ISCE_INSAR/components/zerodop/GPUresampslc/cuda/GPUresamp.cu

370 lines
15 KiB
Plaintext

//
// Author: Joshua Cohen
// Copyright 2017
//
#include <cuda_runtime.h>
#include <cuComplex.h>
#include <complex.h>
#include <math.h>
#include <stdio.h>
#include <sys/time.h>
#define SINC_SUB 8192
#define SINC_LEN 8
#define SINC_HALF (SINC_LEN/2)
#define SINC_ONE (SINC_LEN+1)
#define IDX1D(i,j,w) (((i)*(w))+(j))
#define modulo_f(a,b) fmod(fmod(a,b)+(b),(b))
struct InputData {
cuFloatComplex *imgIn;
cuFloatComplex *imgOut;
double *residAz;
double *residRg;
double *azOffPoly;
double *rgOffPoly;
double *dopPoly;
double *azCarrierPoly;
double *rgCarrierPoly;
float *fintp;
};
__constant__ double ind[6];
__constant__ int ini[8];
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
// GPU Helper Functions
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
// Data usage: 8 doubles/pointers, 2 ints -- 72 bytes/call
__device__ double evalPolyAt(double *polyArr, double azi, double rng) {
// C-style eval method of Poly2d (adjusted to work with the array-format Poly2d where:
// polyArr[0] = azimuthOrder
// polyArr[1] = rangeOrder
// polyArr[2] = azimuthMean
// polyArr[3] = rangeMean
// polyArr[4] = azimuthNorm
// polyArr[5] = rangeNorm
// polyArr[6...] = coeffs (len ([0]+1)*([1]+1))
// Therefore we can guarantee that polyArr has at least 7 elements, and intuitively stores its own length using the orders
double val, scalex, scaley, xval, yval;
int i, j;
val = 0.;
scaley = 1.;
xval = (rng - polyArr[3]) / polyArr[5];
yval = (azi - polyArr[2]) / polyArr[4];
for (i=0; i<=polyArr[0]; i++,scaley*=yval) {
scalex = 1.;
for (j=0; j<=polyArr[1]; j++,scalex*=xval) {
val += scalex * scaley * polyArr[IDX1D(i,j,int(polyArr[1])+1)+6];
}
}
return val;
}
// Data usage: 4 doubles, 1 int -- 36 bytes/call
__device__ cuFloatComplex fintp(int idx) {
// Replaces storing ~4MB and copying in/out of constant storage by calculating an element of fintp on the fly
double weight, filtFact, sinFact, i;
i = int(int(idx / SINC_LEN) + ((idx % SINC_LEN) * (SINC_SUB / 2)) - 1) - 16383.5;
weight = .5 + (.5 * cos((M_PI * i) / 16383.5));
sinFact = i * (.75 / (SINC_SUB / 2));
filtFact = (sin(M_PI * sinFact) / (M_PI * sinFact)) * weight; // No need to check if sinFact is 0 since SINC_SUB != 0 and i-16383.5 != 0 as i is an int
return make_cuFloatComplex(filtFact*weight, 0.);
}
// Data usage: 6 double/complex/pointers, 4 ints -- 64 bytes/call
// Add call to fintp (36 bytes) -- 100 bytes/call
__device__ cuFloatComplex sinc_interp(cuFloatComplex *chip, double fraca, double fracr, double dop, float *fintp) {
// This is a modified/hardwired form of sinc_eval_2d from Interpolator. We eliminate a couple of checks that we know will pass, and
// adjusted the algorithm to account for modifications to the main kernel below. Primarily three things are of interest:
// 1. Chip is actually a pointer to the top-left of the chip location in the main image block, so chip's 'width' actually is
// ini[1], not SINC_ONE.
// 2. We account for removing doppler effects using cval and taking in dop. In the older version of the main kernel, this
// value was calculated and multiplied as the data was copied into the smaller chip. Here we calculate it on the fly using
// the same index for 'ii' as the row index of the chip being operated on (so in this case since it happens from the bottom
// up, we use SINC_LEN-i instead of i).
cuFloatComplex ret = make_cuFloatComplex(0.,0.);
cuFloatComplex cval;
int ifracx, ifracy, i, j;
ifracx = min(max(0,int(fraca*SINC_SUB)), SINC_SUB-1);
ifracy = min(max(0,int(fracr*SINC_SUB)), SINC_SUB-1);
for (i=0; i<SINC_LEN; i++) {
cval = make_cuFloatComplex(cos((SINC_LEN-i-4.)*dop), -sin((SINC_LEN-i-4.)*dop));
for (j=0; j<SINC_LEN; j++) {
ret = cuCaddf(ret,
cuCmulf(
cuCmulf(chip[IDX1D(SINC_LEN-i,SINC_LEN-j,ini[1])], cval),
make_cuFloatComplex(fintp[IDX1D(ifracx,i,SINC_LEN)]*fintp[IDX1D(ifracy,j,SINC_LEN)], 0.)
