Multi-radix with kernel generation

This commit is contained in:
Alexander Karsakov
2014-06-26 10:09:15 +04:00
parent 8f8450793a
commit 5dd9263848
5 changed files with 667 additions and 37 deletions

View File

@@ -1960,7 +1960,7 @@ static void CL_CALLBACK oclCleanupCallback(cl_event e, cl_int, void *p)
}
static bool ocl_dft(InputArray _src, OutputArray _dst, int flags)
static bool ocl_dft_amdfft(InputArray _src, OutputArray _dst, int flags)
{
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
Size ssize = _src.size();
@@ -2029,12 +2029,257 @@ static bool ocl_dft(InputArray _src, OutputArray _dst, int flags)
#endif // HAVE_CLAMDFFT
namespace cv
{
#ifdef HAVE_OPENCL
static bool fft_radixN(InputArray _src, OutputArray _dst, int radix, int block_size, int nonzero_rows, int flags)
{
int N = _src.size().width;
if (N % radix)
return false;
UMat src = _src.getUMat();
UMat dst = _dst.getUMat();
int thread_count = N / radix;
size_t globalsize[2] = { thread_count, nonzero_rows };
String kernel_name = format("fft_radix%d", radix);
ocl::Kernel k(kernel_name.c_str(), ocl::core::fft_oclsrc, (flags & DFT_INVERSE) != 0 ? "-D INVERSE" : "");
if (k.empty())
return false;
k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnlyNoSize(dst), block_size, thread_count, nonzero_rows);
return k.run(2, globalsize, NULL, false);
}
static bool ocl_packToCCS(InputArray _buffer, OutputArray _dst, int flags)
{
UMat buffer = _buffer.getUMat();
UMat dst = _dst.getUMat();
buffer = buffer.reshape(1);
if ((flags & DFT_ROWS) == 0 && buffer.rows > 1)
{
// pack to CCS by rows
if (dst.cols > 2)
buffer.colRange(2, dst.cols + (dst.cols % 2)).copyTo(dst.colRange(1, dst.cols-1 + (dst.cols % 2)));
Mat dst_mat = dst.getMat(ACCESS_WRITE);
Mat buffer_mat = buffer.getMat(ACCESS_READ);
dst_mat.at<float>(0,0) = buffer_mat.at<float>(0,0);
dst_mat.at<float>(dst_mat.rows-1,0) = buffer_mat.at<float>(buffer.rows/2,0);
for (int i=1; i<dst_mat.rows-1; i+=2)
{
dst_mat.at<float>(i,0) = buffer_mat.at<float>((i+1)/2,0);
dst_mat.at<float>(i+1,0) = buffer_mat.at<float>((i+1)/2,1);
}
if (dst_mat.cols % 2 == 0)
{
dst_mat.at<float>(0,dst_mat.cols-1) = buffer_mat.at<float>(0,buffer.cols/2);
dst_mat.at<float>(dst_mat.rows-1,dst_mat.cols-1) = buffer_mat.at<float>(buffer.rows/2,buffer.cols/2);
for (int i=1; i<dst_mat.rows-1; i+=2)
{
dst_mat.at<float>(i,dst_mat.cols-1) = buffer_mat.at<float>((i+1)/2,buffer.cols/2);
dst_mat.at<float>(i+1,dst_mat.cols-1) = buffer_mat.at<float>((i+1)/2,buffer.cols/2+1);
}
}
}
else
{
// pack to CCS each row
buffer.colRange(0,1).copyTo(dst.colRange(0,1));
buffer.colRange(2, (dst.cols+1)).copyTo(dst.colRange(1, dst.cols));
}
return true;
}
static bool ocl_dft_C2C_row(InputArray _src, OutputArray _dst, int nonzero_rows, int flags)
{
int type = _src.type(), depth = CV_MAT_DEPTH(type), channels = CV_MAT_CN(type);
UMat src = _src.getUMat();
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
if (depth == CV_64F && !doubleSupport)
