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 #ifdef __GNUC__
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 
22 
xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_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_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_gemm_minmax_ukernel_4x4c2__xop_ld128(
24     size_t mr,
25     size_t nc,
26     size_t kc,
27     const int8_t* restrict a,
28     size_t a_stride,
29     const void* restrict w,
30     int8_t* restrict c,
31     size_t cm_stride,
32     size_t cn_stride,
33     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
34 {
35   assert(mr != 0);
36   assert(mr <= 4);
37   assert(nc != 0);
38   assert(kc != 0);
39   assert(kc % sizeof(int8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   kc = round_up_po2(kc, 2);
45   const int8_t* a0 = a;
46   int8_t* c0 = c;
47   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
49   if XNN_UNPREDICTABLE(mr < 2) {
50     a1 = a0;
51     c1 = c0;
52   }
53   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
55   if XNN_UNPREDICTABLE(mr <= 2) {
56     a2 = a1;
57     c2 = c1;
58   }
59   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
61   if XNN_UNPREDICTABLE(mr != 4) {
62     a3 = a2;
63     c3 = c2;
64   }
65 
66   do {
67     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
68     __m128i vacc1x0123 = vacc0x0123;
69     __m128i vacc2x0123 = vacc0x0123;
70     __m128i vacc3x0123 = vacc0x0123;
71     w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
72 
73     size_t k = kc;
74     while (k >= 8 * sizeof(int8_t)) {
75       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
76       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
77       a0 += 8;
78       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
79       const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
80       a1 += 8;
81       const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
82       const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
83       a2 += 8;
84       const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
85       const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
86       a3 += 8;
87 
88       const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
89       const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
90       const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
91       const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
92 
93       vacc0x0123 = _mm_maddd_epi16(
94         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
95       vacc1x0123 = _mm_maddd_epi16(
96         _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
97       vacc2x0123 = _mm_maddd_epi16(
98         _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
99       vacc3x0123 = _mm_maddd_epi16(
100         _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
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 vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
111       const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
112       const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
113       const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
114 
115       vacc0x0123 = _mm_maddd_epi16(
116         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
117       vacc1x0123 = _mm_maddd_epi16(
118         _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
119       vacc2x0123 = _mm_maddd_epi16(
120         _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
121       vacc3x0123 = _mm_maddd_epi16(
122         _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
123 
124       vacc0x0123 = _mm_maddd_epi16(
125         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
126       vacc1x0123 = _mm_maddd_epi16(
127         _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
128       vacc2x0123 = _mm_maddd_epi16(
129         _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
130       vacc3x0123 = _mm_maddd_epi16(
131         _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
132 
133       w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
134       k -= 8 * sizeof(int8_t);
135     }
136     if (k != 0) {
137       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
138       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
139       a0 = (const int8_t*) ((uintptr_t) a0 + k);
140       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
141       const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
142       a1 = (const int8_t*) ((uintptr_t) a1 + k);
143       const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
144       const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
145       a2 = (const int8_t*) ((uintptr_t) a2 + k);
146       const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
147       const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
148       a3 = (const int8_t*) ((uintptr_t) a3 + k);
149 
150       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
151       const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
152       w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
153 
154       vacc0x0123 = _mm_maddd_epi16(
155         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
156       vacc1x0123 = _mm_maddd_epi16(
157         _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
158       vacc2x0123 = _mm_maddd_epi16(
159         _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
160       vacc3x0123 = _mm_maddd_epi16(
161         _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
162 
163       if (k > 2 * sizeof(int8_t)) {
164         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
165         const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
166         w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 
168         vacc0x0123 = _mm_maddd_epi16(
169           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
170         vacc1x0123 = _mm_maddd_epi16(
171           _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
172         vacc2x0123 = _mm_maddd_epi16(
173           _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
174         vacc3x0123 = _mm_maddd_epi16(
175           _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
176 
177         if (k > 4 * sizeof(int8_t)) {
178           const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
179           const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
180           w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181 
182           vacc0x0123 = _mm_maddd_epi16(
183             _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
184           vacc1x0123 = _mm_maddd_epi16(
185             _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
186           vacc2x0123 = _mm_maddd_epi16(
187             _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
188           vacc3x0123 = _mm_maddd_epi16(
189             _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
190         }
191       }
192     }
193 
194     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
195     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
196 
197     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
198     const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
199     const __m128i vacc2x1133 = _mm_shuffle_epi32(vacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
200     const __m128i vacc3x1133 = _mm_shuffle_epi32(vacc3x0123, _MM_SHUFFLE(3, 3, 1, 1));
201 
202     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
203     const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
204     const __m128i vprod2x02 = _mm_add_epi64(_mm_mul_epi32(vacc2x0123, vmultiplier), vrounding);
205     const __m128i vprod3x02 = _mm_add_epi64(_mm_mul_epi32(vacc3x0123, vmultiplier), vrounding);
206 
207     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
208     const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
209     const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding);
210     const __m128i vprod3x13 = _mm_add_epi64(_mm_mul_epi32(vacc3x1133, vmultiplier), vrounding);
211 
212     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
213     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
214     const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
215     const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
216     const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31);
217     const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13);
218     const __m128i vq31prod3x02 = _mm_srli_epi64(vprod3x02, 31);
219     const __m128i vq31prod3x13 = _mm_add_epi64(vprod3x13, vprod3x13);
220 
221     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
222     const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
223     const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);
224     const __m128i vq31prod3x0123 = _mm_blend_epi16(vq31prod3x02, vq31prod3x13, 0xCC);
225 
226     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
227     const __m128i vrem0x0123 =
228       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
229     const __m128i vrem1x0123 =
230       _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
231     const __m128i vrem2x0123 =
232       _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
233     const __m128i vrem3x0123 =
234       _mm_add_epi32(_mm_and_si128(vq31prod3x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod3x0123));
235 
236     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
237     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
238     vacc0x0123 =
239       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
240     vacc1x0123 =
241       _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
242     vacc2x0123 =
243       _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
244     vacc3x0123 =
245       _mm_sub_epi32(_mm_sra_epi32(vq31prod3x0123, vshift), _mm_cmpgt_epi32(vrem3x0123, vremainder_threshold));
246 
247     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
248     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
249     __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
250 
251     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
252     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
253     vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
254     vacc23x0123 = _mm_min_epi16(_mm_max_epi16(vacc23x0123, voutput_min), voutput_max);
255 
256     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
257 
258     if (nc >= 4) {
259       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
260       *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
261       *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
262       *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
263 
264       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
265       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
266       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
267       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
268 
269       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
270       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
271       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
272       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
273 
274       nc -= 4;
275     } else {
276       if (nc & 2) {
277         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
278         c0 += 2;
279         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
280         c1 += 2;
281         *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
282         c2 += 2;
283         *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
284         c3 += 2;
285         vout = _mm_srli_epi32(vout, 16);
286       }
287       if (nc & 1) {
288         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
289         *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
290         *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
291         *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
292       }
293 
294       nc = 0;
295     }
296   } while (nc != 0);
297 }
298