• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_fp32_ukernel_4x16c8__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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__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_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(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   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52   int8_t* c3 = (int8_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 int8_t* restrict a0 = a[0];
84       if XNN_UNPREDICTABLE(a0 != zero) {
85         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
86       }
87       const int8_t* restrict a1 = a[1];
88       if XNN_UNPREDICTABLE(a1 != zero) {
89         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
90       }
91       const int8_t* restrict a2 = a[2];
92       if XNN_UNPREDICTABLE(a2 != zero) {
93         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
94       }
95       const int8_t* restrict a3 = a[3];
96       if XNN_UNPREDICTABLE(a3 != zero) {
97         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
98       }
99       a += 4;
100 
101       size_t k = 0;
102       while (k < kc) {
103         const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0)));
104         a0 += 8;
105         const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1)));
106         a1 += 8;
107         const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2)));
108         a2 += 8;
109         const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3)));
110         a3 += 8;
111 
112         const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w));
113 
114         vacc0x0123 = _mm512_add_epi32(vacc0x0123, _mm512_madd_epi16(va0, vb0123));
115         vacc1x0123 = _mm512_add_epi32(vacc1x0123, _mm512_madd_epi16(va1, vb0123));
116         vacc2x0123 = _mm512_add_epi32(vacc2x0123, _mm512_madd_epi16(va2, vb0123));
117         vacc3x0123 = _mm512_add_epi32(vacc3x0123, _mm512_madd_epi16(va3, vb0123));
118         const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 32)));
119 
120         vacc0x4567 = _mm512_add_epi32(vacc0x4567, _mm512_madd_epi16(va0, vb4567));
121         vacc1x4567 = _mm512_add_epi32(vacc1x4567, _mm512_madd_epi16(va1, vb4567));
122         vacc2x4567 = _mm512_add_epi32(vacc2x4567, _mm512_madd_epi16(va2, vb4567));
123         vacc3x4567 = _mm512_add_epi32(vacc3x4567, _mm512_madd_epi16(va3, vb4567));
124         const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 64)));
125 
126         vacc0x89AB = _mm512_add_epi32(vacc0x89AB, _mm512_madd_epi16(va0, vb89AB));
127         vacc1x89AB = _mm512_add_epi32(vacc1x89AB, _mm512_madd_epi16(va1, vb89AB));
128         vacc2x89AB = _mm512_add_epi32(vacc2x89AB, _mm512_madd_epi16(va2, vb89AB));
129         vacc3x89AB = _mm512_add_epi32(vacc3x89AB, _mm512_madd_epi16(va3, vb89AB));
130         const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((const int8_t*) w + 96)));
131 
132         vacc0xCDEF = _mm512_add_epi32(vacc0xCDEF, _mm512_madd_epi16(va0, vbCDEF));
133         vacc1xCDEF = _mm512_add_epi32(vacc1xCDEF, _mm512_madd_epi16(va1, vbCDEF));
134         vacc2xCDEF = _mm512_add_epi32(vacc2xCDEF, _mm512_madd_epi16(va2, vbCDEF));
135         vacc3xCDEF = _mm512_add_epi32(vacc3xCDEF, _mm512_madd_epi16(va3, vbCDEF));
136 
137         w = (const void*) ((const int8_t*) w + 128);
138         k += 8 * sizeof(int8_t);
139       }
140       p -= 4 * sizeof(void*);
141     } while (p != 0);
142 
143     const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm512_unpackhi_epi32(vacc0x0123, vacc0x4567));
144     const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm512_unpackhi_epi32(vacc0x89AB, vacc0xCDEF));
145     const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm512_unpackhi_epi32(vacc1x0123, vacc1x4567));
146     const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm512_unpackhi_epi32(vacc1x89AB, vacc1xCDEF));
147     const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm512_unpackhi_epi32(vacc2x0123, vacc2x4567));
148     const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm512_unpackhi_epi32(vacc2x89AB, vacc2xCDEF));
149     const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm512_unpackhi_epi32(vacc3x0123, vacc3x4567));
