1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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/igemm.h>
20 #include <xnnpack/math.h>
21 
22 
xnn_qs8_igemm_minmax_ukernel_4x4c2__xop_ld128(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)])23 void xnn_qs8_igemm_minmax_ukernel_4x4c2__xop_ld128(
24     size_t mr,
25     size_t nc,
26     size_t kc,
27     size_t ks,
28     const int8_t** restrict a,
29     const void* restrict w,
30     int8_t* restrict c,
31     size_t cm_stride,
32     size_t cn_stride,
33     size_t a_offset,
34     const int8_t* zero,
35     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
36 {
37   assert(mr != 0);
38   assert(mr <= 4);
39   assert(nc != 0);
40   assert(kc != 0);
41   assert(ks != 0);
42   assert(ks % (4 * sizeof(void*)) == 0);
43   assert(a_offset % sizeof(int8_t) == 0);
44   assert(a != NULL);
45   assert(w != NULL);
46   assert(c != NULL);
47 
48   kc = round_up_po2(kc, 2);
49   int8_t* c0 = c;
50   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
51   if XNN_UNPREDICTABLE(mr < 2) {
52     c1 = c0;
53   }
54   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
55   if XNN_UNPREDICTABLE(mr <= 2) {
56     c2 = c1;
57   }
58   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
59   if XNN_UNPREDICTABLE(mr != 4) {
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*) ((uintptr_t) w + 4 * sizeof(int32_t));
69 
70     size_t p = ks;
71     do {
72       const int8_t* restrict a0 = a[0];
73       if XNN_UNPREDICTABLE(a0 != zero) {
74         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
75       }
76       const int8_t* restrict a1 = a[1];
77       if XNN_UNPREDICTABLE(a1 != zero) {
78         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
79       }
80       const int8_t* restrict a2 = a[2];
81       if XNN_UNPREDICTABLE(a2 != zero) {
82         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
83       }
84       const int8_t* restrict a3 = a[3];
85       if XNN_UNPREDICTABLE(a3 != zero) {
86         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
87       }
88       a += 4;
89 
90       size_t k = kc;
91       while (k >= 8 * sizeof(int8_t)) {
92         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
93         const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
94         a0 += 8;
95         const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
96         const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
97         a1 += 8;
98         const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
99         const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
100         a2 += 8;
101         const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
102         const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
103         a3 += 8;
104 
105         const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
106         const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
107         const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
108         const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
109 
110         vacc0x0123 = _mm_maddd_epi16(
111           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
112         vacc1x0123 = _mm_maddd_epi16(
113           _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
114         vacc2x0123 = _mm_maddd_epi16(
115           _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
116         vacc3x0123 = _mm_maddd_epi16(
117           _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
118 
119         vacc0x0123 = _mm_maddd_epi16(
120           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
121         vacc1x0123 = _mm_maddd_epi16(
122           _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
123         vacc2x0123 = _mm_maddd_epi16(
124           _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
125         vacc3x0123 = _mm_maddd_epi16(
126           _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
127         const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16));
128         const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
129         const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
130         const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
131 
132         vacc0x0123 = _mm_maddd_epi16(
133           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
134         vacc1x0123 = _mm_maddd_epi16(
135           _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
136         vacc2x0123 = _mm_maddd_epi16(
137           _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
138         vacc3x0123 = _mm_maddd_epi16(
139           _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
140 
141         vacc0x0123 = _mm_maddd_epi16(
142           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
143         vacc1x0123 = _mm_maddd_epi16(
144           _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
145         vacc2x0123 = _mm_maddd_epi16(
146           _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
147         vacc3x0123 = _mm_maddd_epi16(
148           _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
149 
150         w = (const void*) ((uintptr_t) w + 32);
151         k -= 8 * sizeof(int8_t);
152       }
153       if (k != 0) {
154         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
155         const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
156         a0 = (const int8_t*) ((uintptr_t) a0 + k);
157         const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
158         const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
159         a1 = (const int8_t*) ((uintptr_t) a1 + k);
160         const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
161         const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
162         a2 = (const int8_t*) ((uintptr_t) a2 + k);
163         const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
164         const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
165         a3 = (const int8_t*) ((uintptr_t) a3 + k);
166 
167         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
168         w = (const void*) ((uintptr_t) w + 8);
169         const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
170 
171         vacc0x0123 = _mm_maddd_epi16(
172           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
173         vacc1x0123 = _mm_maddd_epi16(
174           _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
175         vacc2x0123 = _mm_maddd_epi16(
176           _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
