xref: /aosp_15_r20/external/XNNPACK/src/qs8-f32-vcvt/gen/vcvt-sse41-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_qs8_f32_vcvt_ukernel__sse41_x8(size_t n,const int8_t * x,float * y,const union xnn_qs8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_f32_vcvt_ukernel__sse41_x8(
20     size_t n,
21     const int8_t* x,
22     float* y,
23     const union xnn_qs8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25   assert(n != 0);
26   assert(n % sizeof(int8_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 >= 8 * sizeof(int8_t); n -= 8 * sizeof(int8_t)) {
33     __m128i vx0123 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
34     __m128i vx4567 = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x + 4)));
35     x += 8;
36 
37     vx0123 = _mm_add_epi32(vx0123, vminus_zero_point);
38     vx4567 = _mm_add_epi32(vx4567, vminus_zero_point);
39 
40     __m128 vy0123 = _mm_cvtepi32_ps(vx0123);
41     __m128 vy4567 = _mm_cvtepi32_ps(vx4567);
42 
43     vy0123 = _mm_mul_ps(vy0123, vscale);
44     vy4567 = _mm_mul_ps(vy4567, vscale);
45 
46     _mm_storeu_ps(y, vy0123);
47     _mm_storeu_ps(y + 4, vy4567);
48     y += 8;
49   }
50   for (; n >= 4 * sizeof(int8_t); n -= 4 * sizeof(int8_t)) {
51     __m128i vx = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
52     vx = _mm_add_epi32(vx, vminus_zero_point);
53     x += 4;
54 
55     __m128 vy = _mm_cvtepi32_ps(vx);
56     vy = _mm_mul_ps(vy, vscale);
57 
58     _mm_storeu_ps(y, vy);
59     y += 4;
60   }
61   if XNN_UNLIKELY(n != 0) {
62     assert(n >= 1 * sizeof(int8_t));
63     assert(n <= 3 * sizeof(int8_t));
64 
65     __m128i vx = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(x)));
66     vx = _mm_add_epi32(vx, vminus_zero_point);
67 
68     __m128 vy = _mm_cvtepi32_ps(vx);
69     vy = _mm_mul_ps(vy, vscale);
70 
71     if (n & (2 * sizeof(int8_t))) {
72       _mm_storel_pi((__m64*) y, vy);
73       vy = _mm_movehl_ps(vy, vy);
74       y += 2;
75     }
76     if (n & (1 * sizeof(int8_t))) {
77       _mm_store_ss(y, vy);
78     }
79   }
80 }
81