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_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(
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_qc8_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 vb01 = _mm_loadu_si128((const __m128i*) w);
86 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
87 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
88
89 vacc0x0123 = _mm_add_epi32(vacc0x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
91 vacc1x0123 = _mm_add_epi32(vacc1x0123,
92 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
93 vacc2x0123 = _mm_add_epi32(vacc2x0123,
94 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
95 vacc3x0123 = _mm_add_epi32(vacc3x0123,
96 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
97
98 vacc0x0123 = _mm_add_epi32(vacc0x0123,
99 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
100 vacc1x0123 = _mm_add_epi32(vacc1x0123,
101 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
102 vacc2x0123 = _mm_add_epi32(vacc2x0123,
103 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
104 vacc3x0123 = _mm_add_epi32(vacc3x0123,
105 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
106 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
107 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
108 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
109
110 vacc0x0123 = _mm_add_epi32(vacc0x0123,
111 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
112 vacc1x0123 = _mm_add_epi32(vacc1x0123,
113 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
114 vacc2x0123 = _mm_add_epi32(vacc2x0123,
115 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
116 vacc3x0123 = _mm_add_epi32(vacc3x0123,
117 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
118
119 vacc0x0123 = _mm_add_epi32(vacc0x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
121 vacc1x0123 = _mm_add_epi32(vacc1x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
123 vacc2x0123 = _mm_add_epi32(vacc2x0123,
124 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
125 vacc3x0123 = _mm_add_epi32(vacc3x0123,
126 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
127
128 w = (const void*) ((const int8_t*) w + 32);
129 k -= 8 * sizeof(int8_t);
130 }
131 if (k != 0) {
132 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
133 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
134 a0 = (const int8_t*) ((uintptr_t) a0 + k);
135 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
136 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
137 a1 = (const int8_t*) ((uintptr_t) a1 + k);
138 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
139 const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
140 a2 = (const int8_t*) ((uintptr_t) a2 + k);
141 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
142 const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
143 a3 = (const int8_t*) ((uintptr_t) a3 + k);
144
145 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
146 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
147 w = (const void*) ((const int8_t*) w + 8);
148
149 vacc0x0123 = _mm_add_epi32(vacc0x0123,
150 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
151 vacc1x0123 = _mm_add_epi32(vacc1x0123,
152 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
153 vacc2x0123 = _mm_add_epi32(vacc2x0123,
154 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
155 vacc3x0123 = _mm_add_epi32(vacc3x0123,
156 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
157
158 if (k > 2 * sizeof(int8_t)) {
159 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
160 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
161 w = (const void*) ((const int8_t*) w + 8);
162
163 vacc0x0123 = _mm_add_epi32(vacc0x0123,
164 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
165 vacc1x0123 = _mm_add_epi32(vacc1x0123,
166 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
167 vacc2x0123 = _mm_add_epi32(vacc2x0123,
168 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
169 vacc3x0123 = _mm_add_epi32(vacc3x0123,
170 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
171
172 if (k > 4 * sizeof(int8_t)) {
173 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
174 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
175 w = (const void*) ((const int8_t*) w + 8);
176
177 vacc0x0123 = _mm_add_epi32(vacc0x0123,
178 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
179 vacc1x0123 = _mm_add_epi32(vacc1x0123,
180 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
181 vacc2x0123 = _mm_add_epi32(vacc2x0123,
182 _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
183 vacc3x0123 = _mm_add_epi32(vacc3x0123,
184 _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
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 vscale0123 = _mm_loadu_ps((const float*) w);
195 w = (const void*) ((const float*) w + 4);
196 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
197 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
198 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
199 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
200
201 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
202 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
203 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
204 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
205 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
206
207 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
208 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
209 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
210 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
211
212 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
213 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
214 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
215
216
217 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
218
219 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
220
221 if (nc >= 4) {
222 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
223 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
224 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
225 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
226
227 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
228 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
229 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
230 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
231
232 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
233 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
234 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
235 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
236
237 nc -= 4;
238 } else {
239 if (nc & 2) {
240 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
241 c0 += 2;
242 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
243 c1 += 2;
244 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
245 c2 += 2;
246 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
247 c3 += 2;
248 vout = _mm_srli_epi32(vout, 16);
249 }
250 if (nc & 1) {
251 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
252 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
253 *c2 = (int8_t) _mm_extract_epi8(vout, 8);
254 *c3 = (int8_t) _mm_extract_epi8(vout, 12);
255 }
256
257 nc = 0;
258 }
259 } while (nc != 0);
260 }
261