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_x24(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_x24(
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 >= 24 * sizeof(uint8_t); n -= 24 * 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 x += 24;
40
41 vx0123 = _mm_add_epi32(vx0123, vminus_zero_point);
42 vx4567 = _mm_add_epi32(vx4567, vminus_zero_point);
43 vx89AB = _mm_add_epi32(vx89AB, vminus_zero_point);
44 vxCDEF = _mm_add_epi32(vxCDEF, vminus_zero_point);
45 vxGHIJ = _mm_add_epi32(vxGHIJ, vminus_zero_point);
46 vxKLMN = _mm_add_epi32(vxKLMN, vminus_zero_point);
47
48 __m128 vy0123 = _mm_cvtepi32_ps(vx0123);
49 __m128 vy4567 = _mm_cvtepi32_ps(vx4567);
50 __m128 vy89AB = _mm_cvtepi32_ps(vx89AB);
51 __m128 vyCDEF = _mm_cvtepi32_ps(vxCDEF);
52 __m128 vyGHIJ = _mm_cvtepi32_ps(vxGHIJ);
53 __m128 vyKLMN = _mm_cvtepi32_ps(vxKLMN);
54
55 vy0123 = _mm_mul_ps(vy0123, vscale);
56 vy4567 = _mm_mul_ps(vy4567, vscale);
57 vy89AB = _mm_mul_ps(vy89AB, vscale);
58 vyCDEF = _mm_mul_ps(vyCDEF, vscale);
59 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale);
60 vyKLMN = _mm_mul_ps(vyKLMN, vscale);
61
62 _mm_storeu_ps(y, vy0123);
63 _mm_storeu_ps(y + 4, vy4567);
64 _mm_storeu_ps(y + 8, vy89AB);
65 _mm_storeu_ps(y + 12, vyCDEF);
66 _mm_storeu_ps(y + 16, vyGHIJ);
67 _mm_storeu_ps(y + 20, vyKLMN);
68 y += 24;
69 }
70 for (; n >= 4 * sizeof(uint8_t); n -= 4 * sizeof(uint8_t)) {
71 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
72 vx = _mm_add_epi32(vx, vminus_zero_point);
73 x += 4;
74
75 __m128 vy = _mm_cvtepi32_ps(vx);
76 vy = _mm_mul_ps(vy, vscale);
77
78 _mm_storeu_ps(y, vy);
79 y += 4;
80 }
81 if XNN_UNLIKELY(n != 0) {
82 assert(n >= 1 * sizeof(uint8_t));
83 assert(n <= 3 * sizeof(uint8_t));
84
85 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
86 vx = _mm_add_epi32(vx, vminus_zero_point);
87
88 __m128 vy = _mm_cvtepi32_ps(vx);
89 vy = _mm_mul_ps(vy, vscale);
90
91 if (n & (2 * sizeof(uint8_t))) {
92 _mm_storel_pi((__m64*) y, vy);
93 vy = _mm_movehl_ps(vy, vy);
94 y += 2;
95 }
96 if (n & (1 * sizeof(uint8_t))) {
97 _mm_store_ss(y, vy);
98 }
99 }
100 }
101