177         vacc3x0123 = _mm_maddd_epi16(
178           _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
179 
180         if (k > 2 * sizeof(int8_t)) {
181           const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
182           w = (const void*) ((uintptr_t) w + 8);
183           const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
184 
185           vacc0x0123 = _mm_maddd_epi16(
186             _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
187           vacc1x0123 = _mm_maddd_epi16(
188             _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
189           vacc2x0123 = _mm_maddd_epi16(
190             _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
191           vacc3x0123 = _mm_maddd_epi16(
192             _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
193 
194           if (k > 4 * sizeof(int8_t)) {
195             const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
196             w = (const void*) ((uintptr_t) w + 8);
197             const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
198 
199             vacc0x0123 = _mm_maddd_epi16(
200               _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
201             vacc1x0123 = _mm_maddd_epi16(
202               _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
203             vacc2x0123 = _mm_maddd_epi16(
204               _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
205             vacc3x0123 = _mm_maddd_epi16(
206               _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
207           }
208         }
209       }
210       p -= 4 * sizeof(void*);
211     } while (p != 0);
212 
213     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
214     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
215 
216     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
217     const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
218     const __m128i vacc2x1133 = _mm_shuffle_epi32(vacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
219     const __m128i vacc3x1133 = _mm_shuffle_epi32(vacc3x0123, _MM_SHUFFLE(3, 3, 1, 1));
220 
221     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
222     const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
223     const __m128i vprod2x02 = _mm_add_epi64(_mm_mul_epi32(vacc2x0123, vmultiplier), vrounding);
224     const __m128i vprod3x02 = _mm_add_epi64(_mm_mul_epi32(vacc3x0123, vmultiplier), vrounding);
225 
226     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
227     const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
228     const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding);
229     const __m128i vprod3x13 = _mm_add_epi64(_mm_mul_epi32(vacc3x1133, vmultiplier), vrounding);
230 
231     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
232     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
233     const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
234     const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
235     const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31);
236     const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13);
237     const __m128i vq31prod3x02 = _mm_srli_epi64(vprod3x02, 31);
238     const __m128i vq31prod3x13 = _mm_add_epi64(vprod3x13, vprod3x13);
239 
240     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
241     const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
242     const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);
243     const __m128i vq31prod3x0123 = _mm_blend_epi16(vq31prod3x02, vq31prod3x13, 0xCC);
244 
245     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
246     const __m128i vrem0x0123 =
247       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
248     const __m128i vrem1x0123 =
249       _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
250     const __m128i vrem2x0123 =
251       _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
252     const __m128i vrem3x0123 =
253       _mm_add_epi32(_mm_and_si128(vq31prod3x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod3x0123));
254 
255     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
256     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
257     vacc0x0123 =
258       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
259     vacc1x0123 =
260       _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
261     vacc2x0123 =
262       _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
263     vacc3x0123 =
264       _mm_sub_epi32(_mm_sra_epi32(vq31prod3x0123, vshift), _mm_cmpgt_epi32(vrem3x0123, vremainder_threshold));
265 
266     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
267     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
268     __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
269 
270     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
271     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
272     vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
273     vacc23x0123 = _mm_min_epi16(_mm_max_epi16(vacc23x0123, voutput_min), voutput_max);
274 
275     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
276 
277     if (nc >= 4) {
278       *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
279       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
280       *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
281       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
282       *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
283       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
284       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
285       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
286 
287       a = (const int8_t**restrict) ((uintptr_t) a - ks);
288 
289       nc -= 4;
290     } else {
291       if (nc & 2) {
292         *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
293         c3 += 2;
294         *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
295         c2 += 2;
296         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
297         c1 += 2;
298         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
299         c0 += 2;
300         vout = _mm_srli_epi32(vout, 16);
301       }
302       if (nc & 1) {
303         *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
304         *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
305         *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
306         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
307       }
308 
309       nc = 0;
310     }
311   } while (nc != 0);
312 }
313