return false;
int factors[34];
int nf = DFTFactorize( src.cols, factors );
int n = 1;
int factor_index = 0;
String radix_processing;
int min_radix = INT_MAX;
// 1. 2^n transforms
if ( (factors[factor_index] & 1) == 0 )
{
for( ; n < factors[factor_index]; )
{
int radix = 2;
if (8*n <= factors[0])
radix = 8;
else if (4*n <= factors[0])
radix = 4;
radix_processing += format("fft_radix%d(smem,x,%d,%d);", radix, n, src.cols/radix);
min_radix = min(radix, min_radix);
n *= radix;
}
factor_index++;
}
// 2. all the other transforms
for( ; factor_index < nf; factor_index++ )
{
int radix = factors[factor_index];
radix_processing += format("fft_radix%d(smem,x,%d,%d);", radix, n, src.cols/radix);
min_radix = min(radix, min_radix);
n *= radix;
}
UMat dst = _dst.getUMat();
int thread_count = src.cols / min_radix;
size_t globalsize[2] = { thread_count, nonzero_rows };
size_t localsize[2] = { thread_count, 1 };
String buildOptions = format("-D LOCAL_SIZE=%d -D kercn=%d -D RADIX_PROCESS=%s",
src.cols, src.cols/thread_count, radix_processing.c_str());
ocl::Kernel k("fft_multi_radix", ocl::core::fft_oclsrc, buildOptions);
if (k.empty())
return false;
k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnlyNoSize(dst), thread_count, nonzero_rows);
return k.run(2, globalsize, localsize, false);
}
static bool ocl_dft(InputArray _src, OutputArray _dst, int flags, int nonzero_rows)
{
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
Size ssize = _src.size();
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
if ( (!doubleSupport && depth == CV_64F) ||
!(type == CV_32FC1 || type == CV_32FC2 || type == CV_64FC1 || type == CV_64FC2))
return false;
// if is not a multiplication of prime numbers { 2, 3, 5 }
if (ssize.area() != getOptimalDFTSize(ssize.area()))
return false;
UMat src = _src.getUMat();
int complex_input = cn == 2 ? 1 : 0;
int complex_output = (flags & DFT_COMPLEX_OUTPUT) != 0;
int real_input = cn == 1 ? 1 : 0;
int real_output = (flags & DFT_REAL_OUTPUT) != 0;
bool inv = (flags & DFT_INVERSE) != 0 ? 1 : 0;
bool is1d = (flags & DFT_ROWS) != 0 || src.rows == 1;
// if output format is not specified
if (complex_output + real_output == 0)
{
if (!inv)
{
if (real_input)
real_output = 1;
else
complex_output = 1;
}
}
if (complex_output)
{
//if (is1d)
// _dst.create(Size(src.cols/2+1, src.rows), CV_MAKE_TYPE(depth, 2));
//else
_dst.create(src.size(), CV_MAKE_TYPE(depth, 2));
}
else
_dst.create(src.size(), CV_MAKE_TYPE(depth, 1));
UMat dst = _dst.getUMat();
bool inplace = src.u == dst.u;
//UMat buffer;
//if (complex_input)
//{
// if (inplace)
// buffer = src;
// else
// src.copyTo(buffer);
//}
//else
//{
// if (!inv)
// {
// // in case real input convert it to complex
// buffer.create(src.size(), CV_MAKE_TYPE(depth, 2));
// std::vector<UMat> planes;
// planes.push_back(src);
// planes.push_back(UMat::zeros(src.size(), CV_32F));
// merge(planes, buffer);
// }
// else
// {
// // TODO: unpack from CCS format
