potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 15:53:03 +04:00
Code Packages Releases Wiki Activity

Added DFT_SCALE for forward transforms

Browse Source
This commit is contained in:
Alexander Karsakov
2014-07-17 16:20:04 +04:00
parent 6c8b6bd0c7
commit b17bf031f6
4 changed files with 48 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/src/dxt.cpp
+10 -3
View File
@@ -2151,27 +2151,34 @@ struct OCL_FftPlan
size_t localsize[2];
String kernel_name;
bool is1d = (flags & DFT_ROWS) != 0 || dft_size == 1;
String options = buildOptions;
if (rows)
{
globalsize[0] = thread_count; globalsize[1] = dft_size;
localsize[0] = thread_count; localsize[1] = 1;
kernel_name = "fft_multi_radix_rows";
if (is1d && (flags & DFT_SCALE))
options += " -D DFT_SCALE";
}
else
{
globalsize[0] = dft_size; globalsize[1] = thread_count;
localsize[0] = 1; localsize[1] = thread_count;
kernel_name = "fft_multi_radix_cols";
if (flags & DFT_SCALE)
options += " -D DFT_SCALE";
}
bool is1d = (flags & DFT_ROWS) != 0 || dft_size == 1;
String options = buildOptions;
if (src.channels() == 1)
options += " -D REAL_INPUT";
if (dst.channels() == 1)
options += " -D CCS_OUTPUT";
if ((is1d && src.channels() == 1) || (rows && (flags & DFT_REAL_OUTPUT)))
options += " -D NO_CONJUGATE";
if (is1d)
options += " -D IS_1D";
ocl::Kernel k(kernel_name.c_str(), ocl::core::fft_oclsrc, options);
if (k.empty())
modules/core/src/opencl/fft.cl
+22 -9
View File
@@ -301,6 +301,12 @@ void fft_radix5(__local float2* smem, __constant const float2* twiddles, const i
barrier(CLK_LOCAL_MEM_FENCE);
}
#ifdef DFT_SCALE
#define VAL(x, scale) x*scale
#else
#define VAL(x, scale) x
#endif
__kernel void fft_multi_radix_rows(__global const uchar* src_ptr, int src_step, int src_offset, int src_rows, int src_cols,
__global uchar* dst_ptr, int dst_step, int dst_offset, int dst_rows, int dst_cols,
__constant float2 * twiddles_ptr, const int t, const int nz)
@@ -314,6 +320,11 @@ __kernel void fft_multi_radix_rows(__global const uchar* src_ptr, int src_step,
__constant const float2* twiddles = (__constant float2*) twiddles_ptr;
const int ind = x;
const int block_size = LOCAL_SIZE/kercn;
#ifdef IS_1D
float scale = 1.f/dst_cols;
#else
float scale = 1.f/(dst_cols*dst_rows);
#endif
#ifndef REAL_INPUT
__global const float2* src = (__global const float2*)(src_ptr + mad24(y, src_step, mad24(x, (int)(sizeof(float)*2), src_offset)));
@@ -341,15 +352,15 @@ __kernel void fft_multi_radix_rows(__global const uchar* src_ptr, int src_step,
__global float2* dst = (__global float2*)(dst_ptr + mad24(y, dst_step, dst_offset));
#pragma unroll
for (int i=x; i<cols; i+=block_size)
dst[i] = smem[i];
dst[i] = VAL(smem[i], scale);
#else
// pack row to CCS
__local float* smem_1cn = (__local float*) smem;
__global float* dst = (__global float*)(dst_ptr + mad24(y, dst_step, dst_offset));
for (int i=x; i<dst_cols-1; i+=block_size)
dst[i+1] = smem_1cn[i+2];
dst[i+1] = VAL(smem_1cn[i+2], scale);
if (x == 0)
dst[0] = smem_1cn[0];
dst[0] = VAL(smem_1cn[0], scale);
#endif
}
}
@@ -368,6 +379,8 @@ __kernel void fft_multi_radix_cols(__global const uchar* src_ptr, int src_step,
__constant const float2* twiddles = (__constant float2*) twiddles_ptr;
const int ind = y;
const int block_size = LOCAL_SIZE/kercn;
float scale = 1.f/(dst_rows*dst_cols);
#pragma unroll
for (int i=0; i<kercn; i++)
smem[y+i*block_size] = *((__global const float2*)(src + i*block_size*src_step));