)
);
}
}
return ret;
}
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
// GPU Main Kernel
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
// Data Usage: 15 pointers/doubles, 5 ints, 1 bool -- 144 bytes/call (assuming 1 bool ==> 1 int)
// Add call to sinc_interp (100 bytes/call) -- 244 bytes/call (for funsies let's assume ~250 bytes/call)
// NOTE: We ignore calls to evalPolyAt since they have less
// data usage and therefore do not really matter for
// max data usage
__global__ void GPUResamp(struct InputData inData) {
// Main GPU ResampSlc kernel, slightly modified from original algorithm to save significant space
int pix = (blockDim.x * blockIdx.x) + threadIdx.x;
if (pix < (ini[2] * ini[6])) {
//cuFloatComplex chip[SINC_ONE*SINC_ONE];
cuFloatComplex cval;
double ao, ro, fracr, fraca, ph, dop;
int k, kk, idxi, idxj; //chipi, chipj, ii, jj;
bool flag;
flag = false;
idxi = (pix / ini[2]) + ini[4];
idxj = (pix % ini[2]);
ao = evalPolyAt(inData.azOffPoly, idxi+1, idxj+1);
ro = evalPolyAt(inData.rgOffPoly, idxi+1, idxj+1);
k = idxi + ao;
fraca = idxi + ao - k;
if ((k < SINC_HALF) || (k >= (ini[0]-SINC_HALF))) flag = true;
if (!flag) {
kk = idxj + ro;
fracr = idxj + ro - kk;
if ((kk < SINC_HALF) || (kk >= (ini[1]-SINC_HALF))) flag = true;
if (!flag) {
dop = evalPolyAt(inData.dopPoly, idxi+1, idxj+1);
/*
for (ii=0; ii<SINC_ONE; ii++) {
chipi = k - ini[3] + ii - SINC_HALF;
cval = make_cuFloatComplex(cos((ii-4.)*dop), -sin((ii-4.)*dop));
for (jj=0; jj<SINC_ONE; jj++) {
chipj = kk + jj - SINC_HALF;
chip[IDX1D(ii,jj,SINC_ONE)] = cuCmulf(inData.imgIn[IDX1D(chipi,chipj,ini[1])], cval);
}
}
top = k - ini[3] - SINC_HALF;
left = kk - SINC_HALF;
topLeft = IDX1D(top,left,ini[1]);
*/
ph = (dop * fraca) + evalPolyAt(inData.rgCarrierPoly, idxi+ao, idxj+ro) + evalPolyAt(inData.azCarrierPoly, idxi+ao, idxj+ro);
if (ini[7] == 1)
ph = ph + ((4.*(M_PI/ind[0]))*((ind[2]-ind[3])+(idxj*(ind[4]-ind[5]))+(ro*ind[4])))+((4.*M_PI*(ind[3]+(idxj*ind[5])))*((1./ind[1])-(1./ind[0])));
ph = modulo_f(ph, 2.*M_PI);
// NOTE: This has been modified to pass the pointer to the "top left" location of what used to be the 'chip' copy of data from imgIn. This
// saves allocating 81 extra floats per pixel (since chip is SINC_ONE*SINC_ONE), and instead uses logic to determine the offsets. The
// logic simply takes the minimum values of chipi and chipj above and adds the IDX1D index to the pointer. This means that sinc_interp
// will pass this pointer as "chip" and it will still point to the correct values (since chip is just a 2D window subset of imgIn). We
// Also have to account for 'cval' by modifying the sinc_interp function to calculate 'cval' dynamically (no extra cost computationally).