150     const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm512_unpackhi_epi32(vacc3x89AB, vacc3xCDEF));
151 
152     __m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9DAEBF), _mm512_unpackhi_epi32(vacc0x04152637, vacc0x8C9DAEBF));
153     __m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9DAEBF), _mm512_unpackhi_epi32(vacc1x04152637, vacc1x8C9DAEBF));
154     __m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9DAEBF), _mm512_unpackhi_epi32(vacc2x04152637, vacc2x8C9DAEBF));
155     __m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9DAEBF), _mm512_unpackhi_epi32(vacc3x04152637, vacc3x8C9DAEBF));
156 
157     __m512 vscaled0x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc0x084C195D2A6E3B7F);
158     __m512 vscaled1x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc1x084C195D2A6E3B7F);
159     __m512 vscaled2x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc2x084C195D2A6E3B7F);
160     __m512 vscaled3x084C195D2A6E3B7F = _mm512_cvtepi32_ps(vacc3x084C195D2A6E3B7F);
161 
162     vscaled0x084C195D2A6E3B7F = _mm512_mul_ps(vscaled0x084C195D2A6E3B7F, vscale);
163     vscaled1x084C195D2A6E3B7F = _mm512_mul_ps(vscaled1x084C195D2A6E3B7F, vscale);
164     vscaled2x084C195D2A6E3B7F = _mm512_mul_ps(vscaled2x084C195D2A6E3B7F, vscale);
165     vscaled3x084C195D2A6E3B7F = _mm512_mul_ps(vscaled3x084C195D2A6E3B7F, vscale);
166 
167     vscaled0x084C195D2A6E3B7F = _mm512_min_ps(vscaled0x084C195D2A6E3B7F, voutput_max_less_zero_point);
168     vscaled1x084C195D2A6E3B7F = _mm512_min_ps(vscaled1x084C195D2A6E3B7F, voutput_max_less_zero_point);
169     vscaled2x084C195D2A6E3B7F = _mm512_min_ps(vscaled2x084C195D2A6E3B7F, voutput_max_less_zero_point);
170     vscaled3x084C195D2A6E3B7F = _mm512_min_ps(vscaled3x084C195D2A6E3B7F, voutput_max_less_zero_point);
171 
172     vacc0x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled0x084C195D2A6E3B7F);
173     vacc1x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled1x084C195D2A6E3B7F);
174     vacc2x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled2x084C195D2A6E3B7F);
175     vacc3x084C195D2A6E3B7F = _mm512_cvtps_epi32(vscaled3x084C195D2A6E3B7F);
176 
177     const __m512i vacc01x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, vacc1x084C195D2A6E3B7F), voutput_zero_point);
178     const __m512i vacc23x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc2x084C195D2A6E3B7F, vacc3x084C195D2A6E3B7F), voutput_zero_point);
179 
180     __m512i vout0123x084Cx195Dx2A6Ex3B7F = _mm512_packs_epi16(vacc01x084Cx195Dx2A6Ex3B7F, vacc23x084Cx195Dx2A6Ex3B7F);
181     vout0123x084Cx195Dx2A6Ex3B7F = _mm512_permutexvar_epi32(_mm512_set_epi32(15, 11, 7, 3, 14, 10, 6, 2, 13, 9, 5, 1, 12, 8, 4, 0), vout0123x084Cx195Dx2A6Ex3B7F);
182     __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));
183     vout0123x0123456789ABCDEF = _mm512_max_epi8(vout0123x0123456789ABCDEF, voutput_min);
184 
185     if (nc >= 16) {
186       _mm_storeu_si128((__m128i*) c3, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 3));
187       _mm_storeu_si128((__m128i*) c2, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 2));
188       _mm_storeu_si128((__m128i*) c1, _mm512_extracti32x4_epi32(vout0123x0123456789ABCDEF, 1));
189       _mm_storeu_si128((__m128i*) c0, _mm512_castsi512_si128(vout0123x0123456789ABCDEF));
190 
191       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
192       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
193       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
194       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
195 
196       a = (const int8_t**restrict) ((uintptr_t) a - ks);
197 
198       nc -= 16;
199     } else {
200       // Prepare mask for valid 8-bit elements (depends on nc).
201       __mmask64 vmask = _cvtu64_mask64((uint64_t) ((UINT64_C(1) << (nc + 48)) - (UINT64_C(1) << 48)));
202 
203       _mm512_mask_storeu_epi8(c3 - 48, vmask, vout0123x0123456789ABCDEF);
204       vmask = _kshiftri_mask64(vmask, 16);
205       _mm512_mask_storeu_epi8(c2 - 32, vmask, vout0123x0123456789ABCDEF);
206       vmask = _kshiftri_mask64(vmask, 16);
207       _mm512_mask_storeu_epi8(c1 - 16, vmask, vout0123x0123456789ABCDEF);
208       vmask = _kshiftri_mask64(vmask, 16);
209       _mm512_mask_storeu_epi8(c0, vmask, vout0123x0123456789ABCDEF);
210 
211       nc = 0;
212     }
213   } while (nc != 0);
214 }
215