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_x16(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_x16(
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 >= 16 * sizeof(float); n -= 16 * 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 x += 16;
41
42 vx0123 = _mm_mul_ps(vx0123, vscale);
43 vx4567 = _mm_mul_ps(vx4567, vscale);
44 vx89AB = _mm_mul_ps(vx89AB, vscale);
45 vxCDEF = _mm_mul_ps(vxCDEF, vscale);
46
47 vx0123 = _mm_min_ps(vx0123, voutput_max_less_zero_point);
48 vx4567 = _mm_min_ps(vx4567, voutput_max_less_zero_point);
49 vx89AB = _mm_min_ps(vx89AB, voutput_max_less_zero_point);
50 vxCDEF = _mm_min_ps(vxCDEF, voutput_max_less_zero_point);
51
52 const __m128i vy0123 = _mm_cvtps_epi32(vx0123);
53 const __m128i vy4567 = _mm_cvtps_epi32(vx4567);
54 const __m128i vy89AB = _mm_cvtps_epi32(vx89AB);
55 const __m128i vyCDEF = _mm_cvtps_epi32(vxCDEF);
56
57 __m128i vy01234567 = _mm_packs_epi32(vy0123, vy4567);
58 __m128i vy89ABCDEF = _mm_packs_epi32(vy89AB, vyCDEF);
59
60 vy01234567 = _mm_adds_epi16(vy01234567, voutput_zero_point);
61 vy89ABCDEF = _mm_adds_epi16(vy89ABCDEF, voutput_zero_point);
62
63 vy01234567 = _mm_max_epi16(vy01234567, voutput_min);
64 vy89ABCDEF = _mm_max_epi16(vy89ABCDEF, voutput_min);
65
66 __m128i vy0123456789ABCDEF = _mm_packs_epi16(vy01234567, vy89ABCDEF);
67
68
69 _mm_storeu_si128((__m128i*) y, vy0123456789ABCDEF);
70 y += 16;
71 }
72 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
73 __m128 vx_lo = _mm_loadu_ps(x);
74 __m128 vx_hi = _mm_loadu_ps(x + 4);
75 x += 8;
76
77 vx_lo = _mm_mul_ps(vx_lo, vscale);
78 vx_hi = _mm_mul_ps(vx_hi, vscale);
79
80 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
81 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
82
83 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
84 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
85
86 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
87 vy = _mm_adds_epi16(vy, voutput_zero_point);
88 vy = _mm_max_epi16(vy, voutput_min);
89 vy = _mm_packs_epi16(vy, vy);
90
91 _mm_storel_epi64((__m128i*) y, vy);
92 y += 8;
93 }
94 if XNN_UNLIKELY(n != 0) {
95 __m128 vx_lo = _mm_loadu_ps(x);
96 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
97 __m128 vx_hi = _mm_loadu_ps(x_hi);
98
99 vx_lo = _mm_mul_ps(vx_lo, vscale);
100 vx_hi = _mm_mul_ps(vx_hi, vscale);
101
102 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
103 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
104
105 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
106 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
107
108 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
109 vy = _mm_adds_epi16(vy, voutput_zero_point);
110 vy = _mm_max_epi16(vy, voutput_min);
111 vy = _mm_packs_epi16(vy, vy);
112
113 if (n & (4 * sizeof(float))) {
114 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
115 y += 4;
116 vy = _mm_srli_epi64(vy, 32);
117 }
118 {
119 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
120 if (n & (2 * sizeof(float))) {
121 unaligned_store_u16(y, (uint16_t) vy_lo);
122 y += 2;
123 vy_lo >>= 16;
124 }
125 if (n & (1 * sizeof(float))) {
126 *y = (int8_t) vy_lo;
127 }
128 }
129 }
130 }
131