1 /*
2  * Copyright (c) 2017-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #pragma once
25 
26 #ifdef __arm__
27 
28 #include "transpose_interleave_common.hpp"
29 
30 // Generic unblocked transposed 8x32-bit sized specialisation
31 template <>
32 template <typename T>
Transform(T * out,const T * const in,const int stride,const int x0,const int xmax,const int k0,const int kmax)33 void TransformImpl<8, 1, true, 4, 4, VLType::None>::Transform(
34     T* out, const T* const in, const int stride,
35     const int x0, const int xmax, const int k0, const int kmax
36 ) {
37   // Redirect to a 16x uint16_t specialisation
38   TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform(
39     reinterpret_cast<uint16_t *>(out),
40     reinterpret_cast<const uint16_t *>(in),
41     stride*2, x0*2, xmax*2, k0, kmax
42   );
43 }
44 
45 // Generic 16x16-bit sized specialisation
46 template <>
47 template <typename T>
Transform(T * out,const T * const in,const int stride,const int x0,const int xmax,const int k0,const int kmax)48 void TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform(
49     T* out, const T* const in, const int stride,
50     const int x0, const int xmax, const int k0, const int kmax
51 ) {
52   // Redirect to a uint16_t specialisation
53   Transform(
54     reinterpret_cast<uint16_t *>(out),
55     reinterpret_cast<const uint16_t *>(in),
56     stride, x0, xmax, k0, kmax
57   );
58 }
59 
60 // Specialised 16 x uint16_t version
61 template <>
moveblock_1x1(const uint16_t * & in0,uint16_t * out)62 void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x1(const uint16_t *&in0, uint16_t *out) {
63   __asm volatile (
64     "VLD1.32	{d0-d3}, [%[in0]]!\n"
65     "VST1.32	{d0-d3}, [%[out]]\n"
66     ASM_PREFETCH("[%[in0], #192]")
67     : [in0] "+r" (in0),
68       [out] "+r" (out)
69     :
70     : "q0", "q1", "memory"
71   );
72 }
73 
74 template <>
moveblock_1x2(const uint16_t * & in0,const uint16_t * & in1,uint16_t * out)75 void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x2(const uint16_t *&in0, const uint16_t *&in1, uint16_t *out) {
76   __asm volatile (
77     "VLD1.32	{d0-d3}, [%[in0]]!\n"
78     "VST1.32	{d0-d3}, [%[out]]!\n"
79     ASM_PREFETCH("[%[in0], #192]")
80     "VLD1.32	{d0-d3}, [%[in1]]!\n"
81     "VST1.32	{d0-d3}, [%[out]]\n"
82     ASM_PREFETCH("[%[in1], #192]")
83     "SUB	%[out], %[out], #32\n"
84     : [in0] "+r" (in0),
85       [in1] "+r" (in1),
86       [out] "+r" (out)
87     :
88     : "q0", "q1", "memory"
89   );
90 }
91 
92 template <>
moveblock_1x4(const uint16_t * & in0,const uint16_t * & in1,const uint16_t * & in2,const uint16_t * & in3,uint16_t * out)93 void TransposeInterleaveCommon<16, uint16_t, uint16_t>::moveblock_1x4(const uint16_t *&in0, const uint16_t *&in1, const uint16_t *&in2, const uint16_t *&in3, uint16_t *out) {
94   __asm __volatile (
95     "VLD1.32	{d0-d3}, [%[in0]]!\n"
96     "VST1.32	{d0-d3}, [%[out]]!\n"
97     ASM_PREFETCH("[%[in0], #192]")
98     "VLD1.32	{d0-d3}, [%[in1]]!\n"
99     "VST1.32	{d0-d3}, [%[out]]!\n"
100     ASM_PREFETCH("[%[in1], #192]")
101     "VLD1.32	{d0-d3}, [%[in2]]!\n"
102     "VST1.32	{d0-d3}, [%[out]]!\n"
103     ASM_PREFETCH("[%[in2], #192]")
104     "VLD1.32	{d0-d3}, [%[in3]]!\n"
105     "VST1.32	{d0-d3}, [%[out]]\n"
106     ASM_PREFETCH("[%[in3], #192]")
107     "SUB	%[out], %[out], #96\n"
108     : [in0] "+r" (in0),
109       [in1] "+r" (in1),
110       [in2] "+r" (in2),
111       [in3] "+r" (in3),
112       [out] "+r" (out)
113     :
114     : "q0", "q1", "memory"
115   );
116 }
117 
118 template <>
119 template <>
Transform(uint16_t * out,const uint16_t * const in,const int stride,const int x0,const int xmax,const int k0,const int kmax)120 void TransformImpl<16, 1, true, 2, 2, VLType::None>::Transform(
121     uint16_t* out, const uint16_t* const in, const int stride,
122     const int x0, const int xmax, const int k0, const int kmax
123 ) {
124   TransposeInterleaveCommon<16, uint16_t, uint16_t>::Transform(out, in, stride, x0, xmax, k0, kmax);
125 }
126 
127 #endif // __arm__
128