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_minmax_ukernel_2x4c8__xop_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)])23 void xnn_qs8_gemm_minmax_ukernel_2x4c8__xop_ld64(
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 <= 2);
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 
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     w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
64 
65     size_t k = 0;
66     while (k < kc) {
67       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
68       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
69       a0 += 8;
70       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
71       const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
72       a1 += 8;
73 
74       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
75       const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
76 
77       vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
78       vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
79       const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
80       const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
81 
82       vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
83       vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
84       const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
85       const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
86 
87       vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
88       vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
89       const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
90       const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
91 
92       vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
93       vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
94 
95       w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
96       k += 8 * sizeof(int8_t);
97     }
98 
99     const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
100     const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
101     const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
102     const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
103 
104     __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
105     __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
106 
107     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
108     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
109 
110     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
111     const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
112 
113     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
114     const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
115 
116     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
117     const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
118 
119     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
120     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
121     const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
122     const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
123 
124     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
125     const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
126 
127     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
128     const __m128i vrem0x0123 =
129       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
130     const __m128i vrem1x0123 =
131       _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
132 
133     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
134     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
135     vacc0x0123 =
136       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
137     vacc1x0123 =
138       _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
139 
140     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
141     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
142 
143     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
144     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
145     vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
146 
147     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
148 
149     if (nc >= 4) {
150       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
151       *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
152 
153       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
154       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
155 
156       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
157       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
158 
159       nc -= 4;
160     } else {
161       if (nc & 2) {
162         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
163         c0 += 2;
164         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
165         c1 += 2;
166         vout = _mm_srli_epi32(vout, 16);
167       }
168       if (nc & 1) {
169         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
170         *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
171       }
172 
173       nc = 0;
174     }
175   } while (nc != 0);
176 }
177