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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/igemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld128(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 size_t ks,
29 const int8_t** restrict a,
30 const void* restrict w,
31 int8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 size_t a_offset,
35 const int8_t* zero,
36 const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 2);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (2 * sizeof(void*)) == 0);
44 assert(a_offset % sizeof(int8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 2 * sizeof(int8_t));
50 int8_t* c0 = c;
51 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr != 2) {
53 c1 = c0;
54 }
55
56 do {
57 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
58 __m128i vacc1x0123 = vacc0x0123;
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const int8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 a += 2;
72
73 size_t k = kc;
74 while (k >= 8 * sizeof(int8_t)) {
75 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
76 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
77 a0 += 8;
78 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
79 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
80 a1 += 8;
81
82 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
83 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
84 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
85
86 vacc0x0123 = _mm_maddd_epi16(
87 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
88 vacc1x0123 = _mm_maddd_epi16(
89 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
90
91 vacc0x0123 = _mm_maddd_epi16(
92 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
93 vacc1x0123 = _mm_maddd_epi16(
94 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
95 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
96 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
97 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
98
99 vacc0x0123 = _mm_maddd_epi16(
100 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
101 vacc1x0123 = _mm_maddd_epi16(
102 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
103
104 vacc0x0123 = _mm_maddd_epi16(
105 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
106 vacc1x0123 = _mm_maddd_epi16(
107 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
108
109 w = (const void*) ((const int8_t*) w + 32);
110 k -= 8 * sizeof(int8_t);
111 }
112 if (k != 0) {
113 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
114 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
115 a0 = (const int8_t*) ((uintptr_t) a0 + k);
116 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
117 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
118 a1 = (const int8_t*) ((uintptr_t) a1 + k);
119
120 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
121 w = (const void*) ((const int8_t*) w + 8);
122 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
123
124 vacc0x0123 = _mm_maddd_epi16(
125 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
126 vacc1x0123 = _mm_maddd_epi16(
127 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
128
129 if (k > 2 * sizeof(int8_t)) {
130 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
131 w = (const void*) ((const int8_t*) w + 8);
132 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
133
134 vacc0x0123 = _mm_maddd_epi16(
135 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
136 vacc1x0123 = _mm_maddd_epi16(
137 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
138
139 if (k > 4 * sizeof(int8_t)) {
140 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
141 w = (const void*) ((const int8_t*) w + 8);
142 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
143
144 vacc0x0123 = _mm_maddd_epi16(
145 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
146 vacc1x0123 = _mm_maddd_epi16(
147 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
148 }
149 }
150 }
151 p -= 2 * sizeof(void*);
152 } while (p != 0);
153
154 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
155 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
156
157 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
158 w = (const void*) ((const float*) w + 4);
159 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
160 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
161
162 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
163 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
164 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
165
166 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
167 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
168
169 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
170 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
171
172
173 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
174
175 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
176
177 if (nc >= 4) {
178 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
179 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
180 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
181 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
182
183 a = (const int8_t**restrict) ((uintptr_t) a - ks);
184
185 nc -= 4;
186 } else {
187 if (nc & 2) {
188 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
189 c1 += 2;
190 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
191 c0 += 2;
192 vout = _mm_srli_epi32(vout, 16);
193 }
194 if (nc & 1) {
195 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
196 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
197 }
198
199 nc = 0;
200 }
201 } while (nc != 0);
202 }
203