1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-f32-vcvt/sse4.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/unaligned.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_qu8_f32_vcvt_ukernel__sse41_x32(size_t n,const uint8_t * x,float * y,const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_f32_vcvt_ukernel__sse41_x32(
20 size_t n,
21 const uint8_t* x,
22 float* y,
23 const union xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(uint8_t) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m128i vminus_zero_point = _mm_load_si128((const __m128i*) params->sse4.minus_zero_point);
31 const __m128 vscale = _mm_load_ps(params->sse4.scale);
32 for (; n >= 32 * sizeof(uint8_t); n -= 32 * sizeof(uint8_t)) {
33 __m128i vx0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
34 __m128i vx4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 4)));
35 __m128i vx89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 8)));
36 __m128i vxCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 12)));
37 __m128i vxGHIJ = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 16)));
38 __m128i vxKLMN = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 20)));
39 __m128i vxOPQR = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 24)));
40 __m128i vxSTUV = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 28)));
41 x += 32;
42
43 vx0123 = _mm_add_epi32(vx0123, vminus_zero_point);
44 vx4567 = _mm_add_epi32(vx4567, vminus_zero_point);
45 vx89AB = _mm_add_epi32(vx89AB, vminus_zero_point);
46 vxCDEF = _mm_add_epi32(vxCDEF, vminus_zero_point);
47 vxGHIJ = _mm_add_epi32(vxGHIJ, vminus_zero_point);
48 vxKLMN = _mm_add_epi32(vxKLMN, vminus_zero_point);
49 vxOPQR = _mm_add_epi32(vxOPQR, vminus_zero_point);
50 vxSTUV = _mm_add_epi32(vxSTUV, vminus_zero_point);
51
52 __m128 vy0123 = _mm_cvtepi32_ps(vx0123);
53 __m128 vy4567 = _mm_cvtepi32_ps(vx4567);
54 __m128 vy89AB = _mm_cvtepi32_ps(vx89AB);
55 __m128 vyCDEF = _mm_cvtepi32_ps(vxCDEF);
56 __m128 vyGHIJ = _mm_cvtepi32_ps(vxGHIJ);
57 __m128 vyKLMN = _mm_cvtepi32_ps(vxKLMN);
58 __m128 vyOPQR = _mm_cvtepi32_ps(vxOPQR);
59 __m128 vySTUV = _mm_cvtepi32_ps(vxSTUV);
60
61 vy0123 = _mm_mul_ps(vy0123, vscale);
62 vy4567 = _mm_mul_ps(vy4567, vscale);
63 vy89AB = _mm_mul_ps(vy89AB, vscale);
64 vyCDEF = _mm_mul_ps(vyCDEF, vscale);
65 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale);
66 vyKLMN = _mm_mul_ps(vyKLMN, vscale);
67 vyOPQR = _mm_mul_ps(vyOPQR, vscale);
68 vySTUV = _mm_mul_ps(vySTUV, vscale);
69
70 _mm_storeu_ps(y, vy0123);
71 _mm_storeu_ps(y + 4, vy4567);
72 _mm_storeu_ps(y + 8, vy89AB);
73 _mm_storeu_ps(y + 12, vyCDEF);
74 _mm_storeu_ps(y + 16, vyGHIJ);
75 _mm_storeu_ps(y + 20, vyKLMN);
76 _mm_storeu_ps(y + 24, vyOPQR);
77 _mm_storeu_ps(y + 28, vySTUV);
78 y += 32;
79 }
80 for (; n >= 4 * sizeof(uint8_t); n -= 4 * sizeof(uint8_t)) {
81 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
82 vx = _mm_add_epi32(vx, vminus_zero_point);
83 x += 4;
84
85 __m128 vy = _mm_cvtepi32_ps(vx);
86 vy = _mm_mul_ps(vy, vscale);
87
88 _mm_storeu_ps(y, vy);
89 y += 4;
90 }
91 if XNN_UNLIKELY(n != 0) {
92 assert(n >= 1 * sizeof(uint8_t));
93 assert(n <= 3 * sizeof(uint8_t));
94
95 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
96 vx = _mm_add_epi32(vx, vminus_zero_point);
97
98 __m128 vy = _mm_cvtepi32_ps(vx);
99 vy = _mm_mul_ps(vy, vscale);
100
101 if (n & (2 * sizeof(uint8_t))) {
102 _mm_storel_pi((__m64*) y, vy);
103 vy = _mm_movehl_ps(vy, vy);
104 y += 2;
105 }
106 if (n & (1 * sizeof(uint8_t))) {
107 _mm_store_ss(y, vy);
108 }
109 }
110 }
111