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_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,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_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     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_broadcast_f32x4(_mm_load_ps(params->sse.max));
34   const __m512 vmin = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.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     const float* i3 = input[3];
52     assert(i3 != NULL);
53     if XNN_UNPREDICTABLE(i3 != zero) {
54       i3 = (const float*) ((uintptr_t) i3 + input_offset);
55     }
56     input = (const float**) ((uintptr_t) input + input_stride);
57 
58     size_t c = channels;
59     const float* w = weights;
60     for (; c >= 32; c -= 32) {
61       __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
62       __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16);
63 
64 
65       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
66       const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16);
67       i0 += 32;
68 
69       const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
70       const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48);
71       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
72       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
73 
74       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
75       const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16);
76       i1 += 32;
77 
78       const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
79       const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80);
80       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
81       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
82 
83       const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
84       const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16);
85       i2 += 32;
86 
87       const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
88       const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112);
89       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
90       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
91 
92       const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
93       const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16);
94       i3 += 32;
95 
96       const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
97       const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144);
98       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
99       vaccGHIJKLMNOPQRSTUVp0 = _mm512_fmadd_ps(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUVp0);
100 
101       w += 160;
102 
103 
104       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
105       __m512 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUVp0, vmin);
106       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
107       vaccGHIJKLMNOPQRSTUV = _mm512_min_ps(vaccGHIJKLMNOPQRSTUV, vmax);
108 
109       _mm512_storeu_ps(output, vacc0123456789ABCDEF);
110       _mm512_storeu_ps(output + 16, vaccGHIJKLMNOPQRSTUV);
111       output += 32;
112     }
113     for (; c >= 16; c -= 16) {
114       __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
115 
116       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
117       i0 += 16;
118 
119       const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32);
120       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
121 
122       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
123       i1 += 16;
124 
125       const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64);
126       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
127 
128       const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
129       i2 += 16;
130 
131       const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96);
132       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
133 
134       const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
135       i3 += 16;
136 
137       const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128);
138       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
139 
140       w += 16;
141 
142 
143       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
144       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
145 
146       _mm512_storeu_ps(output, vacc0123456789ABCDEF);
147       output += 16;
148     }
149     if XNN_UNLIKELY(c != 0) {
150       assert(c >= 1);
151       assert(c <= 16);
152       // Prepare mask for valid 32-bit elements (depends on nc).
153       const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));
154 
155       __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w);
156 
157       const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0);
158       const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32);
159       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
160 
161       const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1);
162       const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64);
163       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF, vacc0123456789ABCDEFp0);
164 
165       const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2);
166       const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96);
167       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
168 
169       const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3);
170       const __m512 vk3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 128);
171       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp0);
172 
173 
174       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
175       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
176 
177       _mm512_mask_storeu_ps(output, vmask, vacc0123456789ABCDEF);
178       output += c;
179     }
180 
181     output = (float*) ((uintptr_t) output + output_increment);
182   } while (--output_width != 0);
183 }
184