Debian Codegraph @

src/qu8-f32-vcvt/gen/qu8-f32-vcvt-sse41-u8.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_u8(
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 >= 8 * sizeof(uint8_t); batch -= 8 * 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 input += 8;
38
39 vx0123 = _mm_sub_epi32(vx0123, vzero_point);
40 vx4567 = _mm_sub_epi32(vx4567, vzero_point);
41
42 __m128 vy0123 = _mm_cvtepi32_ps(vx0123);
43 __m128 vy4567 = _mm_cvtepi32_ps(vx4567);
44
45 vy0123 = _mm_mul_ps(vy0123, vscale);
46 vy4567 = _mm_mul_ps(vy4567, vscale);
47
48 _mm_storeu_ps(output, vy0123);
49 _mm_storeu_ps(output + 4, vy4567);
50 output += 8;
51 }
52 for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) {
53 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
54 vx = _mm_sub_epi32(vx, vzero_point);
55 input += 4;
56
57 __m128 vy = _mm_cvtepi32_ps(vx);
58 vy = _mm_mul_ps(vy, vscale);
59
60 _mm_storeu_ps(output, vy);
61 output += 4;
62 }
63 if XNN_UNLIKELY(batch != 0) {
64 assert(batch >= 1 * sizeof(uint8_t));
65 assert(batch <= 3 * sizeof(uint8_t));
66
67 __m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input)));
68 vx = _mm_sub_epi32(vx, vzero_point);
69
70 __m128 vy = _mm_cvtepi32_ps(vx);
71 vy = _mm_mul_ps(vy, vscale);
72
73 if (batch & (2 * sizeof(uint8_t))) {
74 _mm_storel_pi((__m64*) output, vy);
75 vy = _mm_movehl_ps(vy, vy);
76 output += 2;
77 }
78 if (batch & (1 * sizeof(uint8_t))) {
79 _mm_store_ss(output, vy);
80 }
81 }
82 }