// This should actually speed up the kernel as well as significantly reduce redundant data usage.
cval = sinc_interp((inData.imgIn + IDX1D(k-ini[3]-SINC_HALF,kk-SINC_HALF,ini[1])), fraca, fracr, dop, inData.fintp);
inData.imgOut[pix] = cuCmulf(cval, make_cuFloatComplex(cos(ph), sin(ph)));
}
}
}
}
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
// CPU Helper Functions
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
double cpuSecond() {
struct timeval tp;
gettimeofday(&tp,NULL);
return (double(tp.tv_sec) + double(tp.tv_usec)*1.e-6);
}
void checkKernelErrors() {
cudaError_t errSync = cudaGetLastError();
cudaError_t errAsync = cudaDeviceSynchronize();
if (errSync != cudaSuccess) printf("\nSync kernel error: %s\n", cudaGetErrorString(errSync));
if (errAsync != cudaSuccess) printf("\nAsync kernel error: %s\n", cudaGetErrorString(errAsync));
}
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
// Main CPU Function
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
void runGPUResamp(double *h_inpts_dbl, int *h_inpts_int, void *imgIn, void *imgOut, double *residAz, double *residRg, double *azOffPoly, double *rgOffPoly,
double *dopPoly, double *azCarrierPoly, double *rgCarrierPoly, float *fintp) {
/* * * * * * * * * * * * * * * * * * * *
* Input mapping -
*
* Double 0 - wvl
* Double 1 - refwvl
* Double 2 - r0
* Double 3 - refr0
* Double 4 - slr
* Double 5 - refslr
*
* Int 0 - inLength
* Int 1 - inWidth
* Int 2 - outWidth
* Int 3 - firstImageRow
* Int 4 - firstTileRow
* Int 5 - nRowsInBlock
* Int 6 - LINES_PER_TILE
* Int 7 - flatten
*
* * * * * * * * * * * * * * * * * * * */
// Casting input/output images to native cuFloatComplex type from complex<float>
cuFloatComplex *h_imgIn = (cuFloatComplex *)imgIn;
cuFloatComplex *h_imgOut = (cuFloatComplex *)imgOut;
// Create handles for device copies of inputs
cuFloatComplex *d_imgIn, *d_imgOut;
double *d_residAz, *d_residRg, *d_azOffPoly, *d_rgOffPoly, *d_dopPoly, *d_azCarrierPoly, *d_rgCarrierPoly;
float *d_fintp;
double startRun, endRun, startKernel, endKernel;
struct InputData inData;
printf("\n Initializing GPU ResampSlc\n");
cudaSetDevice(0);
startRun = cpuSecond();
printf(" Allocating initial memory... ");
fflush(stdout);
int nInPix = h_inpts_int[5] * h_inpts_int[1];
int nOutPix = h_inpts_int[6] * h_inpts_int[2];
int nResidAzPix = 0;
if (residAz != 0) nResidAzPix = h_inpts_int[6] * h_inpts_int[2];
int nResidRgPix = 0;
if (residRg != 0) nResidRgPix = h_inpts_int[6] * h_inpts_int[2];
int nAzOffPix = ((azOffPoly[0]+1) * (azOffPoly[1]+1)) + 6; // [0] and [1] of the Poly2d arrays hold the az/rg orders
int nRgOffPix = ((rgOffPoly[0]+1) * (rgOffPoly[1]+1)) + 6;
int nDopPix = ((dopPoly[0]+1) * (dopPoly[1]+1)) + 6;
int nAzCarryPix = ((azCarrierPoly[0]+1) * (azCarrierPoly[1]+1)) + 6;
int nRgCarryPix = ((rgCarrierPoly[0]+1) * (rgCarrierPoly[1]+1)) + 6;
size_t nb_in = nInPix * sizeof(cuFloatComplex);
size_t nb_out = nOutPix * sizeof(cuFloatComplex);
size_t nb_rsdAz = nResidAzPix * sizeof(double);
size_t nb_rsdRg = nResidRgPix * sizeof(double);
size_t nb_azOff = nAzOffPix * sizeof(double);
size_t nb_rgOff = nRgOffPix * sizeof(double);
