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
22
xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__xop_ld64(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_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2__xop_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_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, 2);
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 do {
56 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
57 __m128i vacc1x0123 = vacc0x0123;
58 w = (const void*) ((const int32_t*) w + 4);
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 }
66 const int8_t* restrict a1 = a[1];
67 if XNN_UNPREDICTABLE(a1 != zero) {
68 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
69 }
70 a += 2;
71
72 size_t k = kc;
73 while (k >= 8 * sizeof(int8_t)) {
74 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
75 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
76 a0 += 8;
77 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
78 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
79 a1 += 8;
80
81 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
82 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
83
84 vacc0x0123 = _mm_maddd_epi16(
85 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
86 vacc1x0123 = _mm_maddd_epi16(
87 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
88 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
89 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
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 vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
96 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
97
98 vacc0x0123 = _mm_maddd_epi16(
99 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
100 vacc1x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
102 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
103 const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
104
105 vacc0x0123 = _mm_maddd_epi16(
106 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
107 vacc1x0123 = _mm_maddd_epi16(
108 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
109
110 w = (const void*) ((const int8_t*) w + 32);
111 k -= 8 * sizeof(int8_t);
112 }
113 if (k != 0) {
114 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
115 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
116 a0 = (const int8_t*) ((uintptr_t) a0 + k);
117 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
118 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
119 a1 = (const int8_t*) ((uintptr_t) a1 + k);
120
121 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
122 w = (const void*) ((const int8_t*) w + 8);
123 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
124
125 vacc0x0123 = _mm_maddd_epi16(
126 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
127 vacc1x0123 = _mm_maddd_epi16(
128 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
129
130 if (k > 2 * sizeof(int8_t)) {
131 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
132 w = (const void*) ((const int8_t*) w + 8);
133 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
134
135 vacc0x0123 = _mm_maddd_epi16(
136 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
137 vacc1x0123 = _mm_maddd_epi16(
138 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
139
140 if (k > 4 * sizeof(int8_t)) {
141 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
142 w = (const void*) ((const int8_t*) w + 8);
143 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
144
145 vacc0x0123 = _mm_maddd_epi16(
146 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
147 vacc1x0123 = _mm_maddd_epi16(
148 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
149 }
150 }
151 }
152 p -= 2 * sizeof(void*);
153 } while (p != 0);
154
155 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
156 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
157
158 const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
159 w = (const void*) ((const float*) w + 4);
160 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
161 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
162
163 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse4.output_max_less_zero_point);
164 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
165 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
166
167 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
168 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
169
170 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
171 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
172
173
174 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
175
176 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
177
178 if (nc >= 4) {
179 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
180 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
181 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
182 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
183
184 a = (const int8_t**restrict) ((uintptr_t) a - ks);
185
186 nc -= 4;
187 } else {
188 if (nc & 2) {
189 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
190 c1 += 2;
191 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
192 c0 += 2;
193 vout = _mm_srli_epi32(vout, 16);
194 }
195 if (nc & 1) {
196 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
197 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
198 }
199
200 nc = 0;
201 }
202 } while (nc != 0);
203 }
204