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/unaligned.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_f32_qs8_vcvt_ukernel__sse2_x24(size_t n,const float * x,int8_t * y,const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_qs8_vcvt_ukernel__sse2_x24(
20 size_t n,
21 const float* x,
22 int8_t* y,
23 const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m128 vscale = _mm_load_ps(params->sse2.scale);
31 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
32 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
33 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
34
35 for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
36 __m128 vx0123 = _mm_loadu_ps(x);
37 __m128 vx4567 = _mm_loadu_ps(x + 4);
38 __m128 vx89AB = _mm_loadu_ps(x + 8);
39 __m128 vxCDEF = _mm_loadu_ps(x + 12);
40 __m128 vxGHIJ = _mm_loadu_ps(x + 16);
41 __m128 vxKLMN = _mm_loadu_ps(x + 20);
42 x += 24;
43
44 vx0123 = _mm_mul_ps(vx0123, vscale);
45 vx4567 = _mm_mul_ps(vx4567, vscale);
46 vx89AB = _mm_mul_ps(vx89AB, vscale);
47 vxCDEF = _mm_mul_ps(vxCDEF, vscale);
48 vxGHIJ = _mm_mul_ps(vxGHIJ, vscale);
49 vxKLMN = _mm_mul_ps(vxKLMN, vscale);
50
51 vx0123 = _mm_min_ps(vx0123, voutput_max_less_zero_point);
52 vx4567 = _mm_min_ps(vx4567, voutput_max_less_zero_point);
53 vx89AB = _mm_min_ps(vx89AB, voutput_max_less_zero_point);
54 vxCDEF = _mm_min_ps(vxCDEF, voutput_max_less_zero_point);
55 vxGHIJ = _mm_min_ps(vxGHIJ, voutput_max_less_zero_point);
56 vxKLMN = _mm_min_ps(vxKLMN, voutput_max_less_zero_point);
57
58 const __m128i vy0123 = _mm_cvtps_epi32(vx0123);
59 const __m128i vy4567 = _mm_cvtps_epi32(vx4567);
60 const __m128i vy89AB = _mm_cvtps_epi32(vx89AB);
61 const __m128i vyCDEF = _mm_cvtps_epi32(vxCDEF);
62 const __m128i vyGHIJ = _mm_cvtps_epi32(vxGHIJ);
63 const __m128i vyKLMN = _mm_cvtps_epi32(vxKLMN);
64
65 __m128i vy01234567 = _mm_packs_epi32(vy0123, vy4567);
66 __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF);
67 __m128i vyGHIJKLMN = _mm_packs_epi32(vyGHIJ, vyKLMN);
68
69 vy01234567 = _mm_adds_epi16(vy01234567, voutput_zero_point);
70 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point);
71 vyGHIJKLMN = _mm_adds_epi16(vyGHIJKLMN, voutput_zero_point);
72
73 vy01234567 = _mm_max_epi16(vy01234567, voutput_min);
74 vy89ABCDEF = _mm_max_epi16(vy89ABCDEF, voutput_min);
75 vyGHIJKLMN = _mm_max_epi16(vyGHIJKLMN, voutput_min);
76
77 __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF);
78 vyGHIJKLMN = _mm_packs_epi16(vyGHIJKLMN, vyGHIJKLMN);
79
80
81 _mm_storeu_si128((__m128i*) y, vy0123456789ABCDEF);
82 _mm_storel_epi64((__m128i*) (y + 16), vyGHIJKLMN);
83 y += 24;
84 }
85 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
86 __m128 vx_lo = _mm_loadu_ps(x);
87 __m128 vx_hi = _mm_loadu_ps(x + 4);
88 x += 8;
89
90 vx_lo = _mm_mul_ps(vx_lo, vscale);
91 vx_hi = _mm_mul_ps(vx_hi, vscale);
92
93 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
94 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
95
96 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
97 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
98
99 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
100 vy = _mm_adds_epi16(vy, voutput_zero_point);
101 vy = _mm_max_epi16(vy, voutput_min);
102 vy = _mm_packs_epi16(vy, vy);
103
104 _mm_storel_epi64((__m128i*) y, vy);
105 y += 8;
106 }
107 if XNN_UNLIKELY(n != 0) {
108 __m128 vx_lo = _mm_loadu_ps(x);
109 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
110 __m128 vx_hi = _mm_loadu_ps(x_hi);
111
112 vx_lo = _mm_mul_ps(vx_lo, vscale);
113 vx_hi = _mm_mul_ps(vx_hi, vscale);
114
115 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
116 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
117
118 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
119 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
120
121 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
122 vy = _mm_adds_epi16(vy, voutput_zero_point);
123 vy = _mm_max_epi16(vy, voutput_min);
124 vy = _mm_packs_epi16(vy, vy);
125
126 if (n & (4 * sizeof(float))) {
127 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
128 y += 4;
129 vy = _mm_srli_epi64(vy, 32);
130 }
131 {
132 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
133 if (n & (2 * sizeof(float))) {
134 unaligned_store_u16(y, (uint16_t) vy_lo);
135 y += 2;
136 vy_lo >>= 16;
137 }
138 if (n & (1 * sizeof(float))) {
139 *y = (int8_t) vy_lo;
140 }
141 }
142 }
143 }
144