1 // Auto-generated file. Do not edit!
2 // Template: src/f32-qs8-vcvt/sse.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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/unaligned.h>
16 #include <xnnpack/vcvt.h>
17
18
xnn_f32_qs8_vcvt_ukernel__sse2_x8(size_t n,const float * x,int8_t * y,const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_qs8_vcvt_ukernel__sse2_x8(
20 size_t n,
21 const float* x,
22 int8_t* y,
23 const union xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m128 vscale = _mm_load_ps(params->sse2.scale);
31 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
32 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
33 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
34
35 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
36 __m128 vx_lo = _mm_loadu_ps(x);
37 __m128 vx_hi = _mm_loadu_ps(x + 4);
38 x += 8;
39
40 vx_lo = _mm_mul_ps(vx_lo, vscale);
41 vx_hi = _mm_mul_ps(vx_hi, vscale);
42
43 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
44 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
45
46 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
47 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
48
49 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
50 vy = _mm_adds_epi16(vy, voutput_zero_point);
51 vy = _mm_max_epi16(vy, voutput_min);
52 vy = _mm_packs_epi16(vy, vy);
53
54 _mm_storel_epi64((__m128i*) y, vy);
55 y += 8;
56 }
57 if XNN_UNLIKELY(n != 0) {
58 __m128 vx_lo = _mm_loadu_ps(x);
59 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
60 __m128 vx_hi = _mm_loadu_ps(x_hi);
61
62 vx_lo = _mm_mul_ps(vx_lo, vscale);
63 vx_hi = _mm_mul_ps(vx_hi, vscale);
64
65 vx_lo = _mm_min_ps(vx_lo, voutput_max_less_zero_point);
66 vx_hi = _mm_min_ps(vx_hi, voutput_max_less_zero_point);
67
68 const __m128i vy_lo = _mm_cvtps_epi32(vx_lo);
69 const __m128i vy_hi = _mm_cvtps_epi32(vx_hi);
70
71 __m128i vy = _mm_packs_epi32(vy_lo, vy_hi);
72 vy = _mm_adds_epi16(vy, voutput_zero_point);
73 vy = _mm_max_epi16(vy, voutput_min);
74 vy = _mm_packs_epi16(vy, vy);
75
76 if (n & (4 * sizeof(float))) {
77 unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
78 y += 4;
79 vy = _mm_srli_epi64(vy, 32);
80 }
81 {
82 uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
83 if (n & (2 * sizeof(float))) {
84 unaligned_store_u16(y, (uint16_t) vy_lo);
85 y += 2;
86 vy_lo >>= 16;
87 }
88 if (n & (1 * sizeof(float))) {
89 *y = (int8_t) vy_lo;
90 }
91 }
92 }
93 }
94