_sub_epi32(vx, vzero_point);
92
93 __m128 vy = _mm_cvtepi32_ps(vx);
94 vy = _mm_mul_ps(vy, _mm256_castps256_ps128(vscale));
95
96 if (batch & (2 * sizeof(uint8_t))) {
97 _mm_storel_pi((__m64*) output, vy);
98 vy = _mm_movehl_ps(vy, vy);
99 output += 2;
100 }
101 if (batch & (1 * sizeof(uint8_t))) {
102 _mm_store_ss(output, vy);
103 }
104 }
105 }