1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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 <tmmintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64(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)])18 void xnn_qs8_gemm_minmax_ukernel_3x4c8__ssse3_ld64(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30   assert(mr != 0);
31   assert(mr <= 3);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 8);
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54 
55   do {
56     __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
57     __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
58     __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
59     __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
60     __m128i vacc1x0 = vacc0x0;
61     __m128i vacc1x1 = vacc0x1;
62     __m128i vacc1x2 = vacc0x2;
63     __m128i vacc1x3 = vacc0x3;
64     __m128i vacc2x0 = vacc0x0;
65     __m128i vacc2x1 = vacc0x1;
66     __m128i vacc2x2 = vacc0x2;
67     __m128i vacc2x3 = vacc0x3;
68     w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
69 
70     size_t k = 0;
71     while (k < kc) {
72       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
73       const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
74       a0 += 8;
75       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
76       const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
77       a1 += 8;
78       const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
79       const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
80       a2 += 8;
81 
82       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
83       const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
84 
85       vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
86       vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
87       vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
88       const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
89       const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
90 
91       vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
92       vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
93       vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
94       const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
95       const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
96 
97       vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
98       vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
99       vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
100       const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
101       const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
102 
103       vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
104       vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
105       vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
106 
107       w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
108       k += 8 * sizeof(int8_t);
109     }
110 
111     const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
112     const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
113     const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
114     const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
115     const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
116     const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
117 
118     __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
119     __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
120     __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
121 
122     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
123     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
124 
125     const __m128i vnmask0x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0x0123);
126     const __m128i vnmask1x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc1x0123);
127     const __m128i vnmask2x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc2x0123);
128 
129     const __m128i vabsacc0x0123 = _mm_abs_epi32(vacc0x0123);
130     const __m128i vabsacc1x0123 = _mm_abs_epi32(vacc1x0123);
131     const __m128i vabsacc2x0123 = _mm_abs_epi32(vacc2x0123);
132 
133     const __m128i vabsacc0x1133 = _mm_shuffle_epi32(vabsacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
134     const __m128i vabsacc1x1133 = _mm_shuffle_epi32(vabsacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
135     const __m128i vabsacc2x1133 = _mm_shuffle_epi32(vabsacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
136 
137     const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier);
138     const __m128i vabsprod1x02 = _mm_mul_epu32(vabsacc1x0123, vmultiplier);
139     const __m128i vabsprod2x02 = _mm_mul_epu32(vabsacc2x0123, vmultiplier);
140 
141     const __m128i vnmask0x02 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(2, 2, 0, 0));
142     const __m128i vnmask1x02 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(2, 2, 0, 0));
143     const __m128i vnmask2x02 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(2, 2, 0, 0));
144 
145     const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02);
146     const __m128i vprod1x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x02, vnmask1x02), vnmask1x02);
147     const __m128i vprod2x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x02, vnmask2x02), vnmask2x02);
148 
149     const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31);
150     const __m128i vq31prod1x02 = _mm_srli_epi64(_mm_add_epi64(vprod1x02, vrounding), 31);
151     const __m128i vq31prod2x02 = _mm_srli_epi64(_mm_add_epi64(vprod2x02, vrounding), 31);
152 
153     const __m128i vabsprod0x13 = _mm_mul_epu32(vabsacc0x1133, vmultiplier);
154     const __m128i vabsprod1x13 = _mm_mul_epu32(vabsacc1x1133, vmultiplier);
155     const __m128i vabsprod2x13 = _mm_mul_epu32(vabsacc2x1133, vmultiplier);
156 
157     const __m128i vnmask0x13 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(3, 3, 1, 1));
158     const __m128i vnmask1x13 = _mm_shuffle_epi32(vnmask1x0123, _MM_SHUFFLE(3, 3, 1, 1));
159     const __m128i vnmask2x13 = _mm_shuffle_epi32(vnmask2x0123, _MM_SHUFFLE(3, 3, 1, 1));
160 
161     const __m128i vprod0x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x13, vnmask0x13), vnmask0x13);
162     const __m128i vprod1x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod1x13, vnmask1x13), vnmask1x13);
163     const __m128i vprod2x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod2x13, vnmask2x13), vnmask2x13);
164 
165     const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31);
166     const __m128i vq31prod1x13 = _mm_srli_epi64(_mm_add_epi64(vprod1x13, vrounding), 31);
167     const __m128i vq31prod2x13 = _mm_srli_epi64(_mm_add_epi64(vprod2x13, vrounding), 31);
168 
169     const __m128i vq31prod0x0213 = _mm_castps_si128(_mm_shuffle_ps(
170         _mm_castsi128_ps(vq31prod0x02), _mm_castsi128_ps(vq31prod0x13), _MM_SHUFFLE(2, 0, 2, 0)));
171     const __m128i vq31prod1x0213 = _mm_castps_si128(_mm_shuffle_ps(
172         _mm_castsi128_ps(vq31prod1x02), _mm_castsi128_ps(vq31prod1x13), _MM_SHUFFLE(2, 0, 2, 0)));
173     const __m128i vq31prod2x0213 = _mm_castps_si128(_mm_shuffle_ps(
174         _mm_castsi128_ps(vq31prod2x02), _mm_castsi128_ps(vq31prod2x13), _MM_SHUFFLE(2, 0, 2, 0)));
175 
176     const __m128i vq31prod0x0123 = _mm_shuffle_epi32(vq31prod0x0213, _MM_SHUFFLE(3, 1, 2, 0));
177     const __m128i vq31prod1x0123 = _mm_shuffle_epi32(vq31prod1x0213, _MM_SHUFFLE(3, 1, 2, 0));
178     const __m128i vq31prod2x0123 = _mm_shuffle_epi32(vq31prod2x0213, _MM_SHUFFLE(3, 1, 2, 0));
179 
180     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
181     const __m128i vrem0x0123 =
182       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
183     const __m128i vrem1x0123 =
184       _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
185     const __m128i vrem2x0123 =
186       _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
187 
188     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
189     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
190     vacc0x0123 =
191       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
192     vacc1x0123 =
193       _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
194     vacc2x0123 =
195       _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
196 
197     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
198     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
199     __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
200 
201     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
202     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
203     vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
204     vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
205 
206     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
207 
208     if (nc >= 4) {
209       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
210       vout = _mm_srli_si128(vout, 4);
211       *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
212       vout = _mm_srli_si128(vout, 4);
213       *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(vout);
214 
215       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
216       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
217       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
218 
219       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
220       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
221       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
222 
223       nc -= 4;
224     } else {
225       if (nc & 2) {
226         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
227         c0 += 2;
228         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
229         c1 += 2;
230         *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
231         c2 += 2;
232         vout = _mm_srli_epi32(vout, 16);
233       }
234       if (nc & 1) {
235         *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
236         *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
237         *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
238       }
239 
240       nc = 0;
241     }
242   } while (nc != 0);
243 }
244