1 // Auto-generated file. Do not edit!
2 // Template: src/f32-dwconv/up-avx512.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/dwconv.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16
17
xnn_f32_dwconv_ukernel_up32x4__avx512f(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,const union xnn_f32_output_params params[restrict static1])18 void xnn_f32_dwconv_ukernel_up32x4__avx512f(
19 size_t channels,
20 size_t output_width,
21 const float** input,
22 const float* weights,
23 float* output,
24 size_t input_stride,
25 size_t output_increment,
26 const union xnn_f32_output_params params[restrict static 1])
27 {
28 assert(channels != 0);
29 assert(output_width != 0);
30
31 const __m512 vmax = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
32 const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
33 do {
34 const float* i0 = input[0];
35 assert(i0 != NULL);
36 const float* i1 = input[1];
37 assert(i1 != NULL);
38 const float* i2 = input[2];
39 assert(i2 != NULL);
40 const float* i3 = input[3];
41 assert(i3 != NULL);
42 input = (const float**) ((uintptr_t) input + input_stride);
43
44 size_t c = channels;
45 const float* w = weights;
46 for (; c >= 32; c -= 32) {
47 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
48 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16);
49
50
51 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
52 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16);
53 i0 += 32;
54
55 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
56 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48);
57 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
58 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
59
60 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
61 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16);
62 i1 += 32;
63
64 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
65 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80);
66 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
67 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
68
69 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
70 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16);
71 i2 += 32;
72
73 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
74 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112);
75 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
76 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
77
78 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
79 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16);
80 i3 += 32;
81
82 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
83 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144);
84 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
85 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
86
87 w += 160;
88
89
90 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
91 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin);
92 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
93 vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax);
94
95 _mm512_storeu_ps(output, vacc0123456789ABCDEF);
96 _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV);
97 output += 32;
98 }
99 for (; c >= 16; c -= 16) {
100 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
101
102 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
103 i0 += 16;
104
105 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
106 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
107
108 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
109 i1 += 16;
110
111 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
112 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
113
114 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
115 i2 += 16;
116
117 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
118 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
119
120 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
121 i3 += 16;
122
123 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
124 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
125
126 w += 16;
127
128
129 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
130 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
131
132 _mm512_storeu_ps(output, vacc0123456789ABCDEF);
133 output += 16;
134 }
135 if XNN_UNLIKELY(c != 0) {
136 assert(c >= 1);
137 assert(c <= 16);
138 // Prepare mask for valid 32-bit elements (depends on nc).
139 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));
140
141 __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w);
142
143 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0);
144 const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32);
145 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
146
147 const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1);
148 const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64);
149 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
150
151 const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2);
152 const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96);
153 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
154
155 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3);
156 const __m512 vk3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 128);
157 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
158
159
160 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
161 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
162
163 _mm512_mask_storeu_ps(output, vmask, vacc0123456789ABCDEF);
164 output += c;
165 }
166
167 output = (float*) ((uintptr_t) output + output_increment);
168 } while (--output_width != 0);
169 }
170