// }
//}
if( nonzero_rows <= 0 || nonzero_rows > _src.rows() )
nonzero_rows = _src.rows();
if (!ocl_dft_C2C_row(src, dst, nonzero_rows, flags))
return false;
if ((flags & DFT_ROWS) == 0 && nonzero_rows > 1)
{
transpose(dst, dst);
if (!ocl_dft_C2C_row(dst, dst, dst.rows, flags))
return false;
transpose(dst, dst);
}
//if (complex_output)
//{
// if (real_input && is1d)
// _dst.assign(buffer.colRange(0, buffer.cols/2+1));
// else
// _dst.assign(buffer);
//}
//else
//{
// if (!inv)
// ocl_packToCCS(buffer, _dst, flags);
// else
// {
// // copy real part to dst
// }
//}
return true;
}
#endif
} // namespace cv;
void cv::dft( InputArray _src0, OutputArray _dst, int flags, int nonzero_rows )
{
#ifdef HAVE_CLAMDFFT
CV_OCL_RUN(ocl::haveAmdFft() && ocl::Device::getDefault().type() != ocl::Device::TYPE_CPU &&
_dst.isUMat() && _src0.dims() <= 2 && nonzero_rows == 0,
ocl_dft(_src0, _dst, flags))
ocl_dft_amdfft(_src0, _dst, flags))
#endif
#ifdef HAVE_OPENCL
CV_OCL_RUN(_dst.isUMat() && _src0.dims() <= 2,
ocl_dft(_src0, _dst, flags, nonzero_rows))
#endif
static DFTFunc dft_tbl[6] =
@@ -2046,10 +2291,8 @@ void cv::dft( InputArray _src0, OutputArray _dst, int flags, int nonzero_rows )
(DFTFunc)RealDFT_64f,
(DFTFunc)CCSIDFT_64f
};
AutoBuffer<uchar> buf;
void *spec = 0;
Mat src0 = _src0.getMat(), src = src0;
int prev_len = 0, stage = 0;
bool inv = (flags & DFT_INVERSE) != 0;
@@ -2058,6 +2301,7 @@ void cv::dft( InputArray _src0, OutputArray _dst, int flags, int nonzero_rows )
int elem_size = (int)src.elemSize1(), complex_elem_size = elem_size*2;
int factors[34];
bool inplace_transform = false;
bool is1d = (flags & DFT_ROWS) != 0 || src.rows == 1;
#ifdef USE_IPP_DFT
AutoBuffer<uchar> ippbuf;
int ipp_norm_flag = !(flags & DFT_SCALE) ? 8 : inv ? 2 : 1;
@@ -2066,7 +2310,10 @@ void cv::dft( InputArray _src0, OutputArray _dst, int flags, int nonzero_rows )
CV_Assert( type == CV_32FC1 || type == CV_32FC2 || type == CV_64FC1 || type == CV_64FC2 );
if( !inv && src.channels() == 1 && (flags & DFT_COMPLEX_OUTPUT) )
_dst.create( src.size(), CV_MAKETYPE(depth, 2) );
if (!is1d)
_dst.create( src.size(), CV_MAKETYPE(depth, 2) );
else
_dst.create( Size(src.cols/2+1, src.rows), CV_MAKETYPE(depth, 2) );
else if( inv && src.channels() == 2 && (flags & DFT_REAL_OUTPUT) )
_dst.create( src.size(), depth );
else

View File

@@ -0,0 +1,297 @@
__constant float PI = 3.14159265f;
__constant float SQRT_2 = 0.707106781188f;
__constant float sin_120 = 0.866025403784f;
__constant float fft5_2 = 0.559016994374f;
__constant float fft5_3 = -0.951056516295f;
__constant float fft5_4 = -1.538841768587f;
__constant float fft5_5 = 0.363271264002f;
inline float2 mul_float2(float2 a, float2 b){
float2 res;
res.x = a.x * b.x - a.y * b.y;
res.y = a.x * b.y + a.y * b.x;
return res;
}
inline float2 sincos_float2(float alpha) {
float cs, sn;
sn = sincos(alpha, &cs); // sincos
return (float2)(cs, sn);
}
inline float2 twiddle(float2 a) {
