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_up16x4__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_up16x4__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     input = (const float**) ((uintptr_t) input + input_stride);
57 
58     size_t c = channels;
59     const float* w = weights;
60     for (; c >= 16; c -= 16) {
61       __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w);
62 
63 
64       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
65       i0 += 16;
66 
67       const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 16);
68       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
69 
70       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
71       i1 += 16;
72 
73       const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 32);
74       __m512 vacc0123456789ABCDEFp1 = _mm512_mul_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
75 
76       const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2);
77       i2 += 16;
78 
79       const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 48);
80       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
81 
82       const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3);
83       i3 += 16;
84 
85       const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 64);
86       vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp1);
87 
88       w += 80;
89 
90       // Add up all accumulators to vacc0123456789ABCDEFp0
91       vacc0123456789ABCDEFp0 = _mm512_add_ps(vacc0123456789ABCDEFp0, vacc0123456789ABCDEFp1);
92 
93       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
94       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
95 
96       _mm512_storeu_ps(output, vacc0123456789ABCDEF);
97       output += 16;
98     }
99     if XNN_UNLIKELY(c != 0) {
100       assert(c >= 1);
101       assert(c <= 16);
102       // Prepare mask for valid 32-bit elements (depends on nc).
103       const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << c) - UINT32_C(1)));
104 
105       __m512 vacc0123456789ABCDEFp0 = _mm512_maskz_loadu_ps(vmask, w);
106 
107       const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0);
108       const __m512 vk0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 16);
109       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc0123456789ABCDEFp0);
110 
111       const __m512 vi1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i1);
112       const __m512 vk1x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 32);
113       __m512 vacc0123456789ABCDEFp1 = _mm512_mul_ps(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF);
114 
115       const __m512 vi2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i2);
116       const __m512 vk2x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 48);
117       vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF, vacc0123456789ABCDEFp0);
118 
119       const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3);
120       const __m512 vk3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 64);
121       vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc0123456789ABCDEFp1);
122 
123       // Add up all accumulators to vacc0123456789ABCDEFp0
124       vacc0123456789ABCDEFp0 = _mm512_add_ps(vacc0123456789ABCDEFp0, vacc0123456789ABCDEFp1);
125 
126       __m512 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEFp0, vmin);
127       vacc0123456789ABCDEF = _mm512_min_ps(vacc0123456789ABCDEF, vmax);
128 
129       _mm512_mask_storeu_ps(output, vmask, vacc0123456789ABCDEF);
130       output += c;
131     }
132 
133     output = (float*) ((uintptr_t) output + output_increment);
134   } while (--output_width != 0);
135 }
136