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_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 4);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8);
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
50 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
56 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr != 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61
62 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
63 const __m512 vscale = _mm512_load_ps(params->fp32_avx512.scale);
64 const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->fp32_avx512.output_max_less_zero_point);
65 const __m512i voutput_zero_point = _mm512_load_si512(params->fp32_avx512.output_zero_point);
66 const __m512i voutput_min = _mm512_load_si512(params->fp32_avx512.output_min);
67 do {
68 __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
69 __m512i vacc0x4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w + 4));
70 __m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w + 8));
71 __m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w + 12));
72 __m512i vacc1x0123 = vacc0x0123;
73 __m512i vacc1x4567 = vacc0x4567;
74 __m512i vacc1x89AB = vacc0x89AB;
75 __m512i vacc1xCDEF = vacc0xCDEF;
76 __m512i vacc2x0123 = vacc0x0123;
77 __m512i vacc2x4567 = vacc0x4567;
78 __m512i vacc2x89AB = vacc0x89AB;
79 __m512i vacc2xCDEF = vacc0xCDEF;
80 __m512i vacc3x0123 = vacc0x0123;
81 __m512i vacc3x4567 = vacc0x4567;
82 __m512i vacc3x89AB = vacc0x89AB;
83 __m512i vacc3xCDEF = vacc0xCDEF;
84 w = (const void*) ((const int32_t*) w + 16);
85
86 size_t k = 0;
87 const __m512i vb_zero_point = _mm512_load_si512(params->fp32_avx512.kernel_zero_point);
88 while (k < kc) {
89 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
90 a0 += 8;
91 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
92 a1 += 8;
93 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
94 a2 += 8;
95 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
96 a3 += 8;
97
98 const __m512i vb0123 = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) w)), vb_zero_point);
99
100 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
101 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
102 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
103 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
104 const __m512i vb4567 = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((const uint8_t*) w + 32))), vb_zero_point);
105
106 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
107 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
108 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
109 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
110 const __m512i vb89AB = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((const uint8_t*) w + 64))), vb_zero_point);
111
112 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
113 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
114 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
115 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
116 const __m512i vbCDEF = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((const uint8_t*) w + 96))), vb_zero_point);
117
118 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
119 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
120 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
121 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
122
123 w = (const void*) ((const uint8_t*) w + 128);
124 k += 8 * sizeof(uint8_t);
125 }
126
127 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
128 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
129 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
130 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
131 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
132 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
133 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
134 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
135
136 __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
137 __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
138 __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
139 __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
140
141 __m512 vscaled0x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc0x084C195D2A6E3B7F);
142 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F);
143 __m512 vscaled2x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc2x084C195D2A6E3B7F);
144 __m512 vscaled3x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc3x084C195D2A6E3B7F);
145
146 vscaled0x084C195D2A6E3B7F = _mm512_mul_ps(vscaled0x084C195D2A6E3B7F, vscale);
147 vscaled1x084C195D2A6E3B7F = _mm512_mul_ps(vscaled1x084C195D2A6E3B7F, vscale);
148 vscaled2x084C195D2A6E3B7F = _mm512_mul_ps(vscaled2x084C195D2A6E3B7F, vscale);
149 vscaled3x084C195D2A6E3B7F = _mm512_mul_ps(vscaled3x084C195D2A6E3B7F, vscale);
150
151 vscaled0x084C195D2A6E3B7F = _mm512_min_ps(vscaled0x084C195D2A6E3B7F, voutput_max_less_zero_point);
152 vscaled1x084C195D2A6E3B7F = _mm512_min_ps(vscaled1x084C195D2A6E3B7F, voutput_max_less_zero_point);
153 vscaled2x084C195D2A6E3B7F = _mm512_min_ps(vscaled2x084C195D2A6E3B7F, voutput_max_less_zero_point);
154 vscaled3x084C195D2A6E3B7F = _mm512_min_ps(vscaled3x084C195D2A6E3B7F, voutput_max_less_zero_point);
155
156 vacc0x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled0x084C195D2A6E3B7F);
157 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F);
158 vacc2x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled2x084C195D2A6E3B7F);
159 vacc3x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled3x084C195D2A6E3B7F);
160
161 const __m512i vacc01x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_zero_point);
162 const __m512i vacc23x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc2x084C195D2A6E3B7F, vacc3x084C195D2A6E3B7F), voutput_zero_point);
163
164 __m512i vout0123x084Cx195Dx2A6Ex3B7F = _mm512_packus_epi16(vacc01x084Cx195Dx2A6Ex3B7F, vacc23x084Cx195Dx2A6Ex3B7F);
165 vout0123x084Cx195Dx2A6Ex3B7F = _mm512_permutexvar_epi32(_mm512_set_epi32(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0), vout0123x084Cx195Dx2A6Ex3B7F);
166 __m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512_set_epi8(15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0, 15, 11, 7, 3, 13, 9, 5, 1, 14, 10, 6, 2, 12, 8, 4, 0));
167 vout0123x0123456789ABCDEF = _mm512_max_epu8(vout0123x0123456789ABCDEF, voutput_min);
168
169 if (nc >= 16) {
170 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF));
171 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1));
172 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2));
173 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3));
174
175 a0 = (const uint8_t*) ((uintptr_t) a0 - k);
176 a1 = (const uint8_t*) ((uintptr_t) a1 - k);
177 a2 = (const uint8_t*) ((uintptr_t) a2 - k);
178 a3 = (const uint8_t*) ((uintptr_t) a3 - k);
179
180 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
181 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
182 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
183 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
184
185 nc -= 16;
186 } else {
187 // Prepare mask for valid 8-bit elements (depends on nc).
188 __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT32_C(1) << nc) - UINT32_C(1)));
189
190 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF);
191 vmask = _kshiftli_mask64(vmask, 16);
192 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF);
193 vmask = _kshiftli_mask64(vmask, 16);
194 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF);
195 vmask = _kshiftli_mask64(vmask, 16);
196 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF);
197
198 nc = 0;
199 }
200 } while (nc != 0);
201 }
202