/opencl/oclfft : revision 1

1

2

//

3

// File: fft_internal.h

4

//

5

// Version: <1.0>

6

//

7

// Disclaimer: IMPORTANT: This Apple software is supplied to you by Apple Inc. ("Apple")

8

// in consideration of your agreement to the following terms, and your use,

9

// installation, modification or redistribution of this Apple software

10

// constitutes acceptance of these terms. If you do not agree with these

11

// terms, please do not use, install, modify or redistribute this Apple

12

// software.

13

//

14

// In consideration of your agreement to abide by the following terms, and

15

// subject to these terms, Apple grants you a personal, non - exclusive

16

// license, under Apple's copyrights in this original Apple software ( the

17

// "Apple Software" ), to use, reproduce, modify and redistribute the Apple

18

// Software, with or without modifications, in source and / or binary forms;

19

// provided that if you redistribute the Apple Software in its entirety and

20

// without modifications, you must retain this notice and the following text

21

// and disclaimers in all such redistributions of the Apple Software. Neither

22

// the name, trademarks, service marks or logos of Apple Inc. may be used to

23

// endorse or promote products derived from the Apple Software without specific

24

// prior written permission from Apple. Except as expressly stated in this

25

// notice, no other rights or licenses, express or implied, are granted by

26

// Apple herein, including but not limited to any patent rights that may be

27

// infringed by your derivative works or by other works in which the Apple

28

// Software may be incorporated.

29

//

30

// The Apple Software is provided by Apple on an "AS IS" basis. APPLE MAKES NO

31

// WARRANTIES, EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION THE IMPLIED

32

// WARRANTIES OF NON - INFRINGEMENT, MERCHANTABILITY AND FITNESS FOR A

33

// PARTICULAR PURPOSE, REGARDING THE APPLE SOFTWARE OR ITS USE AND OPERATION

34

// ALONE OR IN COMBINATION WITH YOUR PRODUCTS.

35

//

36

// IN NO EVENT SHALL APPLE BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL OR

37

// CONSEQUENTIAL DAMAGES ( INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF

38

// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS

39

// INTERRUPTION ) ARISING IN ANY WAY OUT OF THE USE, REPRODUCTION, MODIFICATION

40

// AND / OR DISTRIBUTION OF THE APPLE SOFTWARE, HOWEVER CAUSED AND WHETHER

41

// UNDER THEORY OF CONTRACT, TORT ( INCLUDING NEGLIGENCE ), STRICT LIABILITY OR

42

// OTHERWISE, EVEN IF APPLE HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

43

//

44

45

//

46

////////////////////////////////////////////////////////////////////////////////////////////////////

47

48

49

#ifndef __CLFFT_INTERNAL_H

50

#define __CLFFT_INTERNAL_H

51

52

#include "clFFT.h"

53

#include <sstream>

54

55

using namespace std;

56

57

typedef enum kernel_dir_t

58

{

59

cl_fft_kernel_x,

60

cl_fft_kernel_y,

61

cl_fft_kernel_z

62

}cl_fft_kernel_dir;

63

64

typedef struct kernel_info_t

65

{

66

cl_kernel kernel;

67

char *kernel_name;

68

size_t lmem_size;

69

size_t num_workgroups;

70

size_t num_xforms_per_workgroup;

71

size_t num_workitems_per_workgroup;

72

cl_fft_kernel_dir dir;

73

int in_place_possible;

74

kernel_info_t *next;

75

}cl_fft_kernel_info;

76

77

typedef struct

78

{

79

// context in which fft resources are created and kernels are executed

80

cl_context context;

81

82

// size of signal

83

clFFT_Dim3 n;

84

85

// dimension of transform ... must be either 1D, 2D or 3D

86

clFFT_Dimension dim;

87

88

// data format ... must be either interleaved or plannar

89

clFFT_DataFormat format;

90

91

// string containing kernel source. Generated at runtime based on

92

// n, dim, format and other parameters

93

string *kernel_string;

94

95

// CL program containing source and kernel this particular

96

// n, dim, data format

97

cl_program program;

98

99

// linked list of kernels which needs to be executed for this fft

100

cl_fft_kernel_info *kernel_info;

101

102

// number of kernels

103

int num_kernels;

104

105

// twist kernel for virtualizing fft of very large sizes that do not

106

// fit in GPU global memory

107

cl_kernel twist_kernel;

108

109

// flag indicating if temporary intermediate buffer is needed or not.

110

// this depends on fft kernels being executed and if transform is

111

// in-place or out-of-place. e.g. Local memory fft (say 1D 1024 ...

112

// one that does not require global transpose do not need temporary buffer)

113

// 2D 1024x1024 out-of-place fft however do require intermediate buffer.

114

// If temp buffer is needed, its allocation is lazy i.e. its not allocated

115

// until its needed

116

cl_int temp_buffer_needed;

117

118

// Batch size is runtime parameter and size of temporary buffer (if needed)

119

// depends on batch size. Allocation of temporary buffer is lazy i.e. its

120

// only created when needed. Once its created at first call of clFFT_Executexxx

121

// it is not allocated next time if next time clFFT_Executexxx is called with

122

// batch size different than the first call. last_batch_size caches the last

123

// batch size with which this plan is used so that we dont keep allocating/deallocating

124

// temp buffer if same batch size is used again and again.

125

size_t last_batch_size;

126

127

// temporary buffer for interleaved plan

128

cl_mem tempmemobj;

129

130

// temporary buffer for planner plan. Only one of tempmemobj or

131

// (tempmemobj_real, tempmemobj_imag) pair is valid (allocated) depending

132

// data format of plan (plannar or interleaved)

133

cl_mem tempmemobj_real, tempmemobj_imag;

134

135

// Maximum size of signal for which local memory transposed based

136

// fft is sufficient i.e. no global mem transpose (communication)

137

// is needed

138

size_t max_localmem_fft_size;

139

140

// Maximum work items per work group allowed. This, along with max_radix below controls

141

// maximum local memory being used by fft kernels of this plan. Set to 256 by default

142

size_t max_work_item_per_workgroup;

143

144

// Maximum base radix for local memory fft ... this controls the maximum register

145

// space used by work items. Currently defaults to 16

146

size_t max_radix;

147

148

// Device depended parameter that tells how many work-items need to be read consecutive

149

// values to make sure global memory access by work-items of a work-group result in

150

// coalesced memory access to utilize full bandwidth e.g. on NVidia tesla, this is 16

151

size_t min_mem_coalesce_width;

152

153

// Number of local memory banks. This is used to geneate kernel with local memory

154

// transposes with appropriate padding to avoid bank conflicts to local memory

155

// e.g. on NVidia it is 16.

156

size_t num_local_mem_banks;

157

}cl_fft_plan;

158

159

void FFT1D(cl_fft_plan *plan, cl_fft_kernel_dir dir);

160

161

#endif