Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x4c8-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_2x4c8__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 <= 2);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(ks != 0);
42 assert(ks % (2 * 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 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr != 2) {
52 c1 = c0;
53 }
54
55
56 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);
57 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
58 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
59 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
60 XNN_FORCE_REALIZATION(voutput_zero_point);
61 XNN_FORCE_REALIZATION(voutput_min);
62
63
64 do {
65 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
66 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
67 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
68 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
69 __m128i vacc1x0 = vacc0x0;
70 __m128i vacc1x1 = vacc0x1;
71 __m128i vacc1x2 = vacc0x2;
72 __m128i vacc1x3 = vacc0x3;
73 w = (const int32_t*) w + 4;
74
75 size_t p = ks;
76 do {
77 const int8_t* restrict a0 = a[0];
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
80 }
81 const int8_t* restrict a1 = a[1];
82 if XNN_UNPREDICTABLE(a1 != zero) {
83 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
84 }
85 a += 2;
86
87 size_t k = 0;
88 while (k < kc) {
89 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
90 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
91 a0 += 8;
92 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
93 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
94 a1 += 8;
95
96 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
97 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
98
99 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
100 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
101 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
102 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
103
104 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
105 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
106 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
107 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
108
109 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
110 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
111 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
112 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
113
114 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
115 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
116
117 w = (const void*) ((const int8_t*) w + 32);
118 k += 8 * sizeof(int8_t);
119 }
120 p -= 2 * sizeof(void*);
121 } while (p != 0);
122
123 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
124 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
125 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
126 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
127
128 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
129 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
130
131 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
132 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
133
134 const __m128 vscale0123 = _mm_load_ps((const float*) w);
135 w = (const float*) w + 4;
136 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
137 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
138
139 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
140 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
141
142 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
143 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
144
145 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
146
147
148 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
149
150 vout = _mm_max_epi8(vout, voutput_min);
151
152 if (nc >= 4) {
153 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
154 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
155 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
156 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
157
158 a = (const int8_t**restrict) ((uintptr_t) a - ks);
159
160 nc -= 4;
161 } else {
162 if (nc & 2) {
163 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
164 c1 += 2;
165 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
166 c0 += 2;
167 vout = _mm_srli_epi32(vout, 16);
168 }
169 if (nc & 1) {
170 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
171 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
172 }
173
174 nc = 0;
175 }
176 } while (nc != 0);
177 }