Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x4c2-minmax-fp32-avx-ld64.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c2-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 #include <smmintrin.h>
13
14 #include "xnnpack/igemm.h"
15 #include "xnnpack/math.h"
16 #include "xnnpack/unaligned.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x4c2__avx_ld64(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (1 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 2 * sizeof(int8_t));
45 int8_t* c0 = c;
46
47 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);
48 const __m128i voutput_zero_point = _mm_set1_epi16(params->fp32_scalar.output_zero_point);
49 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
50 XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
51 XNN_FORCE_REALIZATION(voutput_zero_point);
52 XNN_FORCE_REALIZATION(voutput_min);
53
54 do {
55 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
56 w = (const void*) ((const int32_t*) w + 4);
57
58 size_t p = ks;
59 do {
60 const int8_t* restrict a0 = a[0];
61 if XNN_UNPREDICTABLE(a0 != zero) {
62 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
63 }
64 a += 1;
65
66 size_t k = kc;
67 while (k >= 8 * sizeof(int8_t)) {
68 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
69 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
70 a0 += 8;
71
72 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
73 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
74
75 vacc0x0123 = _mm_add_epi32(vacc0x0123,
76 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
77 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
78 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
79
80 vacc0x0123 = _mm_add_epi32(vacc0x0123,
81 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
82 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
83 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
84
85 vacc0x0123 = _mm_add_epi32(vacc0x0123,
86 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
87 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
88 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
89
90 vacc0x0123 = _mm_add_epi32(vacc0x0123,
91 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
92
93 w = (const void*) ((const int8_t*) w + 32);
94 k -= 8 * sizeof(int8_t);
95 }
96 if (k != 0) {
97 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
98 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
99 a0 = (const int8_t*) ((uintptr_t) a0 + k);
100
101 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
102 w = (const int8_t*) w + 8;
103 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
104
105 vacc0x0123 = _mm_add_epi32(vacc0x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
107
108 if (k > 2 * sizeof(int8_t)) {
109 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
110 w = (const int8_t*) w + 8;
111 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
112
113 vacc0x0123 = _mm_add_epi32(vacc0x0123,
114 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
115
116 if (k > 4 * sizeof(int8_t)) {
117 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
118 w = (const int8_t*) w + 8;
119 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
120
121 vacc0x0123 = _mm_add_epi32(vacc0x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
123 }
124 }
125 }
126 p -= 1 * sizeof(void*);
127 } while (p != 0);
128
129 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
130
131 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
132 w = (const float*) w + 4;
133 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
134
135 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
136
137 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
138
139 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
140
141
142 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
143
144 vout = _mm_max_epi8(vout, voutput_min);
145
146 if (nc >= 4) {
147 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
148 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
149
150 a = (const int8_t**restrict) ((uintptr_t) a - ks);
151
152 nc -= 4;
153 } else {
154 if (nc & 2) {
155 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
156 c0 += 2;
157 vout = _mm_srli_epi32(vout, 16);
158 }
159 if (nc & 1) {
160 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
161 }
162
163 nc = 0;
164 }
165 } while (nc != 0);
166 }