Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x4c8-minmax-fp32-avx-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 #ifdef _MSC_VER
13 #include <intrin.h>
14 #else
15 #include <x86intrin.h>
16 #endif
17
18 #include "xnnpack/igemm.h"
19 #include "xnnpack/math.h"
20 #include "xnnpack/unaligned.h"
21
22
23 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x4c8__avx_ld64(
24 size_t mr,
25 size_t nc,
26 size_t kc,
27 size_t ks,
28 const int8_t** restrict a,
29 const void* restrict w,
30 int8_t* restrict c,
31 size_t cm_stride,
32 size_t cn_stride,
33 size_t a_offset,
34 const int8_t* zero,
35 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 1);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(ks != 0);
42 assert(ks % (1 * sizeof(void*)) == 0);
43 assert(a_offset % sizeof(int8_t) == 0);
44 assert(a != NULL);
45 assert(w != NULL);
46 assert(c != NULL);
47
48 kc = round_up_po2(kc, 8 * sizeof(int8_t));
49 int8_t* c0 = c;
50
51
52 const __m128 voutput_max_less_zero_point = _mm_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
53 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
54 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
55 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
56 XNN_FORCE_REALIZATION(voutput_zero_point);
57 XNN_FORCE_REALIZATION(voutput_min);
58
59
60 do {
61 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
62 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
63 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
64 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
65 w = (const int32_t*) w + 4;
66
67 size_t p = ks;
68 do {
69 const int8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72 }
73 a += 1;
74
75 size_t k = 0;
76 while (k < kc) {
77 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
78 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
79 a0 += 8;
80
81 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
82 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
83
84 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
85 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
86 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
87
88 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
89 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
90 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
91
92 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
93 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
94 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
95
96 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
97
98 w = (const void*) ((const int8_t*) w + 32);
99 k += 8 * sizeof(int8_t);
100 }
101 p -= 1 * sizeof(void*);
102 } while (p != 0);
103
104 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
105 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
106
107 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
108
109 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
110
111 const __m128 vscale0123 = _mm_load_ps((const float*) w);
112 w = (const float*) w + 4;
113 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
114
115 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
116
117 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
118
119 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
120
121
122 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
123
124 vout = _mm_max_epi8(vout, voutput_min);
125
126 if (nc >= 4) {
127 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
128 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
129
130 a = (const int8_t**restrict) ((uintptr_t) a - ks);
131
132 nc -= 4;
133 } else {
134 if (nc & 2) {
135 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
136 c0 += 2;
137 vout = _mm_srli_epi32(vout, 16);
138 }
139 if (nc & 1) {
140 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
141 }
142
143 nc = 0;
144 }
145 } while (nc != 0);
146 }