1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-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_f32_qu8_vcvt_ukernel__avx2_x32(size_t n,const float * x,uint8_t * y,const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_qu8_vcvt_ukernel__avx2_x32(
20 size_t n,
21 const float* x,
22 uint8_t* y,
23 const union xnn_f32_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m256 vscale = _mm256_load_ps(params->avx2.scale);
31 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx2.output_max_less_zero_point);
32 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
33 const __m256i vshuffle_mask = _mm256_load_si256((const __m256i*) params->avx2.shuffle_mask);
34 const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->avx2.output_min);
35
36 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
37 __m256 vx01 = _mm256_loadu_ps(x);
38 __m256 vx23 = _mm256_loadu_ps(x + 8);
39 __m256 vx45 = _mm256_loadu_ps(x + 16);
40 __m256 vx67 = _mm256_loadu_ps(x + 24);
41 x += 32;
42
43 vx01 = _mm256_mul_ps(vx01, vscale);
44 vx23 = _mm256_mul_ps(vx23, vscale);
45 vx45 = _mm256_mul_ps(vx45, vscale);
46 vx67 = _mm256_mul_ps(vx67, vscale);
47
48 vx01 = _mm256_min_ps(vx01, voutput_max_less_zero_point);
49 vx23 = _mm256_min_ps(vx23, voutput_max_less_zero_point);
50 vx45 = _mm256_min_ps(vx45, voutput_max_less_zero_point);
51 vx67 = _mm256_min_ps(vx67, voutput_max_less_zero_point);
52
53 const __m256i vacc01 = _mm256_cvtps_epi32(vx01);
54 const __m256i vacc23 = _mm256_cvtps_epi32(vx23);
55 const __m256i vacc45 = _mm256_cvtps_epi32(vx45);
56 const __m256i vacc67 = _mm256_cvtps_epi32(vx67);
57
58 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23);
59 __m256i vacc4657 = _mm256_packs_epi32(vacc45, vacc67);
60
61 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point);
62 vacc4657 = _mm256_adds_epi16(vacc4657, voutput_zero_point);
63
64 const __m256i vy02461357 = _mm256_packus_epi16(vacc0213, vacc4657);
65
66 __m256i vy01234567 = _mm256_permutevar8x32_epi32(vy02461357, vshuffle_mask);
67
68 vy01234567 = _mm256_max_epu8(vy01234567, voutput_min);
69
70 _mm256_storeu_si256((__m256i*) y, vy01234567);
71 y += 32;
72 }
73 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
74 __m256 vx = _mm256_loadu_ps(x);
75 vx = _mm256_mul_ps(vx, vscale);
76 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
77 x += 8;
78
79 const __m256i vacc = _mm256_cvtps_epi32(vx);
80
81 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
82 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
83 vy = _mm_packus_epi16(vy, vy);
84 vy = _mm_max_epu8(vy, _mm256_castsi256_si128(voutput_min));
85
86 _mm_storel_epi64((__m128i*) y, vy);
87 y += 8;
88 }
89 if XNN_UNLIKELY(n != 0) {
90 assert(n >= 1 * sizeof(float));
91 assert(n <= 7 * sizeof(float));
92 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2.mask_table[7] - n));
93
94 __m256 vx = _mm256_maskload_ps(x, vmask);
95 vx = _mm256_mul_ps(vx, vscale);
96 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
97
98 const __m256i vacc = _mm256_cvtps_epi32(vx);
99
100 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
101 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
102 vy = _mm_packus_epi16(vy, vy);
103 vy = _mm_max_epu8(vy, _mm256_castsi256_si128(voutput_min));
104
105 if (n & (4 * sizeof(float))) {
106 _mm_storeu_si32(y, vy);
107 y += 4;
108 vy = _mm_srli_epi64(vy, 32);
109 }
110 if (n & (2 * sizeof(float))) {
111 _mm_storeu_si16(y, vy);
112 y += 2;
113 vy = _mm_srli_epi32(vy, 16);
114 }
115 if (n & (1 * sizeof(float))) {
116 *y = (uint8_t) _mm_extract_epi8(vy, 0);
117 }
118 }
119 }
120