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_up16x9__avx512f_acc2(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_up16x9__avx512f_acc2(
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     const float* i4 = input[4];
57     assert(i4 != NULL);
58     if XNN_UNPREDICTABLE(i4 != zero) {
59       i4 = (const float*) ((uintptr_t) i4 + input_offset);
60     }
61     const float* i5 = input[5];
62     assert(i5 != NULL);
63     if XNN_UNPREDICTABLE(i5 != zero) {
64       i5 = (const float*) ((uintptr_t) i5 + input_offset);
65     }
66     const float* i6 = input[6];
67     assert(i6 != NULL);
68     if XNN_UNPREDICTABLE(i6 != zero) {
69       i6 = (const float*) ((uintptr_t) i6 + input_offset);
70     }
71     const float* i7 = input[7];
72     assert(i7 != NULL);
73     if XNN_UNPREDICTABLE(i7 != zero) {
74       i7 = (const float*) ((uintptr_t) i7 + input_offset);
75     }
76     const float* i8 = input[8];
77     assert(i8 != NULL);
78     if XNN_UNPREDICTABLE(i8 != zero) {
79       i8 = (const float*) ((uintptr_t) i8 + input_offset);
80     }
81     input = (const float**) ((uintptr_t) input + input_stride);
82 
83     size_t c = channels;
84     const float* w = weights;
85     for (; c >= 16; c -= 16) {
86       __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
87 
88 
89       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
90       i0 += 16;
91 
92       const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 16);
93       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
94 
95       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
96       i1 += 16;
97 
98       const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 32);
99       __m512 vacc0123456789ABCDEFp1 = _mm512_mul_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
100 
101       const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
102       i2 += 16;
103 
104       const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 48);
105       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
106 
107       const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
108       i3 += 16;
109 
110       const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 64);
111       vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp1);
112 
113       const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4);
114       i4 += 16;
115 
116       const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 80);
117       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
118 
119       const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5);
120       i5 += 16;
121 
122       const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96);
123       vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp1);
124 
125       const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6);
126       i6 += 16;
127 
128       const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 112);
129       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
130 
131       const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7);
132       i7 += 16;
133 
134       const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 128);
135       vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp1);
136 
137       const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8);
138       i8 += 16;
139 
140       const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144);
141       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
142 
143       w += 160;
144 
145       // Add up all accumulators to vacc0123456789ABCDEFp0
146       vacc0123456789ABCDEFp0 = _mm512_add_ps(vacc0123456789ABCDEFp0, vacc0123456789ABCDEFp1);
147 
148       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
149       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
150 
151       _mm512_storeu_ps(output, vacc0123456789ABCDEF);
152       output += 16;
153     }
154     if XNN_UNLIKELY(c != 0) {
155       assert(c >= 1);
156       assert(c <= 16);
157       // Prepare mask for valid 32-bit elements (depends on nc).
158       const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));
159 
160       __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w);
161 
162       const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0);
163       const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 16);
164       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
165 
166       const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1);
167       const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32);
168       __m512 vacc0123456789ABCDEFp1 = _mm512_mul_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
169 
170       const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2);
171       const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 48);
172       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
173 
174       const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3);
175       const __m512 vk3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64);
176       vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp1);
177 
178       const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4);
179       const __m512 vk4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 80);
180       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc0123456789ABCDEFp0);
181 
182       const __m512 vi5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i5);
183       const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96);
184       vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc0123456789ABCDEFp1);
185 
186       const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6);
187       const __m512 vk6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 112);
188       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc0123456789ABCDEFp0);
189 
190       const __m512 vi7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i7);
191       const __m512 vk7x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 128);
192       vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF, vacc0123456789ABCDEFp1);
193 
194       const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8);
195       const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 144);
196       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc0123456789ABCDEFp0);
197 
198       // Add up all accumulators to vacc0123456789ABCDEFp0
199       vacc0123456789ABCDEFp0 = _mm512_add_ps(vacc0123456789ABCDEFp0, vacc0123456789ABCDEFp1);
200 
201       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
202       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
203 
204       _mm512_mask_storeu_ps(output, vmask, vacc0123456789ABCDEF);
205       output += c;
206     }
207 
208     output = (float*) ((uintptr_t) output + output_increment);
209   } while (--output_width != 0);
210 }
211