return (float2)(a.y, -a.x);
}
inline float2 square(float2 a) {
return (float2)(a.x * a.x - a.y * a.y, 2.0f * a.x * a.y);
}
inline float2 square3(float2 a) {
return (float2)(a.x * a.x - a.y * a.y, 3.0f * a.x * a.y);
}
inline float2 mul_p1q4(float2 a) {
return (float2)(SQRT_2) * (float2)(a.x + a.y, -a.x + a.y);
}
inline float2 mul_p3q4(float2 a) {
return (float2)(SQRT_2) * (float2)(-a.x + a.y, -a.x - a.y);
}
__attribute__((always_inline))
void fft_radix2(__local float2* smem, const int x, const int block_size, const int t)
{
const int k = x & (block_size - 1);
float2 in1, temp;
if (x < t)
{
in1 = smem[x];
float2 in2 = smem[x+t];
float theta = -PI * k / block_size;
float cs;
float sn = sincos(theta, &cs);
temp = (float2) (in2.x * cs - in2.y * sn,
in2.y * cs + in2.x * sn);
}
barrier(CLK_LOCAL_MEM_FENCE);
if (x < t)
{
const int dst_ind = (x << 1) - k;
smem[dst_ind] = in1 + temp;
smem[dst_ind+block_size] = in1 - temp;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
__attribute__((always_inline))
void fft_radix4(__local float2* smem, const int x, const int block_size, const int t)
{
const int k = x & (block_size - 1);
float2 b0, b1, b2, b3;
if (x < t)
{
float theta = -PI * k / (2 * block_size);
float2 tw = sincos_float2(theta);
float2 a0 = smem[x];
float2 a1 = mul_float2(tw, smem[x+t]);
float2 a2 = smem[x + 2*t];
float2 a3 = mul_float2(tw, smem[x + 3*t]);
tw = square(tw);
a2 = mul_float2(tw, a2);
a3 = mul_float2(tw, a3);
b0 = a0 + a2;
b1 = a0 - a2;
b2 = a1 + a3;
b3 = twiddle(a1 - a3);
}
barrier(CLK_LOCAL_MEM_FENCE);
if (x < t)
{
const int dst_ind = ((x - k) << 2) + k;
smem[dst_ind] = b0 + b2;
smem[dst_ind + block_size] = b1 + b3;
smem[dst_ind + 2*block_size] = b0 - b2;
smem[dst_ind + 3*block_size] = b1 - b3;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
__attribute__((always_inline))
void fft_radix8(__local float2* smem, const int x, const int block_size, const int t)
{
const int k = x % block_size;
float2 a0, a1, a2, a3, a4, a5, a6, a7;
if (x < t)
{
float theta = -PI * k / (4 * block_size);
float2 tw = sincos_float2(theta); // W
a0 = smem[x];
a1 = mul_float2(tw, smem[x + t]);
a2 = smem[x + 2 * t];
a3 = mul_float2(tw, smem[x + 3 * t]);
a4 = smem[x + 4 * t];
a5 = mul_float2(tw, smem[x + 5 * t]);
a6 = smem[x + 6 * t];
a7 = mul_float2(tw, smem[x + 7 * t]);
tw = square(tw); // W^2
a2 = mul_float2(tw, a2);
a3 = mul_float2(tw, a3);
a6 = mul_float2(tw, a6);
a7 = mul_float2(tw, a7);
tw = square(tw); // W^4
a4 = mul_float2(tw, a4);
a5 = mul_float2(tw, a5);
a6 = mul_float2(tw, a6);
a7 = mul_float2(tw, a7);
float2 b0 = a0 + a4;
float2 b4 = a0 - a4;
float2 b1 = a1 + a5;
float2 b5 = mul_p1q4(a1 - a5);
float2 b2 = a2 + a6;
float2 b6 = twiddle(a2 - a6);
float2 b3 = a3 + a7;
float2 b7 = mul_p3q4(a3 - a7);
a0 = b0 + b2;
a2 = b0 - b2;
a1 = b1 + b3;
a3 = twiddle(b1 - b3);
a4 = b4 + b6;
a6 = b4 - b6;
a5 = b5 + b7;
a7 = twiddle(b5 - b7);
}
barrier(CLK_LOCAL_MEM_FENCE);
if (x < t)
{
const int dst_ind = ((x - k) << 3) + k;
