1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
32 {
33 assert(mr != 0);
34 assert(mr <= 2);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8);
43 int8_t* c0 = c;
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 c1 = c0;
47 }
48
49 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
50 const __mmask16 vblend_mask = _cvtu32_mask16(0xAAAA);
51 const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.multiplier));
52 const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.rounding));
53 const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
54 const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
55 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
56 const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
57 const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_min));
58 const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.output_max));
59 do {
60 __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
61 __m512i vacc0x4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
62 __m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
63 __m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
64 __m512i vacc1x0123 = vacc0x0123;
65 __m512i vacc1x4567 = vacc0x4567;
66 __m512i vacc1x89AB = vacc0x89AB;
67 __m512i vacc1xCDEF = vacc0xCDEF;
68 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
69
70 size_t p = ks;
71 do {
72 const int8_t* restrict a0 = a[0];
73 if XNN_UNPREDICTABLE(a0 != zero) {
74 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
75 }
76 const int8_t* restrict a1 = a[1];
77 if XNN_UNPREDICTABLE(a1 != zero) {
78 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
79 }
80 a += 2;
81
82 size_t k = 0;
83 while (k < kc) {
84 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
85 a0 += 8;
86 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
87 a1 += 8;
88
89 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
90
91 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
92 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
93 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
94
95 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
96 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
97 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
98
99 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
100 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
101 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
102
103 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
104 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
105
106 w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
107 k += 8 * sizeof(int8_t);
108 }
109 p -= 2 * sizeof(void*);
110 } while (p != 0);
111
112 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
113 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
114 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
115 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
116
117 __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
118 __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
119
120 const __m512i vacc0x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc0x084C195D2A6E3B7F, _MM_SHUFFLE(3, 3, 1, 1));
121 const __m512i vacc1x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc1x084C195D2A6E3B7F, _MM_SHUFFLE(3, 3, 1, 1));
122
123 const __m512i vprod0x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc0x084C195D2A6E3B7F, vmultiplier), vrounding);
124 const __m512i vprod1x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc1x084C195D2A6E3B7F, vmultiplier), vrounding);
125
126 const __m512i vprod0x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc0x88CC99DDAAEEBBFF, vmultiplier), vrounding);
127 const __m512i vprod1x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc1x88CC99DDAAEEBBFF, vmultiplier), vrounding);
128
129 const __m512i vq31prod0x04152637 = _mm512_srli_epi64(vprod0x04152637, 31);
130 const __m512i vq31prod0x8C9DAEBF = _mm512_add_epi64(vprod0x8C9DAEBF, vprod0x8C9DAEBF);
131 const __m512i vq31prod1x04152637 = _mm512_srli_epi64(vprod1x04152637, 31);
132 const __m512i vq31prod1x8C9DAEBF = _mm512_add_epi64(vprod1x8C9DAEBF, vprod1x8C9DAEBF);
133
134 const __m512i vq31prod0x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod0x04152637, vq31prod0x8C9DAEBF);
135 const __m512i vq31prod1x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod1x04152637, vq31prod1x8C9DAEBF);
136
137 const __m512i vrem0x084C195D2A6E3B7F =
138 _mm512_add_epi32(_mm512_and_si512(vq31prod0x084C195D2A6E3B7F, vremainder_mask), _mm512_srai_epi32(vq31prod0x084C195D2A6E3B7F, 31));
139 const __m512i vrem1x084C195D2A6E3B7F =
140 _mm512_add_epi32(_mm512_and_si512(vq31prod1x084C195D2A6E3B7F, vremainder_mask), _mm512_srai_epi32(vq31prod1x084C195D2A6E3B7F, 31));
141
142 vacc0x084C195D2A6E3B7F = _mm512_sra_epi32(vq31prod0x084C195D2A6E3B7F, vshift);
143 vacc1x084C195D2A6E3B7F = _mm512_sra_epi32(vq31prod1x084C195D2A6E3B7F, vshift);
144
145 const __m512i vminus_one = _mm512_set1_epi32(-1);
146 vacc0x084C195D2A6E3B7F =
147 _mm512_mask_sub_epi32(vacc0x084C195D2A6E3B7F, _mm512_cmpgt_epi32_mask(vrem0x084C195D2A6E3B7F, vremainder_threshold), vacc0x084C195D2A6E3B7F, vminus_one);
148 vacc1x084C195D2A6E3B7F =
149 _mm512_mask_sub_epi32(vacc1x084C195D2A6E3B7F, _mm512_cmpgt_epi32_mask(vrem1x084C195D2A6E3B7F, vremainder_threshold), vacc1x084C195D2A6E3B7F, vminus_one);
150
151 __m512i vacc01x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_zero_point);
152
153 vacc01x084Cx195Dx2A6Ex3B7F = _mm512_min_epi16(_mm512_max_epi16(vacc01x084Cx195Dx2A6Ex3B7F, voutput_min), voutput_max);
154
155 const __m256i vout01x084Cx2A6Ex195Dx3B7F = _mm256_packs_epi16(_mm512_castsi512_si256(vacc01x084Cx195Dx2A6Ex3B7F), _mm512_extracti32x8_epi32(vacc01x084Cx195Dx2A6Ex3B7F, 1));
156 const __m256i vout01x084C2A6E195D3B7F = _mm256_permutevar8x32_epi32(vout01x084Cx2A6Ex195Dx3B7F, _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0));
157 const __m256i vout01x0123456789ABCDEF = _mm256_shuffle_epi8(vout01x084C2A6E195D3B7F, _mm256_set_epi8(15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0, 15, 7, 11, 3, 13, 5, 9, 1, 14, 6, 10, 2, 12, 4, 8, 0));
158
159 if (nc >= 16) {
160 _mm_storeu_si128((__m128i*) c1, _mm256_extracti128_si256(vout01x0123456789ABCDEF, 1));
161 _mm_storeu_si128((__m128i*) c0, _mm256_castsi256_si128(vout01x0123456789ABCDEF));
162
163 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
164 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
165
166 a = (const int8_t**restrict) ((uintptr_t) a - ks);
167
168 nc -= 16;
169 } else {
170 // Prepare mask for valid 8-bit elements (depends on nc).
171 __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT32_C(1) << (nc + 16)) - (UINT32_C(1) << 16)));
172
173 _mm256_mask_storeu_epi8(c1 - 16, vmask, vout01x0123456789ABCDEF);
174 vmask = _kshiftri_mask64(vmask, 16);
175 _mm256_mask_storeu_epi8(c0, vmask, vout01x0123456789ABCDEF);
176
177 nc = 0;
178 }
179 } while (nc != 0);
180 }
181