1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
24
xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__xop(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])25 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__xop(
26 size_t mr,
27 size_t nc,
28 size_t kc,
29 const int8_t* restrict a,
30 size_t a_stride,
31 const void* restrict w,
32 int8_t* restrict c,
33 size_t cm_stride,
34 size_t cn_stride,
35 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37 assert(mr != 0);
38 assert(mr <= 4);
39 assert(nc != 0);
40 assert(kc != 0);
41 assert(kc % sizeof(int8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 2 * sizeof(int8_t));
47 const int8_t* a0 = a;
48 int8_t* c0 = c;
49 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
50 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
56 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
62 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
63 if XNN_UNPREDICTABLE(mr != 4) {
64 a3 = a2;
65 c3 = c2;
66 }
67
68 do {
69 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
70 __m128i vacc1x0123 = vacc0x0123;
71 __m128i vacc2x0123 = vacc0x0123;
72 __m128i vacc3x0123 = vacc0x0123;
73 w = (const void*) ((const int32_t*) w + 4);
74
75 size_t k = kc;
76 while (k >= 8 * sizeof(int8_t)) {
77 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
78 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
79 a0 += 8;
80 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
81 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
82 a1 += 8;
83 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
84 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
85 a2 += 8;
86 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
87 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
88 a3 += 8;
89
90 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
91
92 vacc0x0123 = _mm_maddd_epi16(
93 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
94 vacc1x0123 = _mm_maddd_epi16(
95 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
96 vacc2x0123 = _mm_maddd_epi16(
97 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
98 vacc3x0123 = _mm_maddd_epi16(
99 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
100 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
101
102 vacc0x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
104 vacc1x0123 = _mm_maddd_epi16(
105 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
106 vacc2x0123 = _mm_maddd_epi16(
107 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
108 vacc3x0123 = _mm_maddd_epi16(
109 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
110 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
111
112 vacc0x0123 = _mm_maddd_epi16(
113 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
114 vacc1x0123 = _mm_maddd_epi16(
115 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
116 vacc2x0123 = _mm_maddd_epi16(
117 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
118 vacc3x0123 = _mm_maddd_epi16(
119 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
120 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
121
122 vacc0x0123 = _mm_maddd_epi16(
123 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
124 vacc1x0123 = _mm_maddd_epi16(
125 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
126 vacc2x0123 = _mm_maddd_epi16(
127 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
128 vacc3x0123 = _mm_maddd_epi16(
129 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
130
131 w = (const void*) ((const int16_t*) w + 32);
132 k -= 8 * sizeof(int8_t);
133 }
134 if (k != 0) {
135 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
136 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
137 a0 = (const int8_t*) ((uintptr_t) a0 + k);
138 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
139 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
140 a1 = (const int8_t*) ((uintptr_t) a1 + k);
141 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
142 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
143 a2 = (const int8_t*) ((uintptr_t) a2 + k);
144 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
145 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
146 a3 = (const int8_t*) ((uintptr_t) a3 + k);
147
148 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
149 w = (const void*) ((const int16_t*) w + 8);
150
151 vacc0x0123 = _mm_maddd_epi16(
152 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
153 vacc1x0123 = _mm_maddd_epi16(
154 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
155 vacc2x0123 = _mm_maddd_epi16(
156 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
157 vacc3x0123 = _mm_maddd_epi16(
158 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
159
160 if (k > 2 * sizeof(int8_t)) {
161 const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
162 w = (const void*) ((const int16_t*) w + 8);
163
164 vacc0x0123 = _mm_maddd_epi16(
165 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
166 vacc1x0123 = _mm_maddd_epi16(
167 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
168 vacc2x0123 = _mm_maddd_epi16(
169 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
170 vacc3x0123 = _mm_maddd_epi16(
171 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
172
173 if (k > 4 * sizeof(int8_t)) {
174 const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
175 w = (const void*) ((const int16_t*) w + 8);
176
177 vacc0x0123 = _mm_maddd_epi16(
178 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
179 vacc1x0123 = _mm_maddd_epi16(
180 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
181 vacc2x0123 = _mm_maddd_epi16(
182 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
183 vacc3x0123 = _mm_maddd_epi16(
184 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
185 }
186 }
187 }
188
189 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
190 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
191 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
192 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
193
194 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
195 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
196 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
197 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
198 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
199
200 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
201 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
202 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
203 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
204 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
205
206 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
207 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
208 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
209 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
210
211 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
212 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
213 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
214
215
216 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
217
218 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
219
220 if (nc >= 4) {
221 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
222 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
223 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
224 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
225
226 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
227 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
228 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
229 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
230
231 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
232 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
233 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
234 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
235
236 nc -= 4;
237 } else {
238 if (nc & 2) {
239 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
240 c0 += 2;
241 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
242 c1 += 2;
243 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
244 c2 += 2;
245 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
246 c3 += 2;
247 vout = _mm_srli_epi32(vout, 16);
248 }
249 if (nc & 1) {
250 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
251 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
252 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
253 *c3 = (int8_t) _mm_extract_epi8(vout, 12);
254 }
255
256 nc = 0;
257 }
258 } while (nc != 0);
259 }
260