1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx4c8-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 <emmintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32 assert(mr != 0);
33 assert(mr <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8);
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53
54 do {
55 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
56 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
57 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
58 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
59 __m128i vacc1x0 = vacc0x0;
60 __m128i vacc1x1 = vacc0x1;
61 __m128i vacc1x2 = vacc0x2;
62 __m128i vacc1x3 = vacc0x3;
63 __m128i vacc2x0 = vacc0x0;
64 __m128i vacc2x1 = vacc0x1;
65 __m128i vacc2x2 = vacc0x2;
66 __m128i vacc2x3 = vacc0x3;
67 w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
68
69 size_t p = ks;
70 do {
71 const int8_t* restrict a0 = a[0];
72 if XNN_UNPREDICTABLE(a0 != zero) {
73 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74 }
75 const int8_t* restrict a1 = a[1];
76 if XNN_UNPREDICTABLE(a1 != zero) {
77 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78 }
79 const int8_t* restrict a2 = a[2];
80 if XNN_UNPREDICTABLE(a2 != zero) {
81 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
82 }
83 a += 3;
84
85 size_t k = 0;
86 while (k < kc) {
87 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
88 const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
89 a0 += 8;
90 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
91 const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
92 a1 += 8;
93 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
94 const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
95 a2 += 8;
96
97 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
98 const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
99
100 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
101 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
102 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
103 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
104 const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
105
106 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
107 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
108 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
109 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
110 const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
111
112 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
113 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
114 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
115 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
116 const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
117
118 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
119 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
120 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
121
122 w = (const void*) ((uintptr_t) w + 32);
123 k += 8 * sizeof(int8_t);
124 }
125 p -= 3 * sizeof(void*);
126 } while (p != 0);
127
128 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
129 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
130 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
131 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
132 const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
133 const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
134
135 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
136 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
137 __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
138
139 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
140 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
141
142 const __m128i vnmask0x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0x0123);
143 const __m128i vnmask1x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc1x0123);
144 const __m128i vnmask2x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc2x0123);
145
146 const __m128i vabsacc0x0123 = _mm_sub_epi32(_mm_xor_si128(vacc0x0123, vnmask0x0123), vnmask0x0123);
147 const __m128i vabsacc1x0123 = _mm_sub_epi32(_mm_xor_si128(vacc1x0123, vnmask1x0123), vnmask1x0123);
148 const __m128i vabsacc2x0123 = _mm_sub_epi32(_mm_xor_si128(vacc2x0123, vnmask2x0123), vnmask2x0123);
149
150 const __m128i vabsacc0x1133 = _mm_shuffle_epi32(vabsacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
151 const __m128i vabsacc1x1133 = _mm_shuffle_epi32(vabsacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
152 const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
153
154 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier);
155 const __m128i vabsprod1x02 = _mm_mul_epu32(vabsacc1x0123, vmultiplier);
156 const __m128i vabsprod2x02 = _mm_mul_epu32(vabsacc2x0123, vmultiplier);
157
158 const __m128i vnmask0x02 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(2, 2, 0, 0));
159 const __m128i vnmask1x02 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(2, 2, 0, 0));
160 const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0));
161
162 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02);
163 const __m128i vprod1x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x02, vnmask1x02), vnmask1x02);
164 const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02);
165
166 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31);
167 const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31);
168 const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31);
169
170 const __m128i vabsprod0x13 = _mm_mul_epu32(vabsacc0x1133, vmultiplier);
171 const __m128i vabsprod1x13 = _mm_mul_epu32(vabsacc1x1133, vmultiplier);
172 const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier);
173
174 const __m128i vnmask0x13 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(3, 3, 1, 1));
175 const __m128i vnmask1x13 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(3, 3, 1, 1));
176 const __m128i vnmask2x13 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(3, 3, 1, 1));
177
178 const __m128i vprod0x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x13, vnmask0x13), vnmask0x13);
179 const __m128i vprod1x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x13, vnmask1x13), vnmask1x13);
180 const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13);
181
182 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31);
183 const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31);
184 const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31);
185
186 const __m128i vq31prod0x0213 = _mm_castps_si128(_mm_shuffle_ps(
187 _mm_castsi128_ps(vq31prod0x02), _mm_castsi128_ps(vq31prod0x13), _MM_SHUFFLE(2, 0, 2, 0)));
188 const __m128i vq31prod1x0213 = _mm_castps_si128(_mm_shuffle_ps(
189 _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0)));
190 const __m128i vq31prod2x0213 = _mm_castps_si128(_mm_shuffle_ps(
191 _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0)));
192
193 const __m128i vq31prod0x0123 = _mm_shuffle_epi32(vq31prod0x0213, _MM_SHUFFLE(3, 1, 2, 0));
194 const __m128i vq31prod1x0123 = _mm_shuffle_epi32(vq31prod1x0213, _MM_SHUFFLE(3, 1, 2, 0));
195 const __m128i vq31prod2x0123 = _mm_shuffle_epi32(vq31prod2x0213, _MM_SHUFFLE(3, 1, 2, 0));
196
197 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
198 const __m128i vrem0x0123 =
199 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
200 const __m128i vrem1x0123 =
201 _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
202 const __m128i vrem2x0123 =
203 _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
204
205 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
206 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
207 vacc0x0123 =
208 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
209 vacc1x0123 =
210 _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
211 vacc2x0123 =
212 _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
213
214 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
215 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
216 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
217
218 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
219 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
220 vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
221 vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
222
223 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
224
225 if (nc >= 4) {
226 *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(2, 2, 2, 2)));
227 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
228 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1)));
229 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
230 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
231 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
232
233 a = (const int8_t**restrict) ((uintptr_t) a - ks);
234
235 nc -= 4;
236 } else {
237 if (nc & 2) {
238 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
239 c2 += 2;
240 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
241 c1 += 2;
242 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
243 c0 += 2;
244 vout = _mm_srli_epi32(vout, 16);
245 }
246 if (nc & 1) {
247 *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
248 *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
249 *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
250 }
251
252 nc = 0;
253 }
254 } while (nc != 0);
255 }
256