1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vmulc/sse-mul16-ld64.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 <smmintrin.h>
13
14 #include <xnnpack/unaligned.h>
15 #include <xnnpack/vmul.h>
16
17
xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16(size_t n,const uint8_t * input_a,const uint8_t * input_b,uint8_t * output,const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16(
19 size_t n,
20 const uint8_t* input_a,
21 const uint8_t* input_b,
22 uint8_t* output,
23 const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
24
25 {
26 const __m128i va_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.a_zero_point);
27 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
28 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
29 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
30 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->fp32_sse2.output_max);
31
32 __m128i vxb = _mm_sub_epi16(
33 _mm_shuffle_epi32(_mm_cvtsi32_si128(UINT32_C(0x00010001) * (uint32_t) (uint16_t) (int16_t) *input_b), 0),
34 _mm_load_si128((const __m128i*) params->fp32_sse2.b_zero_point));
35 for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
36 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
37 const __m128i va89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + 8)));
38 input_a += 16;
39
40
41 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
42 const __m128i vxa89ABCDEF = _mm_sub_epi16(va89ABCDEF, va_zero_point);
43
44 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
45 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
46 const __m128i vprod89ABCDEFlo = _mm_mullo_epi16(vxa89ABCDEF, vxb);
47 const __m128i vprod89ABCDEFhi = _mm_mulhi_epi16(vxa89ABCDEF, vxb);
48
49 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
50 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
51 const __m128i vprod89AB = _mm_unpacklo_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
52 const __m128i vprodCDEF = _mm_unpackhi_epi16(vprod89ABCDEFlo, vprod89ABCDEFhi);
53
54 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
55 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
56 __m128 vfpacc89AB = _mm_cvtepi32_ps(vprod89AB);
57 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vprodCDEF);
58
59 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
60 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
61 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
62 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
63
64 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
65 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
66 const __m128i vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
67 const __m128i vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
68
69 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
70 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
71
72
73 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
74
75 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
76
77 vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max);
78
79 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
80 output += 16;
81 }
82 if XNN_UNLIKELY(n != 0) {
83 do {
84 const __m128i va01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
85 input_a += 8;
86
87
88 const __m128i vxa01234567 = _mm_sub_epi16(va01234567, va_zero_point);
89
90 const __m128i vprod01234567lo = _mm_mullo_epi16(vxa01234567, vxb);
91 const __m128i vprod01234567hi = _mm_mulhi_epi16(vxa01234567, vxb);
92
93 const __m128i vprod0123 = _mm_unpacklo_epi16(vprod01234567lo, vprod01234567hi);
94 const __m128i vprod4567 = _mm_unpackhi_epi16(vprod01234567lo, vprod01234567hi);
95
96 __m128 vfpacc0123 = _mm_cvtepi32_ps(vprod0123);
97 __m128 vfpacc4567 = _mm_cvtepi32_ps(vprod4567);
98
99 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
100 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
101
102 const __m128i vacc0123 = _mm_cvtps_epi32(vfpacc0123);
103 const __m128i vacc4567 = _mm_cvtps_epi32(vfpacc4567);
104
105 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
106
107 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
108 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
109 vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max);
110
111 if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
112 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
113 output += 8;
114 n -= 8 * sizeof(uint8_t);
115 } else {
116 if (n & (4 * sizeof(uint8_t))) {
117 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
118 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
119 output += 4;
120 }
121 if (n & (2 * sizeof(uint8_t))) {
122 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
123 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
124 output += 2;
125 }
126 if (n & (1 * sizeof(uint8_t))) {
127 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
128 }
129 n = 0;
130 }
131 } while (n != 0);
132 }
133 }
134