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_qu8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])25 void xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128(
26 size_t mr,
27 size_t nc,
28 size_t kc,
29 const uint8_t* restrict a,
30 size_t a_stride,
31 const void* restrict w,
32 uint8_t* restrict c,
33 size_t cm_stride,
34 size_t cn_stride,
35 const union xnn_qu8_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(uint8_t) == 0);
42 assert(a != NULL);
43 assert(w != NULL);
44 assert(c != NULL);
45
46 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
47 const uint8_t* a0 = a;
48 uint8_t* c0 = c;
49 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
50 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
51 if XNN_UNPREDICTABLE(mr < 2) {
52 a1 = a0;
53 c1 = c0;
54 }
55 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
56 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
57 if XNN_UNPREDICTABLE(mr <= 2) {
58 a2 = a1;
59 c2 = c1;
60 }
61 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
62 uint8_t* c3 = (uint8_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 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
77 const __m128i vzero = _mm_setzero_si128();
78 while (k >= 8 * sizeof(uint8_t)) {
79 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
80 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
81 a0 += 8;
82 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
83 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
84 a1 += 8;
85 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
86 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
87 a2 += 8;
88 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
89 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
90 a3 += 8;
91
92 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
93 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
94 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
95
96 vacc0x0123 = _mm_maddd_epi16(
97 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
98 vacc1x0123 = _mm_maddd_epi16(
99 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
100 vacc2x0123 = _mm_maddd_epi16(
101 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
102 vacc3x0123 = _mm_maddd_epi16(
103 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
104
105 vacc0x0123 = _mm_maddd_epi16(
106 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
107 vacc1x0123 = _mm_maddd_epi16(
108 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
109 vacc2x0123 = _mm_maddd_epi16(
110 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
111 vacc3x0123 = _mm_maddd_epi16(
112 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
113 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
114 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
115 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
116
117 vacc0x0123 = _mm_maddd_epi16(
118 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
119 vacc1x0123 = _mm_maddd_epi16(
120 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
121 vacc2x0123 = _mm_maddd_epi16(
122 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
123 vacc3x0123 = _mm_maddd_epi16(
124 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
125
126 vacc0x0123 = _mm_maddd_epi16(
127 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
128 vacc1x0123 = _mm_maddd_epi16(
129 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
130 vacc2x0123 = _mm_maddd_epi16(
131 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
132 vacc3x0123 = _mm_maddd_epi16(
133 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
134
135 w = (const void*) ((const uint8_t*) w + 32);
136 k -= 8 * sizeof(uint8_t);
137 }
138 if (k != 0) {
139 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
140 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
141 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
142 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
143 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
144 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
145 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
146 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
147 a2 = (const uint8_t*) ((uintptr_t) a2 + k);
148 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
149 const __m128i vxa3 = _mm_cvtepu8_epi16(va3);
150 a3 = (const uint8_t*) ((uintptr_t) a3 + k);
151
152 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
153 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
154 w = (const void*) ((const uint8_t*) w + 8);
155
156 vacc0x0123 = _mm_maddd_epi16(
157 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
158 vacc1x0123 = _mm_maddd_epi16(
159 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
160 vacc2x0123 = _mm_maddd_epi16(
161 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
162 vacc3x0123 = _mm_maddd_epi16(
163 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
164
165 if (k > 2 * sizeof(uint8_t)) {
166 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
167 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
168 w = (const void*) ((const uint8_t*) w + 8);
169
170 vacc0x0123 = _mm_maddd_epi16(
171 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
172 vacc1x0123 = _mm_maddd_epi16(
173 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
174 vacc2x0123 = _mm_maddd_epi16(
175 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
176 vacc3x0123 = _mm_maddd_epi16(
177 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
178
179 if (k > 4 * sizeof(uint8_t)) {
180 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
181 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
182 w = (const void*) ((const uint8_t*) w + 8);
183
184 vacc0x0123 = _mm_maddd_epi16(
185 _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
186 vacc1x0123 = _mm_maddd_epi16(
187 _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
188 vacc2x0123 = _mm_maddd_epi16(
189 _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
190 vacc3x0123 = _mm_maddd_epi16(
191 _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
192 }
193 }
194 }
195
196 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
197 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
198 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
199 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
200
201 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
202 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
203 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
204 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
205 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
206
207 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
208 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
209 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
210 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
211 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
212
213 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
214 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
215 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
216 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
217
218 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
219 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
220 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
221
222 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
223
224 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
225
226 if (nc >= 4) {
227 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
228 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
229 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
230 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
231
232 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
233 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
234 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
235 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
236
237 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
238 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
239 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
240 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
241
242 nc -= 4;
243 } else {
244 if (nc & 2) {
245 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
246 c0 += 2;
247 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
248 c1 += 2;
249 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
250 c2 += 2;
251 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
252 c3 += 2;
253 vout = _mm_srli_epi32(vout, 16);
254 }
255 if (nc & 1) {
256 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
257 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
258 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
259 *c3 = (uint8_t) _mm_extract_epi8(vout, 12);
260 }
261
262 nc = 0;
263 }
264 } while (nc != 0);
265 }
266