• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_1x4c8__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_1x4c8__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 <= 1);
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 
48   do {
49     __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
50     __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
51     __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
52     __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
53     w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
54 
55     size_t k = 0;
56     while (k < kc) {
57       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
58       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
59       a0 += 8;
60 
61       const __m128i vb01 = _mm_load_si128((const __m128i*) w);
62       const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
63       const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
64       const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
65 
66       vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
67       vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
68       const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
69       const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
70       const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
71       const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
72 
73       vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
74       vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
75 
76       w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
77       k += 8 * sizeof(int8_t);
78     }
79 
80     const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
81     const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
82 
83     __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
84 
85     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
86     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
87 
88     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
89 
90     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
91 
92     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
93 
94     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
95     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
96 
97     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
98 
99     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
100     const __m128i vrem0x0123 =
101       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
102 
103     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
104     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
105     vacc0x0123 =
106       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
107 
108     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
109     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
110 
111     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
112     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
113     vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
114 
115     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
116 
117     if (nc >= 4) {
118       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
119 
120       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
121 
122       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
123 
124       nc -= 4;
125     } else {
126       if (nc & 2) {
127         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
128         c0 += 2;
129         vout = _mm_srli_epi32(vout, 16);
130       }
131       if (nc & 1) {
132         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
133       }
134 
135       nc = 0;
136     }
137   } while (nc != 0);
138 }
139