@@ -380,7 +393,7 @@ __kernel void fft_multi_radix_cols(__global const uchar* src_ptr, int src_step,
__global uchar* dst = dst_ptr + mad24(y, dst_step, mad24(x, (int)(sizeof(float)*2), dst_offset));
#pragma unroll
for (int i=0; i<kercn; i++)
*((__global float2*)(dst + i*block_size*dst_step)) = smem[y + i*block_size];
*((__global float2*)(dst + i*block_size*dst_step)) = VAL(smem[y + i*block_size], scale);
#else
if (x == 0)
{
@@ -388,9 +401,9 @@ __kernel void fft_multi_radix_cols(__global const uchar* src_ptr, int src_step,
__local float* smem_1cn = (__local float*) smem;
__global uchar* dst = dst_ptr + mad24(y+1, dst_step, dst_offset);
for (int i=y; i<dst_rows-1; i+=block_size, dst+=dst_step*block_size)
*((__global float*) dst) = smem_1cn[i+2];
*((__global float*) dst) = VAL(smem_1cn[i+2], scale);
if (y == 0)
*((__global float*) (dst_ptr + dst_offset)) = smem_1cn[0];
*((__global float*) (dst_ptr + dst_offset)) = VAL(smem_1cn[0], scale);
}
else if (x == (dst_cols+1)/2)
{
@@ -398,16 +411,16 @@ __kernel void fft_multi_radix_cols(__global const uchar* src_ptr, int src_step,
__local float* smem_1cn = (__local float*) smem;
__global uchar* dst = dst_ptr + mad24(dst_cols-1, (int)sizeof(float), mad24(y+1, dst_step, dst_offset));
for (int i=y; i<dst_rows-1; i+=block_size, dst+=dst_step*block_size)
*((__global float*) dst) = smem_1cn[i+2];
*((__global float*) dst) = VAL(smem_1cn[i+2], scale);
if (y == 0)
*((__global float*) (dst_ptr + mad24(dst_cols-1, (int)sizeof(float), dst_offset))) = smem_1cn[0];
*((__global float*) (dst_ptr + mad24(dst_cols-1, (int)sizeof(float), dst_offset))) = VAL(smem_1cn[0], scale);
}
else
{
__global uchar* dst = dst_ptr + mad24(x, (int)sizeof(float)*2, mad24(y, dst_step, dst_offset - (int)sizeof(float)));
#pragma unroll
for (int i=y; i<dst_rows; i+=block_size, dst+=block_size*dst_step)
vstore2(smem[i], 0, (__global float*) dst);
vstore2(VAL(smem[i], scale), 0, (__global float*) dst);
}
#endif
}
modules/core/test/ocl/test_dft.cpp
+10 -10
View File
@@ -62,7 +62,7 @@ namespace ocl {
////////////////////////////////////////////////////////////////////////////
// Dft
PARAM_TEST_CASE(Dft, cv::Size, OCL_FFT_TYPE, bool, bool)
PARAM_TEST_CASE(Dft, cv::Size, OCL_FFT_TYPE, bool, bool, bool)
{
cv::Size dft_size;
int dft_flags, depth, cn, dft_type;
@@ -88,15 +88,14 @@ PARAM_TEST_CASE(Dft, cv::Size, OCL_FFT_TYPE, bool, bool)
}
if (GET_PARAM(2))
dft_flags |= cv::DFT_ROWS;
//if (GET_PARAM(3))
// if (dft_type == C2C) dft_flags |= cv::DFT_INVERSE;
//if (GET_PARAM(3))
// dft_flags |= cv::DFT_SCALE;
dft_flags |= cv::DFT_ROWS;
if (GET_PARAM(3))
dft_flags |= cv::DFT_SCALE;
//if (GET_PARAM(4))
// dft_flags |= cv::DFT_INVERSE;
inplace = GET_PARAM(4);
inplace = GET_PARAM(3);
if (inplace && dft_type == 0)
inplace = 0;
is1d = (dft_flags & DFT_ROWS) != 0 || dft_size.height == 1;
}
@@ -123,7 +122,7 @@ OCL_TEST_P(Dft, Mat)
udst = udst(cv::Range(0, udst.rows), cv::Range(0, udst.cols/2 + 1));
}
Mat gpu = udst.getMat(ACCESS_READ);
//Mat gpu = udst.getMat(ACCESS_READ);
//std::cout << src << std::endl;
//std::cout << dst << std::endl;
//std::cout << gpu << std::endl;
@@ -193,6 +192,7 @@ OCL_INSTANTIATE_TEST_CASE_P(Core, Dft, Combine(Values(cv::Size(6, 4), cv::Size(5
cv::Size(512, 1), cv::Size(1280, 768)),
Values((OCL_FFT_TYPE) R2C, (OCL_FFT_TYPE) C2C, (OCL_FFT_TYPE) R2R, (OCL_FFT_TYPE) C2R),
Bool(), // DFT_ROWS
Bool(), // DFT_SCALE
Bool() // inplace
)
);