Debian Codegraph @

src/qu8-f32-vcvt/gen/qu8-f32-vcvt-avx-u16.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_u16(
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 >= 16 * sizeof(uint8_t); batch -= 16 * 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 input += 16;
40
41 vx0123 = _mm_sub_epi32(vx0123, vzero_point);
42 vx4567 = _mm_sub_epi32(vx4567, vzero_point);
43 vx89AB = _mm_sub_epi32(vx89AB, vzero_point);
44 vxCDEF = _mm_sub_epi32(vxCDEF, vzero_point);
45
46 const __m256i vx01234567 = _mm256_insertf128_si256(_mm256_castsi128_si256(vx0123), vx4567, 1);
47 const __m256i vx89ABCDEF = _mm256_insertf128_si256(_mm256_castsi128_si256(vx89AB), vxCDEF, 1);
48
49 __m256 vy01234567 = _mm256_cvtepi32_ps(vx01234567);
50 __m256 vy89ABCDEF = _mm256_cvtepi32_ps(vx89ABCDEF);
51
52 vy01234567 = _mm256_mul_ps(vy01234567, vscale);
53 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vscale);
54
55 _mm256_storeu_ps(output, vy01234567);
56 _mm256_storeu_ps(output + 8, vy89ABCDEF);
57 output += 16;
58 }
59 for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) {
60 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
61 vx = _mm_sub_epi32(vx, vzero_point);
62 input += 4;
63
64 __m128 vy = _mm_cvtepi32_ps(vx);
65 vy = _mm_mul_ps(vy, _mm256_castps256_ps128(vscale));
66
67 _mm_storeu_ps(output, vy);
68 output += 4;
69 }
70 if XNN_UNLIKELY(batch != 0) {
71 assert(batch >= 1 * sizeof(uint8_t));
72 assert(batch <= 3 * sizeof(uint8_t));
73
74 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
75 vx = _mm_sub_epi32(vx, vzero_point);
76
77 __m128 vy = _mm_cvtepi32_ps(vx);
78 vy = _mm_mul_ps(vy, _mm256_castps256_ps128(vscale));
79
80 if (batch & (2 * sizeof(uint8_t))) {
81 _mm_storel_pi((__m64*) output, vy);
82 vy = _mm_movehl_ps(vy, vy);
83 output += 2;
84 }
85 if (batch & (1 * sizeof(uint8_t))) {
86 _mm_store_ss(output, vy);
87 }
88 }
89 }