1 // Auto-generated file. Do not edit!
2 // Template: src/f32-prelu/sse.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 <emmintrin.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/prelu.h>
16
17
xnn_f32_prelu_ukernel__sse2_2x4(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)18 void xnn_f32_prelu_ukernel__sse2_2x4(
19 size_t rows,
20 size_t channels,
21 const float*restrict input,
22 size_t input_stride,
23 const float*restrict weights,
24 float*restrict output,
25 size_t output_stride) XNN_OOB_READS
26 {
27 assert(rows != 0);
28 assert(channels != 0);
29 assert(channels % sizeof(float) == 0);
30
31 const float* i0 = input;
32 float* o0 = output;
33 const float* i1 = (const float*) ((uintptr_t) i0 + input_stride);
34 float* o1 = (float*) ((uintptr_t) o0 + output_stride);
35
36 const size_t input_increment = input_stride * 2 - channels;
37 const size_t output_increment = output_stride * 2 - channels;
38
39 do {
40 if XNN_UNPREDICTABLE(rows < 2) {
41 i1 = i0;
42 o1 = o0;
43 }
44
45 const float* w = weights;
46 size_t c = channels;
47 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
48 const __m128 vw0123 = _mm_load_ps(w);
49 w += 4;
50
51 const __m128 vi0x0123 = _mm_loadu_ps(i0);
52 i0 += 4;
53 const __m128 vi1x0123 = _mm_loadu_ps(i1);
54 i1 += 4;
55
56 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123);
57 const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123)));
58 const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123);
59 const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x0123)));
60
61 const __m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123));
62 const __m128 vacc1x0123 = _mm_or_ps(_mm_and_ps(vprod1x0123, vmask1x0123), _mm_andnot_ps(vmask1x0123, vi1x0123));
63
64 _mm_storeu_ps(o0, vacc0x0123);
65 o0 += 4;
66 _mm_storeu_ps(o1, vacc1x0123);
67 o1 += 4;
68 }
69 if XNN_UNLIKELY(c != 0) {
70 const __m128 vw0123 = _mm_load_ps(w);
71 w = (const float*) ((uintptr_t) w + c);
72
73 const __m128 vi0x0123 = _mm_loadu_ps(i0);
74 i0 = (const float*) ((uintptr_t) i0 + c);
75 const __m128 vi1x0123 = _mm_loadu_ps(i1);
76 i1 = (const float*) ((uintptr_t) i1 + c);
77
78 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123);
79 const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123)));
80 const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123);
81 const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x0123)));
82
83 __m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123));
84 __m128 vacc1x0123 = _mm_or_ps(_mm_and_ps(vprod1x0123, vmask1x0123), _mm_andnot_ps(vmask1x0123, vi1x0123));
85
86 if (c & (2 * sizeof(float))) {
87 _mm_storel_pi((__m64*) o0, vacc0x0123);
88 _mm_storel_pi((__m64*) o1, vacc1x0123);
89
90 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
91 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
92
93 o0 += 2;
94 o1 += 2;
95 }
96 if (c & (1 * sizeof(float))) {
97 _mm_store_ss(o0, vacc0x0123);
98 _mm_store_ss(o1, vacc1x0123);
99
100 o0 += 1;
101 o1 += 1;
102 }
103 }
104 i0 = (const float*) ((uintptr_t) i0 + input_increment);
105 o0 = (float*) ((uintptr_t) o0 + output_increment);
106 i1 = (const float*) ((uintptr_t) i1 + input_increment);
107 o1 = (float*) ((uintptr_t) o1 + output_increment);
108 rows = doz(rows, 2);
109 } while (rows != 0);
110 }
111