1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/sse.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 <smmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16
17
xnn_f32_qs8_vcvt_ukernel__sse41_x24(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__sse41_x24(
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)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m128 vscale = _mm_load_ps(params->sse4.scale);
30 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4.output_min);
33
34 for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
35 __m128 vx0123 = _mm_loadu_ps(x);
36 __m128 vx4567 = _mm_loadu_ps(x + 4);
37 __m128 vx89AB = _mm_loadu_ps(x + 8);
38 __m128 vxCDEF = _mm_loadu_ps(x + 12);
39 __m128 vxGHIJ = _mm_loadu_ps(x + 16);
40 __m128 vxKLMN = _mm_loadu_ps(x + 20);
41 x += 24;
42
43 vx0123 = _mm_mul_ps(vx0123, vscale);
44 vx4567 = _mm_mul_ps(vx4567, vscale);
45 vx89AB = _mm_mul_ps(vx89AB, vscale);
46 vxCDEF = _mm_mul_ps(vxCDEF, vscale);
47 vxGHIJ = _mm_mul_ps(vxGHIJ, vscale);
48 vxKLMN = _mm_mul_ps(vxKLMN, vscale);
49
50 vx0123 = _mm_min_ps(vx0123, voutput_max_less_zero_point);
51 vx4567 = _mm_min_ps(vx4567, voutput_max_less_zero_point);
52 vx89AB = _mm_min_ps(vx89AB, voutput_max_less_zero_point);
53 vxCDEF = _mm_min_ps(vxCDEF, voutput_max_less_zero_point);
54 vxGHIJ = _mm_min_ps(vxGHIJ, voutput_max_less_zero_point);
55 vxKLMN = _mm_min_ps(vxKLMN, voutput_max_less_zero_point);
56
57 const __m128i vy0123 = _mm_cvtps_epi32(vx0123);
58 const __m128i vy4567 = _mm_cvtps_epi32(vx4567);
59 const __m128i vy89AB = _mm_cvtps_epi32(vx89AB);
60 const __m128i vyCDEF = _mm_cvtps_epi32(vxCDEF);
61 const __m128i vyGHIJ = _mm_cvtps_epi32(vxGHIJ);
62 const __m128i vyKLMN = _mm_cvtps_epi32(vxKLMN);
63
64 __m128i vy01234567 = _mm_packs_epi32(vy0123, vy4567);
65 __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF);
66 __m128i vyGHIJKLMN = _mm_packs_epi32(vyGHIJ, vyKLMN);
67
68 vy01234567 = _mm_adds_epi16(vy01234567, voutput_zero_point);
69 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point);
70 vyGHIJKLMN = _mm_adds_epi16(vyGHIJKLMN, voutput_zero_point);
71
72
73 __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF);
74 vyGHIJKLMN = _mm_packs_epi16(vyGHIJKLMN, vyGHIJKLMN);
75
76 vy0123456789ABCDEF = _mm_max_epi8(vy0123456789ABCDEF, voutput_min);
77 vyGHIJKLMN = _mm_max_epi8(vyGHIJKLMN, voutput_min);
78
79 _mm_storeu_si128((__m128i*) y, vy0123456789ABCDEF);
80 _mm_storel_epi64((__m128i*) (y + 16), vyGHIJKLMN);
81 y += 24;
82 }
83 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
84 __m128 vx_lo = _mm_loadu_ps(x);
85 __m128 vx_hi = _mm_loadu_ps(x + 4);
86 x += 8;
87
88 vx_lo = _mm_mul_ps(vx_lo, vscale);
89 vx_hi = _mm_mul_ps(vx_hi, vscale);
90
91 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
92 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
93
94 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
95 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
96
97 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
98 vy = _mm_adds_epi16(vy, voutput_zero_point);
99 vy = _mm_packs_epi16(vy, vy);
100 vy = _mm_max_epi8(vy, voutput_min);
101
102 _mm_storel_epi64((__m128i*) y, vy);
103 y += 8;
104 }
105 if XNN_UNLIKELY(n != 0) {
106 __m128 vx_lo = _mm_loadu_ps(x);
107 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
108 __m128 vx_hi = _mm_loadu_ps(x_hi);
109
110 vx_lo = _mm_mul_ps(vx_lo, vscale);
111 vx_hi = _mm_mul_ps(vx_hi, vscale);
112
113 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
114 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
115
116 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
117 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
118
119 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
120 vy = _mm_adds_epi16(vy, voutput_zero_point);
121 vy = _mm_packs_epi16(vy, vy);
122 vy = _mm_max_epi8(vy, voutput_min);
123
124 if (n & (4 * sizeof(float))) {
125 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
126 y += 4;
127 vy = _mm_srli_epi64(vy, 32);
128 }
129 if (n & (2 * sizeof(float))) {
130 *((uint16_t*) y) = (uint16_t) _mm_extract_epi16(vy, 0);
131 y += 2;
132 vy = _mm_srli_epi32(vy, 16);
133 }
134 if (n & (1 * sizeof(float))) {
135 *y = (int8_t) _mm_extract_epi8(vy, 0);
136 }
137 }
138 }
139