• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/MRx16c8-avx512skx.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 <immintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx(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)])19 void xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const int8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
30 {
31   assert(mr != 0);
32   assert(mr <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   kc = round_up_po2(kc, 8);
41   const int8_t* a0 = a;
42   int8_t* c0 = c;
43 
44   const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
45   const __mmask16 vblend_mask = _cvtu32_mask16(0xAAAA);
46   const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.multiplier));
47   const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.rounding));
48   const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
49   const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
50   const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
51   const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
52   const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
53   const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
54   do {
55     __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
56     __m512i vacc0x4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
57     __m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
58     __m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
59     w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
60 
61     size_t k = 0;
62     while (k < kc) {
63       const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
64       a0 += 8;
65 
66       const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
67 
68       vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
69       const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
70 
71       vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
72       const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
73 
74       vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
75       const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
76 
77       vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
78 
79       w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
80       k += 8 * sizeof(int8_t);
81     }
82 
83     const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
84     const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
85 
86     __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
87 
88     const __m512i vacc0x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc0x084C195D2A6E3B7F, _MM_SHUFFLE(3, 3, 1, 1));
89 
90     const __m512i vprod0x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc0x084C195D2A6E3B7F, vmultiplier), vrounding);
91 
92     const __m512i vprod0x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc0x88CC99DDAAEEBBFF, vmultiplier), vrounding);
93 
94     const __m512i vq31prod0x04152637 = _mm512_srli_epi64(vprod0x04152637, 31);
95     const __m512i vq31prod0x8C9DAEBF = _mm512_add_epi64(vprod0x8C9DAEBF, vprod0x8C9DAEBF);
96 
97     const __m512i vq31prod0x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod0x04152637, vq31prod0x8C9DAEBF);
98 
99     const __m512i vrem0x084C195D2A6E3B7F =
100       _mm512_add_epi32(_mm512_and_si512(vq31prod0x084C195D2A6E3B7F, vremainder_mask), _mm512_srai_epi32(vq31prod0x084C195D2A6E3B7F, 31));
101 
102     vacc0x084C195D2A6E3B7F = _mm512_sra_epi32(vq31prod0x084C195D2A6E3B7F, vshift);
103 
104     const __m512i vminus_one = _mm512_set1_epi32(-1);
105     vacc0x084C195D2A6E3B7F =
106       _mm512_mask_sub_epi32(vacc0x084C195D2A6E3B7F, _mm512_cmpgt_epi32_mask(vrem0x084C195D2A6E3B7F, vremainder_threshold), vacc0x084C195D2A6E3B7F, vminus_one);
107 
108     __m256i vacc0x084C2A6E195D3B7F = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0x084C195D2A6E3B7F), _mm512_extracti32x8_epi32(vacc0x084C195D2A6E3B7F, 1)), voutput_zero_point);
109     vacc0x084C2A6E195D3B7F = _mm256_min_epi16(_mm256_max_epi16(vacc0x084C2A6E195D3B7F, voutput_min), voutput_max);
110 
111     const __m128i vout0x084C2A6E195D3B7F = _mm_packs_epi16(_mm256_castsi256_si128(vacc0x084C2A6E195D3B7F), _mm256_extracti128_si256(vacc0x084C2A6E195D3B7F, 1));
112     const __m128i vout0x0123456789ABCDEF = _mm_shuffle_epi8(vout0x084C2A6E195D3B7F, _mm_set_epi8(15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0));
113 
114     if (nc >= 16) {
115       _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
116 
117       a0 = (const int8_t*) ((uintptr_t) a0 - k);
118 
119       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
120 
121       nc -= 16;
122     } else {
123       // Prepare mask for valid 8-bit elements (depends on nc).
124       const __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT32_C(1) << nc) - UINT32_C(1)));
125 
126       _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
127 
128       nc = 0;
129     }
130   } while (nc != 0);
131 }
132