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 #include <smmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
19
xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__sse41(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)])20 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_4x4c2__sse41(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 4);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 2 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 int8_t* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
57 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr != 4) {
59 a3 = a2;
60 c3 = c2;
61 }
62
63 do {
64 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
65 __m128i vacc1x0123 = vacc0x0123;
66 __m128i vacc2x0123 = vacc0x0123;
67 __m128i vacc3x0123 = vacc0x0123;
68 w = (const void*) ((const int32_t*) w + 4);
69
70 size_t k = kc;
71 while (k >= 8 * sizeof(int8_t)) {
72 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
73 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
74 a0 += 8;
75 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
76 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
77 a1 += 8;
78 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
79 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
80 a2 += 8;
81 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
82 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
83 a3 += 8;
84
85 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
86
87 vacc0x0123 = _mm_add_epi32(vacc0x0123,
88 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
89 vacc1x0123 = _mm_add_epi32(vacc1x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
91 vacc2x0123 = _mm_add_epi32(vacc2x0123,
92 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
93 vacc3x0123 = _mm_add_epi32(vacc3x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
95 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
96
97 vacc0x0123 = _mm_add_epi32(vacc0x0123,
98 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
99 vacc1x0123 = _mm_add_epi32(vacc1x0123,
100 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
101 vacc2x0123 = _mm_add_epi32(vacc2x0123,
102 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
103 vacc3x0123 = _mm_add_epi32(vacc3x0123,
104 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
105 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
106
107 vacc0x0123 = _mm_add_epi32(vacc0x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
109 vacc1x0123 = _mm_add_epi32(vacc1x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
111 vacc2x0123 = _mm_add_epi32(vacc2x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
113 vacc3x0123 = _mm_add_epi32(vacc3x0123,
114 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
115 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
116
117 vacc0x0123 = _mm_add_epi32(vacc0x0123,
118 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
119 vacc1x0123 = _mm_add_epi32(vacc1x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
121 vacc2x0123 = _mm_add_epi32(vacc2x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
123 vacc3x0123 = _mm_add_epi32(vacc3x0123,
124 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
125
126 w = (const void*) ((const int16_t*) w + 32);
127 k -= 8 * sizeof(int8_t);
128 }
129 if (k != 0) {
130 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
131 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
132 a0 = (const int8_t*) ((uintptr_t) a0 + k);
133 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
134 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
135 a1 = (const int8_t*) ((uintptr_t) a1 + k);
136 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
137 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
138 a2 = (const int8_t*) ((uintptr_t) a2 + k);
139 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
140 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
141 a3 = (const int8_t*) ((uintptr_t) a3 + k);
142
143 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
144 w = (const void*) ((const int16_t*) w + 8);
145
146 vacc0x0123 = _mm_add_epi32(vacc0x0123,
147 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
148 vacc1x0123 = _mm_add_epi32(vacc1x0123,
149 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
150 vacc2x0123 = _mm_add_epi32(vacc2x0123,
151 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
152 vacc3x0123 = _mm_add_epi32(vacc3x0123,
153 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
154
155 if (k > 2 * sizeof(int8_t)) {
156 const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
157 w = (const void*) ((const int16_t*) w + 8);
158
159 vacc0x0123 = _mm_add_epi32(vacc0x0123,
160 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
161 vacc1x0123 = _mm_add_epi32(vacc1x0123,
162 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
163 vacc2x0123 = _mm_add_epi32(vacc2x0123,
164 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
165 vacc3x0123 = _mm_add_epi32(vacc3x0123,
166 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
167
168 if (k > 4 * sizeof(int8_t)) {
169 const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
170 w = (const void*) ((const int16_t*) w + 8);
171
172 vacc0x0123 = _mm_add_epi32(vacc0x0123,
173 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
174 vacc1x0123 = _mm_add_epi32(vacc1x0123,
175 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
176 vacc2x0123 = _mm_add_epi32(vacc2x0123,
177 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
178 vacc3x0123 = _mm_add_epi32(vacc3x0123,
179 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
180 }
181 }
182 }
183
184 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
185 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
186 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
187 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
188
189 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
190 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
191 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
192 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
193 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
194
195 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
196 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
197 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
198 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
199 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
200
201 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
202 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
203 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
204 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
205
206 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
207 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
208 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
209
210
211 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
212
213 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
214
215 if (nc >= 4) {
216 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
217 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
218 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
219 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
220
221 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
222 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
223 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
224 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
225
226 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
227 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
228 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
229 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
230
231 nc -= 4;
232 } else {
233 if (nc & 2) {
234 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
235 c0 += 2;
236 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
237 c1 += 2;
238 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
239 c2 += 2;
240 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
241 c3 += 2;
242 vout = _mm_srli_epi32(vout, 16);
243 }
244 if (nc & 1) {
245 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
246 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
247 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
248 *c3 = (int8_t) _mm_extract_epi8(vout, 12);
249 }
250
251 nc = 0;
252 }
253 } while (nc != 0);
254 }
255