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_x64(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_x64(
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 >= 64 * sizeof(float); n -= 64 * 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 __m256 vx89 = _mm256_loadu_ps(x + 32);
41 __m256 vxAB = _mm256_loadu_ps(x + 40);
42 __m256 vxCD = _mm256_loadu_ps(x + 48);
43 __m256 vxEF = _mm256_loadu_ps(x + 56);
44 x += 64;
45
46 vx01 = _mm256_mul_ps(vx01, vscale);
47 vx23 = _mm256_mul_ps(vx23, vscale);
48 vx45 = _mm256_mul_ps(vx45, vscale);
49 vx67 = _mm256_mul_ps(vx67, vscale);
50 vx89 = _mm256_mul_ps(vx89, vscale);
51 vxAB = _mm256_mul_ps(vxAB, vscale);
52 vxCD = _mm256_mul_ps(vxCD, vscale);
53 vxEF = _mm256_mul_ps(vxEF, vscale);
54
55 vx01 = _mm256_min_ps(vx01, voutput_max_less_zero_point);
56 vx23 = _mm256_min_ps(vx23, voutput_max_less_zero_point);
57 vx45 = _mm256_min_ps(vx45, voutput_max_less_zero_point);
58 vx67 = _mm256_min_ps(vx67, voutput_max_less_zero_point);
59 vx89 = _mm256_min_ps(vx89, voutput_max_less_zero_point);
60 vxAB = _mm256_min_ps(vxAB, voutput_max_less_zero_point);
61 vxCD = _mm256_min_ps(vxCD, voutput_max_less_zero_point);
62 vxEF = _mm256_min_ps(vxEF, voutput_max_less_zero_point);
63
64 const __m256i vacc01 = _mm256_cvtps_epi32(vx01);
65 const __m256i vacc23 = _mm256_cvtps_epi32(vx23);
66 const __m256i vacc45 = _mm256_cvtps_epi32(vx45);
67 const __m256i vacc67 = _mm256_cvtps_epi32(vx67);
68 const __m256i vacc89 = _mm256_cvtps_epi32(vx89);
69 const __m256i vaccAB = _mm256_cvtps_epi32(vxAB);
70 const __m256i vaccCD = _mm256_cvtps_epi32(vxCD);
71 const __m256i vaccEF = _mm256_cvtps_epi32(vxEF);
72
73 __m256i vacc0213 = _mm256_packs_epi32(vacc01, vacc23);
74 __m256i vacc4657 = _mm256_packs_epi32(vacc45, vacc67);
75 __m256i vacc8A9B = _mm256_packs_epi32(vacc89, vaccAB);
76 __m256i vaccCEDF = _mm256_packs_epi32(vaccCD, vaccEF);
77
78 vacc0213 = _mm256_adds_epi16(vacc0213, voutput_zero_point);
79 vacc4657 = _mm256_adds_epi16(vacc4657, voutput_zero_point);
80 vacc8A9B = _mm256_adds_epi16(vacc8A9B, voutput_zero_point);
81 vaccCEDF = _mm256_adds_epi16(vaccCEDF, voutput_zero_point);
82
83 const __m256i vy02461357 = _mm256_packs_epi16(vacc0213, vacc4657);
84 const __m256i vy8ACE9BDF = _mm256_packs_epi16(vacc8A9B, vaccCEDF);
85
86 __m256i vy01234567 = _mm256_permutevar8x32_epi32(vy02461357, vshuffle_mask);
87 __m256i vy89ABCDEF = _mm256_permutevar8x32_epi32(vy8ACE9BDF, vshuffle_mask);
88
89 vy01234567 = _mm256_max_epi8(vy01234567, voutput_min);
90 vy89ABCDEF = _mm256_max_epi8(vy89ABCDEF, voutput_min);
91
92 _mm256_storeu_si256((__m256i*) y, vy01234567);
93 _mm256_storeu_si256((__m256i*) (y + 32), vy89ABCDEF);
94 y += 64;
95 }
96 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
97 __m256 vx = _mm256_loadu_ps(x);
98 vx = _mm256_mul_ps(vx, vscale);
99 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
100 x += 8;
101
102 const __m256i vacc = _mm256_cvtps_epi32(vx);
103
104 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
105 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
106 vy = _mm_packs_epi16(vy, vy);
107 vy = _mm_max_epi8(vy, _mm256_castsi256_si128(voutput_min));
108
109 _mm_storel_epi64((__m128i*) y, vy);
110 y += 8;
111 }
112 if XNN_UNLIKELY(n != 0) {
113 assert(n >= 1 * sizeof(float));
114 assert(n <= 7 * sizeof(float));
115 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2.mask_table[7] - n));
116
117 __m256 vx = _mm256_maskload_ps(x, vmask);
118 vx = _mm256_mul_ps(vx, vscale);
119 vx = _mm256_min_ps(vx, voutput_max_less_zero_point);
120
121 const __m256i vacc = _mm256_cvtps_epi32(vx);
122
123 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1));
124 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point));
125 vy = _mm_packs_epi16(vy, vy);
126 vy = _mm_max_epi8(vy, _mm256_castsi256_si128(voutput_min));
127
128 if (n & (4 * sizeof(float))) {
129 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
130 y += 4;
131 vy = _mm_srli_epi64(vy, 32);
132 }
133 if (n & (2 * sizeof(float))) {
134 *((uint16_t*) y) = (uint16_t) _mm_extract_epi16(vy, 0);
135 y += 2;
136 vy = _mm_srli_epi32(vy, 16);
137 }
138 if (n & (1 * sizeof(float))) {
139 *y = (int8_t) _mm_extract_epi8(vy, 0);
140 }
141 }
142 }
143