size_t nb_dop = nDopPix * sizeof(double);
size_t nb_azCarry = nAzCarryPix * sizeof(double);
size_t nb_rgCarry = nRgCarryPix * sizeof(double);
cudaMalloc((cuFloatComplex**)&d_imgIn, nb_in);
cudaMalloc((cuFloatComplex**)&d_imgOut, nb_out);
if (residAz != 0) cudaMalloc((double**)&d_residAz, nb_rsdAz);
if (residRg != 0) cudaMalloc((double**)&d_residRg, nb_rsdRg);
cudaMalloc((double**)&d_azOffPoly, nb_azOff);
cudaMalloc((double**)&d_rgOffPoly, nb_rgOff);
cudaMalloc((double**)&d_dopPoly, nb_dop);
cudaMalloc((double**)&d_azCarrierPoly, nb_azCarry);
cudaMalloc((double**)&d_rgCarrierPoly, nb_rgCarry);
cudaMalloc((float**)&d_fintp, (SINC_LEN*SINC_SUB*sizeof(float)));
printf("Done.\n Copying data to GPU... ");
fflush(stdout);
startKernel = cpuSecond();
cudaMemcpy(d_imgIn, h_imgIn, nb_in, cudaMemcpyHostToDevice);
if (residAz != 0) cudaMemcpy(d_residAz, residAz, nb_rsdAz, cudaMemcpyHostToDevice);
if (residRg != 0) cudaMemcpy(d_residRg, residRg, nb_rsdRg, cudaMemcpyHostToDevice);
cudaMemcpy(d_azOffPoly, azOffPoly, nb_azOff, cudaMemcpyHostToDevice);
cudaMemcpy(d_rgOffPoly, rgOffPoly, nb_rgOff, cudaMemcpyHostToDevice);
cudaMemcpy(d_dopPoly, dopPoly, nb_dop, cudaMemcpyHostToDevice);
cudaMemcpy(d_azCarrierPoly, azCarrierPoly, nb_azCarry, cudaMemcpyHostToDevice);
cudaMemcpy(d_rgCarrierPoly, rgCarrierPoly, nb_rgCarry, cudaMemcpyHostToDevice);
cudaMemcpy(d_fintp, fintp, (SINC_LEN*SINC_SUB*sizeof(float)), cudaMemcpyHostToDevice);
cudaMemcpyToSymbol(ind, h_inpts_dbl, (6*sizeof(double)));
cudaMemcpyToSymbol(ini, h_inpts_int, (8*sizeof(int)));
cudaMemset(d_imgOut, 0, nb_out);
endKernel = cpuSecond();
printf("Done. (%f s.)\n", (endKernel-startKernel));
dim3 block(32);
dim3 grid(int((nInPix + 31) / 32));
if ((grid.x * 32) > nInPix) printf(" (DEBUG: There will be %d 'empty' threads in the last thread block).\n", ((grid.x*32)-nInPix));
printf(" Running GPU ResampSlc... ");
fflush(stdout);
startKernel = cpuSecond();
inData.imgIn = d_imgIn;
inData.imgOut = d_imgOut;
inData.residAz = 0;
if (residAz != 0) inData.residAz = d_residAz;
inData.residRg = 0;
if (residRg != 0) inData.residRg = d_residRg;
inData.azOffPoly = d_azOffPoly;
inData.rgOffPoly = d_rgOffPoly;
inData.dopPoly = d_dopPoly;
inData.azCarrierPoly = d_azCarrierPoly;
inData.rgCarrierPoly = d_rgCarrierPoly;
inData.fintp = d_fintp;
GPUResamp <<<grid, block>>>(inData);
checkKernelErrors();
endKernel = cpuSecond();
printf("Done. (%f s.)\n", (endKernel-startKernel));
printf(" Copying memory back to host... ");
fflush(stdout);
startKernel = cpuSecond();
cudaMemcpy(h_imgOut, d_imgOut, nb_out, cudaMemcpyDeviceToHost);
endKernel = cpuSecond();
endRun = cpuSecond();
printf("Done. (%f s.)\n", (endKernel-startKernel));
printf(" Finished GPU ResampSlc in %f s.\n", (endRun-startRun));
printf(" Cleaning device memory and returning to main ResampSlc function...\n");
cudaFree(d_imgIn);
cudaFree(d_imgOut);
if (residAz != 0) cudaFree(d_residAz);
if (residRg != 0) cudaFree(d_residRg);
cudaFree(d_azOffPoly);
cudaFree(d_rgOffPoly);
cudaFree(d_dopPoly);
cudaFree(d_azCarrierPoly);
cudaFree(d_rgCarrierPoly);
cudaFree(d_fintp);
cudaDeviceReset();
printf(" Exiting GPU ResampSlc\n\n");
}