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_minmax_ukernel_up32x3__avx512f(size_t channels,size_t output_width,const float ** input,const float * weights,float * output,size_t input_stride,size_t output_increment,size_t input_offset,const float * zero,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_dwconv_minmax_ukernel_up32x3__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 size_t input_offset,
27 const float* zero,
28 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
29 {
30 assert(channels != 0);
31 assert(output_width != 0);
32
33 const __m512 vmax = _mm512_set1_ps(params->scalar.max);
34 const __m512 vmin = _mm512_set1_ps(params->scalar.min);
35 do {
36 const float* i0 = input[0];
37 assert(i0 != NULL);
38 if XNN_UNPREDICTABLE(i0 != zero) {
39 i0 = (const float*) ((uintptr_t) i0 + input_offset);
40 }
41 const float* i1 = input[1];
42 assert(i1 != NULL);
43 if XNN_UNPREDICTABLE(i1 != zero) {
44 i1 = (const float*) ((uintptr_t) i1 + input_offset);
45 }
46 const float* i2 = input[2];
47 assert(i2 != NULL);
48 if XNN_UNPREDICTABLE(i2 != zero) {
49 i2 = (const float*) ((uintptr_t) i2 + input_offset);
50 }
51 input = (const float**) ((uintptr_t) input + input_stride);
52
53 size_t c = channels;
54 const float* w = weights;
55 for (; c >= 32; c -= 32) {
56 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
57 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16);
58
59
60 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
61 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16);
62 i0 += 32;
63
64 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
65 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48);
66 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
67 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
68
69 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
70 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16);
71 i1 += 32;
72
73 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
74 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80);
75 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
76 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
77
78 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
79 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16);
80 i2 += 32;
81
82 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
83 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112);
84 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
85 vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
86
87 w += 128;
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 w += 16;
121
122
123 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
124 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
125
126 _mm512_storeu_ps(output, vacc0123456789ABCDEF);
127 output += 16;
128 }
129 if XNN_UNLIKELY(c != 0) {
130 assert(c >= 1);
131 assert(c <= 16);
132 // Prepare mask for valid 32-bit elements (depends on nc).
133 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));
134
135 __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w);
136
137 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0);
138 const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32);
139 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
140
141 const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1);
142 const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64);
143 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
144
145 const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2);
146 const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96);
147 vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
148
149
150 __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
151 vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
152
153 _mm512_mask_storeu_ps(output, vmask, vacc0123456789ABCDEF);
154 output += c;
155 }
156
157 output = (float*) ((uintptr_t) output + output_increment);
158 } while (--output_width != 0);
159 }
160