Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x4c8-minmax-fp32-sse41-ld128.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 #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_2x4c8__sse41_ld128(
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 <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (2 * 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, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr != 2) {
48 c1 = c0;
49 }
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 __m128i vacc1x0 = vacc0x0;
66 __m128i vacc1x1 = vacc0x1;
67 __m128i vacc1x2 = vacc0x2;
68 __m128i vacc1x3 = vacc0x3;
69 w = (const int32_t*) w + 4;
70
71 size_t p = ks;
72 do {
73 const int8_t* restrict a0 = a[0];
74 if XNN_UNPREDICTABLE(a0 != zero) {
75 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
76 }
77 const int8_t* restrict a1 = a[1];
78 if XNN_UNPREDICTABLE(a1 != zero) {
79 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
80 }
81 a += 2;
82
83 size_t k = 0;
84 while (k < kc) {
85 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
86 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
87 a0 += 8;
88 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
89 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
90 a1 += 8;
91
92 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
93 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
94 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
95
96 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
97 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
98 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
99 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
100 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
101 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
102 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
103
104 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
105 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
106 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
107 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
108
109 w = (const void*) ((const int8_t*) w + 32);
110 k += 8 * sizeof(int8_t);
111 }
112 p -= 2 * sizeof(void*);
113 } while (p != 0);
114
115 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
116 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
117 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
118 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
119
120 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
121 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
122
123 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
124 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
125
126 const __m128 vscale0123 = _mm_load_ps((const float*) w);
127 w = (const float*) w + 4;
128 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
129 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
130
131 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
132 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
133
134 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
135 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
136
137 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
138
139
140 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
141
142 vout = _mm_max_epi8(vout, voutput_min);
143
144 if (nc >= 4) {
145 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
146 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
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(c1, (uint16_t) _mm_extract_epi16(vout, 2));
156 c1 += 2;
157 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
158 c0 += 2;
159 vout = _mm_srli_epi32(vout, 16);
160 }
161 if (nc & 1) {
162 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
163 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
164 }
165
166 nc = 0;
167 }
168 } while (nc != 0);
169 }