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/vcvt.h>
16
17
xnn_f32_qs8_vcvt_ukernel__avx2_x32(size_t n,const float * x,int8_t * y,const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_qs8_vcvt_ukernel__avx2_x32(
19 size_t n,
20 const float* x,
21 int8_t* y,
22 const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m256 vscale = _mm256_load_ps(params->avx2.scale);
30 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx2.output_max_less_zero_point);
31 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
32 const __m256i vshuffle_mask = _mm256_load_si256((const __m256i*) params->avx2.shuffle_mask);
33 const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->avx2.output_min);
34
35 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
36 __m256 vx01 = _mm256_loadu_ps(x);
37 __m256 vx23 = _mm256_loadu_ps(x + 8);
38 __m256 vx45 = _mm256_loadu_ps(x + 16);
39 __m256 vx67 = _mm256_loadu_ps(x + 24);
40 x += 32;
41
42 vx01 = _mm256_mul_ps(vx01, vscale);
43 vx23 = _mm256_mul_ps(vx23, vscale);
44 vx45 = _mm256_mul_ps(vx45, vscale);
45 vx67 = _mm256_mul_ps(vx67, vscale);
46
47 vx01 = _mm256_min_ps(vx01, voutput_max_less_zero_point);
48 vx23 = _mm256_min_ps(vx23, voutput_max_less_zero_point);
49 vx45 = _mm256_min_ps(vx45, voutput_max_less_zero_point);
50 vx67 = _mm256_min_ps(vx67, voutput_max_less_zero_point);
51
52 const __m256i vacc01 = _mm256_cvtps_epi32(vx01);
53 const __m256i vacc23 = _mm256_cvtps_epi32(vx23);
54 const __m256i vacc45 = _mm256_cvtps_epi32(vx45);
55 const __m256i vacc67 = _mm256_cvtps_epi32(vx67);
56
57 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23);
58 __m256i vacc4657 = _mm256_packs_epi32(vacc45, vacc67);
59
60 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point);
61 vacc4657 = _mm256_adds_epi16(vacc4657, voutput_zero_point);
62
63 const __m256i vy02461357 = _mm256_packs_epi16(vacc0213, vacc4657);
64
65 __m256i vy01234567 = _mm256_permutevar8x32_epi32(vy02461357, vshuffle_mask);
66
67 vy01234567 = _mm256_max_epi8(vy01234567, voutput_min);
68
69 _mm256_storeu_si256((__m256i*) y, vy01234567);
70 y += 32;
71 }
72 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
73 __m256 vx = _mm256_loadu_ps(x);
74 vx = _mm256_mul_ps(vx, vscale);
75 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
76 x += 8;
77
78 const __m256i vacc = _mm256_cvtps_epi32(vx);
79
80 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
81 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
82 vy = _mm_packs_epi16(vy, vy);
83 vy = _mm_max_epi8(vy, _mm256_castsi256_si128(voutput_min));
84
85 _mm_storel_epi64((__m128i*) y, vy);
86 y += 8;
87 }
88 if XNN_UNLIKELY(n != 0) {
89 assert(n >= 1 * sizeof(float));
90 assert(n <= 7 * sizeof(float));
91 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2.mask_table[7] - n));
92
93 __m256 vx = _mm256_maskload_ps(x, vmask);
94 vx = _mm256_mul_ps(vx, vscale);
95 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
96
97 const __m256i vacc = _mm256_cvtps_epi32(vx);
98
99 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
100 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
101 vy = _mm_packs_epi16(vy, vy);
102 vy = _mm_max_epi8(vy, _mm256_castsi256_si128(voutput_min));
103
104 if (n & (4 * sizeof(float))) {
105 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
106 y += 4;
107 vy = _mm_srli_epi64(vy, 32);
108 }
109 if (n & (2 * sizeof(float))) {
110 *((uint16_t*) y) = (uint16_t) _mm_extract_epi16(vy, 0);
111 y += 2;
112 vy = _mm_srli_epi32(vy, 16);
113 }
114 if (n & (1 * sizeof(float))) {
115 *y = (int8_t) _mm_extract_epi8(vy, 0);
116 }
117 }
118 }
119