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_1x16c8__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_1x16c8__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 <= 1);
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
45 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
46 const __mmask16 vblend_mask = _cvtu32_mask16(0xAAAA);
47 const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.multiplier));
48 const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.rounding));
49 const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
50 const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
51 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
52 const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
53 const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
54 const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
55 do {
56 __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
57 __m512i vacc0x4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
58 __m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 * sizeof(int32_t)));
59 __m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 12 * sizeof(int32_t)));
60 w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t));
61
62 size_t p = ks;
63 do {
64 const int8_t* restrict a0 = a[0];
65 if XNN_UNPREDICTABLE(a0 != zero) {
66 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
67 }
68 a += 1;
69
70 size_t k = 0;
71 while (k < kc) {
72 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
73 a0 += 8;
74
75 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
76
77 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
78 const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 32 * sizeof(int8_t))));
79
80 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
81 const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 64 * sizeof(int8_t))));
82
83 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
84 const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 96 * sizeof(int8_t))));
85
86 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
87
88 w = (const void*) ((uintptr_t) w + 128 * sizeof(int8_t));
89 k += 8 * sizeof(int8_t);
90 }
91 p -= 1 * sizeof(void*);
92 } while (p != 0);
93
94 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
95 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
96
97 __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
98
99 const __m512i vacc0x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc0x084C195D2A6E3B7F, _MM_SHUFFLE(3, 3, 1, 1));
100
101 const __m512i vprod0x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc0x084C195D2A6E3B7F, vmultiplier), vrounding);
102
103 const __m512i vprod0x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc0x88CC99DDAAEEBBFF, vmultiplier), vrounding);
104
105 const __m512i vq31prod0x04152637 = _mm512_srli_epi64(vprod0x04152637, 31);
106 const __m512i vq31prod0x8C9DAEBF = _mm512_add_epi64(vprod0x8C9DAEBF, vprod0x8C9DAEBF);
107
108 const __m512i vq31prod0x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod0x04152637, vq31prod0x8C9DAEBF);
109
110 const __m512i vrem0x084C195D2A6E3B7F =
111 _mm512_add_epi32(_mm512_and_si512(vq31prod0x084C195D2A6E3B7F, vremainder_mask), _mm512_srai_epi32(vq31prod0x084C195D2A6E3B7F, 31));
112
113 vacc0x084C195D2A6E3B7F = _mm512_sra_epi32(vq31prod0x084C195D2A6E3B7F, vshift);
114
115 const __m512i vminus_one = _mm512_set1_epi32(-1);
116 vacc0x084C195D2A6E3B7F =
117 _mm512_mask_sub_epi32(vacc0x084C195D2A6E3B7F, _mm512_cmpgt_epi32_mask(vrem0x084C195D2A6E3B7F, vremainder_threshold), vacc0x084C195D2A6E3B7F, vminus_one);
118
119 __m256i vacc0x084C2A6E195D3B7F = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0x084C195D2A6E3B7F), _mm512_extracti32x8_epi32(vacc0x084C195D2A6E3B7F, 1)), voutput_zero_point);
120 vacc0x084C2A6E195D3B7F = _mm256_min_epi16(_mm256_max_epi16(vacc0x084C2A6E195D3B7F, voutput_min), voutput_max);
121
122 const __m128i vout0x084C2A6E195D3B7F = _mm_packs_epi16(_mm256_castsi256_si128(vacc0x084C2A6E195D3B7F), _mm256_extracti128_si256(vacc0x084C2A6E195D3B7F, 1));
123 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));
124
125 if (nc >= 16) {
126 _mm_storeu_si128((__m128i*) c0, vout0x0123456789ABCDEF);
127
128 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
129
130 a = (const int8_t**restrict) ((uintptr_t) a - ks);
131
132 nc -= 16;
133 } else {
134 // Prepare mask for valid 8-bit elements (depends on nc).
135 const __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT32_C(1) << nc) - UINT32_C(1)));
136
137 _mm_mask_storeu_epi8(c0, vmask, vout0x0123456789ABCDEF);
138
139 nc = 0;
140 }
141 } while (nc != 0);
142 }
143