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 #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_xw_minmax_ukernel_3x4c8__xop(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_xw_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__xop(
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_xw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
34 {
35   assert(mr != 0);
36   assert(mr <= 3);
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, 8);
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 
60   do {
61     __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
62     __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
63     __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
64     __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
65     __m128i vacc1x0 = vacc0x0;
66     __m128i vacc1x1 = vacc0x1;
67     __m128i vacc1x2 = vacc0x2;
68     __m128i vacc1x3 = vacc0x3;
69     __m128i vacc2x0 = vacc0x0;
70     __m128i vacc2x1 = vacc0x1;
71     __m128i vacc2x2 = vacc0x2;
72     __m128i vacc2x3 = vacc0x3;
73     w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
74 
75     size_t k = 0;
76     while (k < kc) {
77       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
78       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
79       a0 += 8;
80       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
81       const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
82       a1 += 8;
83       const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
84       const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
85       a2 += 8;
86 
87       const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
88 
89       vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
90       vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
91       vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
92       const __m128i vxb1 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 8 * sizeof(int16_t)));
93 
94       vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
95       vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
96       vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
97       const __m128i vxb2 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int16_t)));
98 
99       vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
100       vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
101       vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
102       const __m128i vxb3 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 24 * sizeof(int16_t)));
103 
104       vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
105       vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
106       vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
107 
108       w = (const void*) ((uintptr_t) w + 32 * sizeof(int16_t));
109       k += 8 * sizeof(int8_t);
110     }
111 
112     const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
113     const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
114     const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
115     const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
116     const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
117     const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
118 
119     __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
120     __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
121     __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
122 
123     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
124     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
125 
126     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
127     const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
128     const __m128i vacc2x1133 = _mm_shuffle_epi32(vacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
129 
130     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
131     const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
132     const __m128i vprod2x02 = _mm_add_epi64(_mm_mul_epi32(vacc2x0123, vmultiplier), vrounding);
133 
134     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
135     const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
136     const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding);
137 
138     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
139     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
140     const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
141     const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
142     const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31);
143     const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13);
144 
145     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
146     const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
147     const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);
148 
149     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
150     const __m128i vrem0x0123 =
151       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
152     const __m128i vrem1x0123 =
153       _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
154     const __m128i vrem2x0123 =
155       _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
156 
157     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
158     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
159     vacc0x0123 =
160       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
161     vacc1x0123 =
162       _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
163     vacc2x0123 =
164       _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
165 
166     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
167     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
168     __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
169 
170     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
171     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
172     vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
173     vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
174 
175     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
176 
177     if (nc >= 4) {
178       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
179       *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
180       *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
181 
182       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
183       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
184       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
185 
186       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
187       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
188       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
189 
190       nc -= 4;
191     } else {
192       if (nc & 2) {
193         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
194         c0 += 2;
195         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
196         c1 += 2;
197         *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
198         c2 += 2;
199         vout = _mm_srli_epi32(vout, 16);
200       }
201       if (nc & 1) {
202         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
203         *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
204         *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
205       }
206 
207       nc = 0;
208     }
209   } while (nc != 0);
210 }
211