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_2x32(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_2x32(
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
37 const size_t input_increment = input_stride * 2 - channels;
38 const size_t output_increment = output_stride * 2 - channels;
39
40 const __m512 vzero = _mm512_setzero_ps();
41 do {
42 if XNN_UNPREDICTABLE(rows < 2) {
43 i1 = i0;
44 o1 = o0;
45 }
46
47 const float* w = weights;
48 size_t c = channels;
49 for (; c >= 32 * sizeof(float); c -= 32 * sizeof(float)) {
50 const __m512 vw0123456789ABCDEF = _mm512_load_ps(w);
51 const __m512 vwGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 16);
52 w += 32;
53
54 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0);
55 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16);
56 i0 += 32;
57 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1);
58 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16);
59 i1 += 32;
60
61 const __mmask16 vsign0x0123456789ABCDEF = _mm512_cmp_ps_mask(vi0x0123456789ABCDEF, vzero, _CMP_LT_OQ);
62 const __m512 vacc0x0123456789ABCDEF = _mm512_mask_mul_ps(vi0x0123456789ABCDEF, vsign0x0123456789ABCDEF, vi0x0123456789ABCDEF, vw0123456789ABCDEF);
63 const __mmask16 vsign0xGHIJKLMNOPQRSTUV = _mm512_cmp_ps_mask(vi0xGHIJKLMNOPQRSTUV, vzero, _CMP_LT_OQ);
64 const __m512 vacc0xGHIJKLMNOPQRSTUV = _mm512_mask_mul_ps(vi0xGHIJKLMNOPQRSTUV, vsign0xGHIJKLMNOPQRSTUV, vi0xGHIJKLMNOPQRSTUV, vwGHIJKLMNOPQRSTUV);
65 const __mmask16 vsign1x0123456789ABCDEF = _mm512_cmp_ps_mask(vi1x0123456789ABCDEF, vzero, _CMP_LT_OQ);
66 const __m512 vacc1x0123456789ABCDEF = _mm512_mask_mul_ps(vi1x0123456789ABCDEF, vsign1x0123456789ABCDEF, vi1x0123456789ABCDEF, vw0123456789ABCDEF);
67 const __mmask16 vsign1xGHIJKLMNOPQRSTUV = _mm512_cmp_ps_mask(vi1xGHIJKLMNOPQRSTUV, vzero, _CMP_LT_OQ);
68 const __m512 vacc1xGHIJKLMNOPQRSTUV = _mm512_mask_mul_ps(vi1xGHIJKLMNOPQRSTUV, vsign1xGHIJKLMNOPQRSTUV, vi1xGHIJKLMNOPQRSTUV, vwGHIJKLMNOPQRSTUV);
69
70 _mm512_storeu_ps(o0, vacc0x0123456789ABCDEF);
71 _mm512_storeu_ps(o0 + 16, vacc0xGHIJKLMNOPQRSTUV);
72 o0 += 32;
73 _mm512_storeu_ps(o1, vacc1x0123456789ABCDEF);
74 _mm512_storeu_ps(o1 + 16, vacc1xGHIJKLMNOPQRSTUV);
75 o1 += 32;
76 }
77 for (; c >= 16 * sizeof(float); c -= 16 * sizeof(float)) {
78 const __m512 vw = _mm512_load_ps(w);
79 w += 16;
80
81 const __m512 vi0 = _mm512_loadu_ps(i0);
82 i0 += 16;
83 const __m512 vi1 = _mm512_loadu_ps(i1);
84 i1 += 16;
85
86 const __mmask16 vsign0 = _mm512_cmp_ps_mask(vi0, vzero, _CMP_LT_OQ);
87 const __m512 vacc0 = _mm512_mask_mul_ps(vi0, vsign0, vi0, vw);
88 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vi1, vzero, _CMP_LT_OQ);
89 const __m512 vacc1 = _mm512_mask_mul_ps(vi1, vsign1, vi1, vw);
90
91 _mm512_storeu_ps(o0, vacc0);
92 o0 += 16;
93 _mm512_storeu_ps(o1, vacc1);
94 o1 += 16;
95 }
96 if XNN_UNLIKELY(c != 0) {
97 assert(c >= 1 * sizeof(float));
98 assert(c <= 15 * sizeof(float));
99 // Prepare mask for valid 32-bit elements (depends on c).
100 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << (c >> 2 /* log2(sizeof(float))*/)) - UINT32_C(1)));
101
102 const __m512 vw = _mm512_maskz_loadu_ps(vmask, w);
103
104 const __m512 vi0 = _mm512_maskz_loadu_ps(vmask, i0);
105 i0 = (const float*) ((uintptr_t) i0 + c);
106 const __m512 vi1 = _mm512_maskz_loadu_ps(vmask, i1);
107 i1 = (const float*) ((uintptr_t) i1 + c);
108
109 const __mmask16 vsign0 = _mm512_cmp_ps_mask(vi0, vzero, _CMP_LT_OQ);
110 const __m512 vacc0 = _mm512_mask_mul_ps(vi0, vsign0, vi0, vw);
111 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vi1, vzero, _CMP_LT_OQ);
112 const __m512 vacc1 = _mm512_mask_mul_ps(vi1, vsign1, vi1, vw);
113
114 _mm512_mask_storeu_ps(o0, vmask, vacc0);
115 o0 = (float*) ((uintptr_t) o0 + c);
116 _mm512_mask_storeu_ps(o1, vmask, vacc1);
117 o1 = (float*) ((uintptr_t) o1 + c);
118 }
119 i0 = (const float*) ((uintptr_t) i0 + input_increment);
120 o0 = (float*) ((uintptr_t) o0 + output_increment);
121 i1 = (const float*) ((uintptr_t) i1 + input_increment);
122 o1 = (float*) ((uintptr_t) o1 + output_increment);
123 rows = doz(rows, 2);
124 } while (rows != 0);
125 }
126