Debian Codegraph @

src/qu8-f32-vcvt/gen/qu8-f32-vcvt-sse41-u24.c

raw
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
19 void xnn_qu8_f32_vcvt_ukernel__sse41_u24(
20 size_t batch,
21 const uint8_t* input,
22 float* output,
23 const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(uint8_t) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const __m128i vzero_point = _mm_set1_epi32(params->scalar.zero_point);
31 const __m128 vscale = _mm_set1_ps(params->scalar.scale);
32 XNN_FORCE_REALIZATION(vzero_point);
33 XNN_FORCE_REALIZATION(vscale);
34 for (; batch >= 24 * sizeof(uint8_t); batch -= 24 * sizeof(uint8_t)) {
35 __m128i vx0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
36 __m128i vx4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 4)));
37 __m128i vx89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 8)));
38 __m128i vxCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 12)));
39 __m128i vxGHIJ = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 16)));
40 __m128i vxKLMN = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 20)));
41 input += 24;
42
43 vx0123 = _mm_sub_epi32(vx0123, vzero_point);
44 vx4567 = _mm_sub_epi32(vx4567, vzero_point);
45 vx89AB = _mm_sub_epi32(vx89AB, vzero_point);
46 vxCDEF = _mm_sub_epi32(vxCDEF, vzero_point);
47 vxGHIJ = _mm_sub_epi32(vxGHIJ, vzero_point);
48 vxKLMN = _mm_sub_epi32(vxKLMN, vzero_point);
49
50 __m128 vy0123 = _mm_cvtepi32_ps(vx0123);
51 __m128 vy4567 = _mm_cvtepi32_ps(vx4567);
52 __m128 vy89AB = _mm_cvtepi32_ps(vx89AB);
53 __m128 vyCDEF = _mm_cvtepi32_ps(vxCDEF);
54 __m128 vyGHIJ = _mm_cvtepi32_ps(vxGHIJ);
55 __m128 vyKLMN = _mm_cvtepi32_ps(vxKLMN);
56
57 vy0123 = _mm_mul_ps(vy0123, vscale);
58 vy4567 = _mm_mul_ps(vy4567, vscale);
59 vy89AB = _mm_mul_ps(vy89AB, vscale);
60 vyCDEF = _mm_mul_ps(vyCDEF, vscale);
61 vyGHIJ = _mm_mul_ps(vyGHIJ, vscale);
62 vyKLMN = _mm_mul_ps(vyKLMN, vscale);
63
64 _mm_storeu_ps(output, vy0123);
65 _mm_storeu_ps(output + 4, vy4567);
66 _mm_storeu_ps(output + 8, vy89AB);
67 _mm_storeu_ps(output + 12, vyCDEF);
68 _mm_storeu_ps(output + 16, vyGHIJ);
69 _mm_storeu_ps(output + 20, vyKLMN);
70 output += 24;
71 }
72 for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) {
73 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
74 vx = _mm_sub_epi32(vx, vzero_point);
75 input += 4;
76
77 __m128 vy = _mm_cvtepi32_ps(vx);
78 vy = _mm_mul_ps(vy, vscale);
79
80 _mm_storeu_ps(output, vy);
81 output += 4;
82 }
83 if XNN_UNLIKELY(batch != 0) {
84 assert(batch >= 1 * sizeof(uint8_t));
85 assert(batch <= 3 * sizeof(uint8_t));
86
87 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
88 vx = _mm_sub_epi32(vx, vzero_point);
89
90 __m128 vy = _mm_cvtepi32_ps(vx);
91 vy = _mm_mul_ps(vy, vscale);
92
93 if (batch & (2 * sizeof(uint8_t))) {
94 _mm_storel_pi((__m64*) output, vy);
95 vy = _mm_movehl_ps(vy, vy);
96 output += 2;
97 }
98 if (batch & (1 * sizeof(uint8_t))) {
99 _mm_store_ss(output, vy);
100 }
101 }
102 }