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_2x8(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,const union xnn_f32_output_params params[restrict static1])18 void xnn_f32_prelu_ukernel__sse2_2x8(
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,
26 const union xnn_f32_output_params params[restrict static 1])
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 __m128 vmin = _mm_load_ps(params->sse.min);
45 const __m128 vmax = _mm_load_ps(params->sse.max);
46 do {
47 const float* w = weights;
48 size_t c = channels;
49 for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
50 const __m128 vw0123 = _mm_load_ps(w);
51 const __m128 vw4567 = _mm_load_ps(w + 4);
52 w += 8;
53
54 const __m128 vi0x0123 = _mm_loadu_ps(i0);
55 const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4);
56 i0 += 8;
57 const __m128 vi1x0123 = _mm_loadu_ps(i1);
58 const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4);
59 i1 += 8;
60
61 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123);
62 const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123)));
63 const __m128 vprod0x4567 = _mm_mul_ps(vi0x4567, vw4567);
64 const __m128 vmask0x4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x4567)));
65 const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123);
66 const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x0123)));
67 const __m128 vprod1x4567 = _mm_mul_ps(vi1x4567, vw4567);
68 const __m128 vmask1x4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x4567)));
69
70 __m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123));
71 __m128 vacc0x4567 = _mm_or_ps(_mm_and_ps(vprod0x4567, vmask0x4567), _mm_andnot_ps(vmask0x4567, vi0x4567));
72 __m128 vacc1x0123 = _mm_or_ps(_mm_and_ps(vprod1x0123, vmask1x0123), _mm_andnot_ps(vmask1x0123, vi1x0123));
73 __m128 vacc1x4567 = _mm_or_ps(_mm_and_ps(vprod1x4567, vmask1x4567), _mm_andnot_ps(vmask1x4567, vi1x4567));
74
75 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
76 vacc0x4567 = _mm_max_ps(vacc0x4567, vmin);
77 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
78 vacc1x4567 = _mm_max_ps(vacc1x4567, vmin);
79
80 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
81 vacc0x4567 = _mm_min_ps(vacc0x4567, vmax);
82 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
83 vacc1x4567 = _mm_min_ps(vacc1x4567, vmax);
84
85 _mm_storeu_ps(o0, vacc0x0123);
86 _mm_storeu_ps(o0 + 4, vacc0x4567);
87 o0 += 8;
88 _mm_storeu_ps(o1, vacc1x0123);
89 _mm_storeu_ps(o1 + 4, vacc1x4567);
90 o1 += 8;
91 }
92 for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) {
93 const __m128 vw0123 = _mm_load_ps(w);
94 w += 4;
95
96 const __m128 vi0x0123 = _mm_loadu_ps(i0);
97 i0 += 4;
98 const __m128 vi1x0123 = _mm_loadu_ps(i1);
99 i1 += 4;
100
101 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123);
102 const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123)));
103 const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123);
104 const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x0123)));
105
106 __m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123));
107 __m128 vacc1x0123 = _mm_or_ps(_mm_and_ps(vprod1x0123, vmask1x0123), _mm_andnot_ps(vmask1x0123, vi1x0123));
108
109 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
110 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
111
112 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
113 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
114
115 _mm_storeu_ps(o0, vacc0x0123);
116 o0 += 4;
117 _mm_storeu_ps(o1, vacc1x0123);
118 o1 += 4;
119 }
120 if XNN_UNLIKELY(c != 0) {
121 const __m128 vw0123 = _mm_load_ps(w);
122 w = (const float*) ((uintptr_t) w + c);
123
124 const __m128 vi0x0123 = _mm_loadu_ps(i0);
125 i0 = (const float*) ((uintptr_t) i0 + c);
126 const __m128 vi1x0123 = _mm_loadu_ps(i1);
127 i1 = (const float*) ((uintptr_t) i1 + c);
128
129 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123);
130 const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123)));
131 const __m128 vprod1x0123 = _mm_mul_ps(vi1x0123, vw0123);
132 const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi1x0123)));
133
134 __m128 vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123));
135 __m128 vacc1x0123 = _mm_or_ps(_mm_and_ps(vprod1x0123, vmask1x0123), _mm_andnot_ps(vmask1x0123, vi1x0123));
136
137 vacc0x0123 = _mm_max_ps(vacc0x0123, vmin);
138 vacc1x0123 = _mm_max_ps(vacc1x0123, vmin);
139
140 vacc0x0123 = _mm_min_ps(vacc0x0123, vmax);
141 vacc1x0123 = _mm_min_ps(vacc1x0123, vmax);
142
143 if (c & (2 * sizeof(float))) {
144 _mm_storel_pi((__m64*) o0, vacc0x0123);
145 _mm_storel_pi((__m64*) o1, vacc1x0123);
146
147 vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);
148 vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
149
150 o0 += 2;
151 o1 += 2;
152 }
153 if (c & (1 * sizeof(float))) {
154 _mm_store_ss(o0, vacc0x0123);
155 _mm_store_ss(o1, vacc1x0123);
156
157 o0 += 1;
158 o1 += 1;
159 }
160 }
161 i0 = (const float*) ((uintptr_t) i0 + input_increment);
162 o0 = (float*) ((uintptr_t) o0 + output_increment);
163 i1 = (const float*) ((uintptr_t) i1 + input_increment);
164 o1 = (float*) ((uintptr_t) o1 + output_increment);
165 if XNN_UNPREDICTABLE(rows < 4) {
166 i1 = i0;
167 o1 = o0;
168 }
169 rows = doz(rows, 2);
170 } while (rows != 0);
171 }
172