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_x16(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_x16(
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 __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
33
34 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
35 __m256 vx01 = _mm256_loadu_ps(x);
36 __m256 vx23 = _mm256_loadu_ps(x + 8);
37 x += 16;
38
39 vx01 = _mm256_mul_ps(vx01, vscale);
40 vx23 = _mm256_mul_ps(vx23, vscale);
41
42 vx01 = _mm256_min_ps(vx01, voutput_max_less_zero_point);
43 vx23 = _mm256_min_ps(vx23, voutput_max_less_zero_point);
44
45 const __m256i vacc01 = _mm256_cvtps_epi32(vx01);
46 const __m256i vacc23 = _mm256_cvtps_epi32(vx23);
47
48 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23);
49
50 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point);
51
52 const __m128i vy0213 = _mm_packs_epi16(_mm256_castsi256_si128(vacc0213), _mm256_extracti128_si256(vacc0213, 1));
53
54 __m128i vy0123 = _mm_shuffle_epi32(vy0213, _MM_SHUFFLE(3, 1, 2, 0));
55
56 vy0123 = _mm_max_epi8(vy0123, voutput_min);
57
58 _mm_storeu_si128((__m128i*) y, vy0123);
59 y += 16;
60 }
61 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
62 __m256 vx = _mm256_loadu_ps(x);
63 vx = _mm256_mul_ps(vx, vscale);
64 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
65 x += 8;
66
67 const __m256i vacc = _mm256_cvtps_epi32(vx);
68
69 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
70 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
71 vy = _mm_packs_epi16(vy, vy);
72 vy = _mm_max_epi8(vy, voutput_min);
73
74 _mm_storel_epi64((__m128i*) y, vy);
75 y += 8;
76 }
77 if XNN_UNLIKELY(n != 0) {
78 assert(n >= 1 * sizeof(float));
79 assert(n <= 7 * sizeof(float));
80 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2.mask_table[7] - n));
81
82 __m256 vx = _mm256_maskload_ps(x, vmask);
83 vx = _mm256_mul_ps(vx, vscale);
84 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
85
86 const __m256i vacc = _mm256_cvtps_epi32(vx);
87
88 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
89 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
90 vy = _mm_packs_epi16(vy, vy);
91 vy = _mm_max_epi8(vy, voutput_min);
92
93 if (n & (4 * sizeof(float))) {
94 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
95 y += 4;
96 vy = _mm_srli_epi64(vy, 32);
97 }
98 if (n & (2 * sizeof(float))) {
99 *((uint16_t*) y) = (uint16_t) _mm_extract_epi16(vy, 0);
100 y += 2;
101 vy = _mm_srli_epi32(vy, 16);
102 }
103 if (n & (1 * sizeof(float))) {
104 *y = (int8_t) _mm_extract_epi8(vy, 0);
105 }
106 }
107 }
108