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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vcvt.h>
16
17
xnn_f32_qs8_vcvt_ukernel__sse2_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__sse2_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)]) 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->sse2.scale);
30 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
31 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
32 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
33
34 for (; n >= 32 * sizeof(float); n -= 32 * 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 __m128 vxOPQR = _mm_loadu_ps(x + 24);
42 __m128 vxSTUV = _mm_loadu_ps(x + 28);
43 x += 32;
44
45 vx0123 = _mm_mul_ps(vx0123, vscale);
46 vx4567 = _mm_mul_ps(vx4567, vscale);
47 vx89AB = _mm_mul_ps(vx89AB, vscale);
48 vxCDEF = _mm_mul_ps(vxCDEF, vscale);
49 vxGHIJ = _mm_mul_ps(vxGHIJ, vscale);
50 vxKLMN = _mm_mul_ps(vxKLMN, vscale);
51 vxOPQR = _mm_mul_ps(vxOPQR, vscale);
52 vxSTUV = _mm_mul_ps(vxSTUV, vscale);
53
54 vx0123 = _mm_min_ps(vx0123, voutput_max_less_zero_point);
55 vx4567 = _mm_min_ps(vx4567, voutput_max_less_zero_point);
56 vx89AB = _mm_min_ps(vx89AB, voutput_max_less_zero_point);
57 vxCDEF = _mm_min_ps(vxCDEF, voutput_max_less_zero_point);
58 vxGHIJ = _mm_min_ps(vxGHIJ, voutput_max_less_zero_point);
59 vxKLMN = _mm_min_ps(vxKLMN, voutput_max_less_zero_point);
60 vxOPQR = _mm_min_ps(vxOPQR, voutput_max_less_zero_point);
61 vxSTUV = _mm_min_ps(vxSTUV, voutput_max_less_zero_point);
62
63 const __m128i vy0123 = _mm_cvtps_epi32(vx0123);
64 const __m128i vy4567 = _mm_cvtps_epi32(vx4567);
65 const __m128i vy89AB = _mm_cvtps_epi32(vx89AB);
66 const __m128i vyCDEF = _mm_cvtps_epi32(vxCDEF);
67 const __m128i vyGHIJ = _mm_cvtps_epi32(vxGHIJ);
68 const __m128i vyKLMN = _mm_cvtps_epi32(vxKLMN);
69 const __m128i vyOPQR = _mm_cvtps_epi32(vxOPQR);
70 const __m128i vySTUV = _mm_cvtps_epi32(vxSTUV);
71
72 __m128i vy01234567 = _mm_packs_epi32(vy0123, vy4567);
73 __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF);
74 __m128i vyGHIJKLMN = _mm_packs_epi32(vyGHIJ, vyKLMN);
75 __m128i vyOPQRSTUV = _mm_packs_epi32(vyOPQR, vySTUV);
76
77 vy01234567 = _mm_adds_epi16(vy01234567, voutput_zero_point);
78 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point);
79 vyGHIJKLMN = _mm_adds_epi16(vyGHIJKLMN, voutput_zero_point);
80 vyOPQRSTUV = _mm_adds_epi16(vyOPQRSTUV, voutput_zero_point);
81
82 vy01234567 = _mm_max_epi16(vy01234567, voutput_min);
83 vy89ABCDEF = _mm_max_epi16(vy89ABCDEF, voutput_min);
84 vyGHIJKLMN = _mm_max_epi16(vyGHIJKLMN, voutput_min);
85 vyOPQRSTUV = _mm_max_epi16(vyOPQRSTUV, voutput_min);
86
87 __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF);
88 __m128i vyGHIJKLMNOPQRSTUV = _mm_packs_epi16(vyGHIJKLMN, vyOPQRSTUV);
89
90
91 _mm_storeu_si128((__m128i*) y, vy0123456789ABCDEF);
92 _mm_storeu_si128((__m128i*) (y + 16), vyGHIJKLMNOPQRSTUV);
93 y += 32;
94 }
95 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
96 __m128 vx_lo = _mm_loadu_ps(x);
97 __m128 vx_hi = _mm_loadu_ps(x + 4);
98 x += 8;
99
100 vx_lo = _mm_mul_ps(vx_lo, vscale);
101 vx_hi = _mm_mul_ps(vx_hi, vscale);
102
103 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
104 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
105
106 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
107 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
108
109 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
110 vy = _mm_adds_epi16(vy, voutput_zero_point);
111 vy = _mm_max_epi16(vy, voutput_min);
112 vy = _mm_packs_epi16(vy, vy);
113
114 _mm_storel_epi64((__m128i*) y, vy);
115 y += 8;
116 }
117 if XNN_UNLIKELY(n != 0) {
118 __m128 vx_lo = _mm_loadu_ps(x);
119 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
120 __m128 vx_hi = _mm_loadu_ps(x_hi);
121
122 vx_lo = _mm_mul_ps(vx_lo, vscale);
123 vx_hi = _mm_mul_ps(vx_hi, vscale);
124
125 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
126 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
127
128 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
129 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
130
131 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
132 vy = _mm_adds_epi16(vy, voutput_zero_point);
133 vy = _mm_max_epi16(vy, voutput_min);
134 vy = _mm_packs_epi16(vy, vy);
135
136 if (n & (4 * sizeof(float))) {
137 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
138 y += 4;
139 vy = _mm_srli_epi64(vy, 32);
140 }
141 {
142 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
143 if (n & (2 * sizeof(float))) {
144 *((uint16_t*) y) = (uint16_t) vy_lo;
145 y += 2;
146 vy_lo >>= 16;
147 }
148 if (n & (1 * sizeof(float))) {
149 *y = (int8_t) vy_lo;
150 }
151 }
152 }
153 }
154