__local float2* dst = smem + dst_ind;
dst[0] = a0 + a1;
dst[block_size] = a4 + a5;
dst[2 * block_size] = a2 + a3;
dst[3 * block_size] = a6 + a7;
dst[4 * block_size] = a0 - a1;
dst[5 * block_size] = a4 - a5;
dst[6 * block_size] = a2 - a3;
dst[7 * block_size] = a6 - a7;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
__attribute__((always_inline))
void fft_radix3(__local float2* smem, const int x, const int block_size, const int t)
{
const int k = x % block_size;
float2 a0, a1, a2, b0, b1;
if (x < t)
{
const float theta = -PI * k * 2 / (3 * block_size);
a0 = smem[x];
a1 = mul_float2(sincos_float2(theta), smem[x+t]);
a2 = mul_float2(sincos_float2(2 * theta), smem[x+2*t]);
b1 = a1 + a2;
a2 = twiddle((float2)sin_120*(a1 - a2));
b0 = a0 - (float2)(0.5f)*b1;
}
barrier(CLK_LOCAL_MEM_FENCE);
if (x < t)
{
const int dst_ind = ((x - k) * 3) + k;
smem[dst_ind] = a0 + b1;
smem[dst_ind + block_size] = b0 + a2;
smem[dst_ind + 2*block_size] = b0 - a2;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
__attribute__((always_inline))
void fft_radix5(__local float2* smem, const int x, const int block_size, const int t)
{
const int k = x % block_size;
float2 a0, a1, a2, a3, a4, b0, b1, b2, b5;
if (x < t)
{
const float theta = -PI * k * 2 / (5 * block_size);
a0 = smem[x];
a1 = mul_float2(sincos_float2(theta), smem[x + t]);
a2 = mul_float2(sincos_float2(theta*2),smem[x+2*t]);
a3 = mul_float2(sincos_float2(theta*3),smem[x+3*t]);
a4 = mul_float2(sincos_float2(theta*4),smem[x+4*t]);
b1 = a1 + a4;
a1 -= a4;
a4 = a3 + a2;
a3 -= a2;
b2 = b1 + a4;
b0 = a0 - (float2)0.25f * b2;
b1 = (float2)fft5_2 * (b1 - a4);
a4 = -(float2)fft5_3 * (a1 + a3);
a4 = twiddle(a4);
b5 = (float2)(a4.x - fft5_5 * a1.y, a4.y + fft5_5 * a1.x);
a4.x += fft5_4 * a3.y;
a4.y -= fft5_4 * a3.x;
a1 = b0 + b1;
b0 -= b1;
}
barrier(CLK_LOCAL_MEM_FENCE);
if (x < t)
{
const int dst_ind = ((x - k) * 5) + k;
__local float2* dst = smem + dst_ind;
dst[0] = a0 + b2;
dst[block_size] = a1 + a4;
dst[2 * block_size] = b0 + b5;
dst[3 * block_size] = b0 - b5;
dst[4 * block_size] = a1 - a4;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
__kernel void fft_multi_radix(__global const uchar* srcptr, int src_step, int src_offset,
__global uchar* dstptr, int dst_step, int dst_offset,
const int t, const int nz)
{
const int x = get_global_id(0);
const int y = get_group_id(1);
if (y < nz)
{
__local float2 smem[LOCAL_SIZE];
__global const float2* src = (__global const float2*)(srcptr + mad24(y, src_step, mad24(x, (int)(sizeof(float)*2), src_offset)));
__global float2* dst = (__global float2*)(dstptr + mad24(y, dst_step, mad24(x, (int)(sizeof(float)*2), dst_offset)));
const int block_size = LOCAL_SIZE/kercn;
#pragma unroll
for (int i=0; i<kercn; i++)
smem[x+i*block_size] = src[i*block_size];
barrier(CLK_LOCAL_MEM_FENCE);
RADIX_PROCESS;
// copy data to dst
#pragma unroll
for (int i=0; i<kercn; i++)
dst[i*block_size] = smem[x + i*block_size];
}
}