• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-prelu/avx512f.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/intrinsics-polyfill.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/prelu.h>
17 
18 
xnn_f32_prelu_ukernel__avx512f_2x16(size_t rows,size_t channels,const float * restrict input,size_t input_stride,const float * restrict weights,float * restrict output,size_t output_stride)19 void xnn_f32_prelu_ukernel__avx512f_2x16(
20     size_t rows,
21     size_t channels,
22     const float*restrict input,
23     size_t input_stride,
24     const float*restrict weights,
25     float*restrict output,
26     size_t output_stride)
27 {
28   assert(rows != 0);
29   assert(channels != 0);
30   assert(channels % sizeof(float) == 0);
31 
32   const float* i0 = input;
33   float* o0 = output;
34   const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
35   float* o1 = (float*) ((uintptr_t) o0 + output_stride);
36   if XNN_UNPREDICTABLE(rows < 2) {
37     i1 = i0;
38     o1 = o0;
39   }
40 
41   const size_t input_increment = input_stride * 2 - channels;
42   const size_t output_increment = output_stride * 2 - channels;
43 
44   const __m512 vzero = _mm512_setzero_ps();
45   do {
46     const float* w = weights;
47     size_t c = channels;
48     for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
49       const __m512 vw0123456789ABCDEF = _mm512_load_ps(w);
50       w += 16;
51 
52       const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
53       i0 += 16;
54       const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
55       i1 += 16;
56 
57       const __mmask16 vsign0x0123456789ABCDEF = _mm512_cmp_ps_mask(vi0x0123456789ABCDEF, vzero, _CMP_LT_OQ);
58       const __m512 vacc0x0123456789ABCDEF = _mm512_mask_mul_ps(vi0x0123456789ABCDEF, vsign0x0123456789ABCDEF, vi0x0123456789ABCDEF, vw0123456789ABCDEF);
59       const __mmask16 vsign1x0123456789ABCDEF = _mm512_cmp_ps_mask(vi1x0123456789ABCDEF, vzero, _CMP_LT_OQ);
60       const __m512 vacc1x0123456789ABCDEF = _mm512_mask_mul_ps(vi1x0123456789ABCDEF, vsign1x0123456789ABCDEF, vi1x0123456789ABCDEF, vw0123456789ABCDEF);
61 
62       _mm512_storeu_ps(o0, vacc0x0123456789ABCDEF);
63       o0 += 16;
64       _mm512_storeu_ps(o1, vacc1x0123456789ABCDEF);
65       o1 += 16;
66     }
67     if XNN_UNLIKELY(c != 0) {
68       assert(c >= 1 * sizeof(float));
69       assert(c <= 15 * sizeof(float));
70       // Prepare mask for valid 32-bit elements (depends on c).
71       const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << (c >> 2 /* log2(sizeof(float))*/)) - UINT32_C(1)));
72 
73       const __m512 vw = _mm512_maskz_loadu_ps(vmask, w);
74 
75       const __m512 vi0 = _mm512_maskz_loadu_ps(vmask, i0);
76       i0 = (const float*) ((uintptr_t) i0 + c);
77       const __m512 vi1 = _mm512_maskz_loadu_ps(vmask, i1);
78       i1 = (const float*) ((uintptr_t) i1 + c);
79 
80       const __mmask16 vsign0 = _mm512_cmp_ps_mask(vi0, vzero, _CMP_LT_OQ);
81       const __m512 vacc0 = _mm512_mask_mul_ps(vi0, vsign0, vi0, vw);
82       const __mmask16 vsign1 = _mm512_cmp_ps_mask(vi1, vzero, _CMP_LT_OQ);
83       const __m512 vacc1 = _mm512_mask_mul_ps(vi1, vsign1, vi1, vw);
84 
85       _mm512_mask_storeu_ps(o0, vmask, vacc0);
86       o0 = (float*) ((uintptr_t) o0 + c);
87       _mm512_mask_storeu_ps(o1, vmask, vacc1);
88       o1 = (float*) ((uintptr_t) o1 + c);
89     }
90     i0 = (const float*) ((uintptr_t) i0 + input_increment);
91     o0 = (float*) ((uintptr_t) o0 + output_increment);
92     i1 = (const float*) ((uintptr_t) i1 + input_increment);
93     o1 = (float*) ((uintptr_t) o1 + output_increment);
94     if XNN_UNPREDICTABLE(rows < 4) {
95       i1 = i0;
96       o1 = o0;
97     }
98     rows = doz(rows, 2);
99   } while (rows != 0);
100 }
101