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_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const uint8_t** restrict a,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const uint8_t* zero,
31 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 4);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(uint8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8);
43 uint8_t* c0 = c;
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 c1 = c0;
47 }
48 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
49 if XNN_UNPREDICTABLE(mr <= 2) {
50 c2 = c1;
51 }
52 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
53 if XNN_UNPREDICTABLE(mr != 4) {
54 c3 = c2;
55 }
56
57 const __mmask16 vbias_mask = _cvtu32_mask16(0x1111);
58 const __m512 vscale = _mm512_load_ps(params->fp32_avx512.scale);
59 const __m512 voutput_max_less_zero_point = _mm512_load_ps(params->fp32_avx512.output_max_less_zero_point);
60 const __m512i voutput_zero_point = _mm512_load_si512(params->fp32_avx512.output_zero_point);
61 const __m512i voutput_min = _mm512_load_si512(params->fp32_avx512.output_min);
62 do {
63 __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w);
64 __m512i vacc0x4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w + 4));
65 __m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w + 8));
66 __m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((const int32_t*) w + 12));
67 __m512i vacc1x0123 = vacc0x0123;
68 __m512i vacc1x4567 = vacc0x4567;
69 __m512i vacc1x89AB = vacc0x89AB;
70 __m512i vacc1xCDEF = vacc0xCDEF;
71 __m512i vacc2x0123 = vacc0x0123;
72 __m512i vacc2x4567 = vacc0x4567;
73 __m512i vacc2x89AB = vacc0x89AB;
74 __m512i vacc2xCDEF = vacc0xCDEF;
75 __m512i vacc3x0123 = vacc0x0123;
76 __m512i vacc3x4567 = vacc0x4567;
77 __m512i vacc3x89AB = vacc0x89AB;
78 __m512i vacc3xCDEF = vacc0xCDEF;
79 w = (const void*) ((const int32_t*) w + 16);
80
81 size_t p = ks;
82 do {
83 const uint8_t* restrict a0 = a[0];
84 if XNN_UNPREDICTABLE(a0 != zero) {
85 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
86 }
87 const uint8_t* restrict a1 = a[1];
88 if XNN_UNPREDICTABLE(a1 != zero) {
89 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
90 }
91 const uint8_t* restrict a2 = a[2];
92 if XNN_UNPREDICTABLE(a2 != zero) {
93 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
94 }
95 const uint8_t* restrict a3 = a[3];
96 if XNN_UNPREDICTABLE(a3 != zero) {
97 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
98 }
99 a += 4;
100
101 size_t k = 0;
102 const __m512i vb_zero_point = _mm512_load_si512(params->fp32_avx512.kernel_zero_point);
103 while (k < kc) {
104 const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
105 a0 += 8;
106 const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
107 a1 += 8;
108 const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
109 a2 += 8;
110 const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
111 a3 += 8;
112
113 const __m512i vb0123 = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) w)), vb_zero_point);
114
115 vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
116 vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
117 vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
118 vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
119 const __m512i vb4567 = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((const uint8_t*) w + 32))), vb_zero_point);
120
121 vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
122 vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
123 vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
124 vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
125 const __m512i vb89AB = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((const uint8_t*) w + 64))), vb_zero_point);
126
127 vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
128 vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
129 vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
130 vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
131 const __m512i vbCDEF = _mm512_sub_epi16(_mm512_cvtepu8_epi16(_mm256_load_si256((const __m256i*) ((const uint8_t*) w + 96))), vb_zero_point);
132
133 vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
134 vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
135 vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
136 vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
137
138 w = (const void*) ((const uint8_t*) w + 128);
139 k += 8 * sizeof(uint8_t);
140 }
141 p -= 4 * sizeof(void*);
142 } while (p != 0);
143
144 const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
145 const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
146 const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
147 const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
148 const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
149 const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
150 const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
151 const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
152
153 __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
154 __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
155 __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
156 __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
157
158 __m512 vscaled0x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc0x084C195D2A6E3B7F);
159 __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F);
160 __m512 vscaled2x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc2x084C195D2A6E3B7F);
161 __m512 vscaled3x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc3x084C195D2A6E3B7F);
162
163 vscaled0x084C195D2A6E3B7F = _mm512_mul_ps(vscaled0x084C195D2A6E3B7F, vscale);
164 vscaled1x084C195D2A6E3B7F = _mm512_mul_ps(vscaled1x084C195D2A6E3B7F, vscale);
165 vscaled2x084C195D2A6E3B7F = _mm512_mul_ps(vscaled2x084C195D2A6E3B7F, vscale);
166 vscaled3x084C195D2A6E3B7F = _mm512_mul_ps(vscaled3x084C195D2A6E3B7F, vscale);
167
168 vscaled0x084C195D2A6E3B7F = _mm512_min_ps(vscaled0x084C195D2A6E3B7F, voutput_max_less_zero_point);
169 vscaled1x084C195D2A6E3B7F = _mm512_min_ps(vscaled1x084C195D2A6E3B7F, voutput_max_less_zero_point);
170 vscaled2x084C195D2A6E3B7F = _mm512_min_ps(vscaled2x084C195D2A6E3B7F, voutput_max_less_zero_point);
171 vscaled3x084C195D2A6E3B7F = _mm512_min_ps(vscaled3x084C195D2A6E3B7F, voutput_max_less_zero_point);
172
173 vacc0x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled0x084C195D2A6E3B7F);
174 vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F);
175 vacc2x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled2x084C195D2A6E3B7F);
176 vacc3x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled3x084C195D2A6E3B7F);
177
178 const __m512i vacc01x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_zero_point);
179 const __m512i vacc23x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc2x084C195D2A6E3B7F, vacc3x084C195D2A6E3B7F), voutput_zero_point);
180
181 __m512i vout0123x084Cx195Dx2A6Ex3B7F = _mm512_packus_epi16(vacc01x084Cx195Dx2A6Ex3B7F, vacc23x084Cx195Dx2A6Ex3B7F);
182 vout0123x084Cx195Dx2A6Ex3B7F = _mm512_permutexvar_epi32(_mm512_set_epi32(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0), vout0123x084Cx195Dx2A6Ex3B7F);
183 __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));
184 vout0123x0123456789ABCDEF = _mm512_max_epu8(vout0123x0123456789ABCDEF, voutput_min);
185
186 if (nc >= 16) {
187 _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3));
188 _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2));
189 _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1));
190 _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF));
191
192 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
193 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
194 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
195 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
196
197 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
198
199 nc -= 16;
200 } else {
201 // Prepare mask for valid 8-bit elements (depends on nc).
202 __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT64_C(1) << (nc + 48)) - (UINT64_C(1) << 48)));
203
204 _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF);
205 vmask = _kshiftri_mask64(vmask, 16);
206 _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF);
207 vmask = _kshiftri_mask64(vmask, 16);
208 _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF);
209 vmask = _kshiftri_mask64(vmask, 16);
210 _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF);
211
212 nc = 0;
213 }
214 } while (nc != 0);
215 }
216