1 /*
2 * Copyright (c) Facebook, Inc. and its affiliates.
3 * All rights reserved.
4 *
5 * This source code is licensed under the BSD-style license found in the
6 * LICENSE file in the root directory of this source tree.
7 */
8
9 #include <assert.h>
10
11 #include <emmintrin.h>
12
13 #include <qnnpack/q8avgpool.h>
14
pytorch_q8avgpool_ukernel_up8xm__sse2(size_t n,size_t ks,size_t kc,const uint8_t ** input,const uint8_t * zero,uint8_t * output,size_t input_increment,size_t output_increment,const union pytorch_qnnp_avgpool_quantization_params quantization_params[RESTRICT_STATIC1])15 void pytorch_q8avgpool_ukernel_up8xm__sse2(
16 size_t n,
17 size_t ks,
18 size_t kc,
19 const uint8_t** input,
20 const uint8_t* zero,
21 uint8_t* output,
22 size_t input_increment,
23 size_t output_increment,
24 const union pytorch_qnnp_avgpool_quantization_params
25 quantization_params[RESTRICT_STATIC 1]) {
26 assert(n != 0);
27 assert(ks != 0);
28 assert(kc < 8);
29
30 const __m128i vbias =
31 _mm_load_si128((const __m128i*)&quantization_params->sse2.bias);
32 const __m128i vzero = _mm_setzero_si128();
33 const __m128 vscale = _mm_loadu_ps(quantization_params->sse2.scale);
34
35 do {
36 const uint8_t** next_input =
37 (const uint8_t**)((uintptr_t)input + input_increment);
38 __m128i vacc_lo = vbias;
39 __m128i vacc_hi = vbias;
40
41 size_t m = ks;
42 do {
43 const uint8_t* i = *input++;
44 i += kc;
45 __m128i vi = _mm_setzero_si128();
46 if (kc & 1) {
47 i -= 1;
48 vi = _mm_cvtsi32_si128((int)(uint32_t)*i);
49 }
50 if (kc & 2) {
51 vi = _mm_slli_epi32(vi, 16);
52 i -= 2;
53 vi = _mm_insert_epi16(vi, *((const uint16_t*)i), 0);
54 }
55 if (kc & 4) {
56 i -= 4;
57 vi = _mm_unpacklo_epi32(
58 _mm_cvtsi32_si128((int)*((const uint32_t*)i)), vi);
59 }
60
61 const __m128i vxi = _mm_unpacklo_epi8(vi, vzero);
62 vacc_lo = _mm_add_epi32(vacc_lo, _mm_unpacklo_epi16(vxi, vzero));
63 vacc_hi = _mm_add_epi32(vacc_hi, _mm_unpackhi_epi16(vxi, vzero));
64 } while (--m != 0);
65 input = next_input;
66
67 const __m128 vacc_lo_f = _mm_mul_ps(_mm_cvtepi32_ps(vacc_lo), vscale);
68 const __m128 vacc_hi_f = _mm_mul_ps(_mm_cvtepi32_ps(vacc_hi), vscale);
69
70 const __m128i vscaled_lo = _mm_cvtps_epi32(vacc_lo_f);
71 const __m128i vscaled_hi = _mm_cvtps_epi32(vacc_hi_f);
72
73 __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);
74 vout = _mm_adds_epi16(
75 vout,
76 _mm_load_si128(
77 (const __m128i*)quantization_params->sse2.output_zero_point));
78 vout = _mm_packus_epi16(vout, vout);
79 vout = _mm_min_epu8(
80 vout,
81 _mm_load_si128((const __m128i*)quantization_params->sse2.output_max));
82 vout = _mm_max_epu8(
83 vout,
84 _mm_load_si128((const __m128i*)quantization_params->sse2.output_min));
85
86 if (kc & 4) {
87 *((uint32_t*)output) = (uint32_t)_mm_cvtsi128_si32(vout);
88 output += 4;
89 vout = _mm_srli_epi64(vout, 32);
90 }
91 if (kc & 2) {
92 *((uint16_t*)output) = (uint16_t)_mm_extract_epi16(vout, 0);
93 output += 2;
94 vout = _mm_srli_epi32(vout, 16);
95 }
96 if (kc & 1) {
97 *((uint8_t*)output) = (uint8_t)_mm_cvtsi128_si32(vout);
98 output += 1;
99 }
100 output = (uint8_t*)((uintptr_t)output + output_increment);
101 } while (--n != 0);
102 }
103