aboutsummaryrefslogtreecommitdiffstats
path: root/src/com/jogamp/opencl/demos/fft/CLFFTPlan.java
diff options
context:
space:
mode:
authorMichael Bien <[email protected]>2010-08-31 01:25:59 +0200
committerMichael Bien <[email protected]>2010-08-31 01:25:59 +0200
commit4455385dd411375345688685dd561652708a7024 (patch)
tree1ddd2240d94e8c0571b29562bde411fb72fb2c82 /src/com/jogamp/opencl/demos/fft/CLFFTPlan.java
parent38641e01d3540e279ee91c9c6cfbc28414a6ed8d (diff)
initial import of Michael Zucchi's port of Apples FFT example (RFE 408).
small modifications: - pick CLPlatform with a GPU - replaced tabs with spaces :)
Diffstat (limited to 'src/com/jogamp/opencl/demos/fft/CLFFTPlan.java')
-rw-r--r--src/com/jogamp/opencl/demos/fft/CLFFTPlan.java1968
1 files changed, 1968 insertions, 0 deletions
diff --git a/src/com/jogamp/opencl/demos/fft/CLFFTPlan.java b/src/com/jogamp/opencl/demos/fft/CLFFTPlan.java
new file mode 100644
index 0000000..f35ce14
--- /dev/null
+++ b/src/com/jogamp/opencl/demos/fft/CLFFTPlan.java
@@ -0,0 +1,1968 @@
+
+/*
+ * sample is based on Apple's FFT example.
+ * initial port to JOCL Copyright 2010 Michael Zucchi
+ *
+ * TODO: The execute functions may allocate/use temporary memory per call hence they are
+ * neither thread safe nor multiple-queue safe. Perhaps some per-queue allocation
+ * system would suffice.
+ * TODO: The dynamic device-dependent variables should be dynamic and device-dependent and not
+ * hardcoded. Where possible.
+ * TODO: CPU support?
+ */
+package com.jogamp.opencl.demos.fft;
+
+import com.jogamp.opencl.CLBuffer;
+import com.jogamp.opencl.CLCommandQueue;
+import com.jogamp.opencl.CLContext;
+import com.jogamp.opencl.CLDevice;
+import com.jogamp.opencl.CLEventList;
+import com.jogamp.opencl.CLKernel;
+import com.jogamp.opencl.CLMemory;
+import com.jogamp.opencl.CLMemory.Mem;
+import com.jogamp.opencl.CLProgram;
+import java.io.OutputStream;
+import java.io.PrintStream;
+import java.nio.FloatBuffer;
+import java.util.LinkedList;
+
+/**
+ *
+ * @author notzed
+ */
+public class CLFFTPlan {
+
+ private class CLFFTDim3 {
+
+ int x;
+ int y;
+ int z;
+
+ CLFFTDim3(int x, int y, int z) {
+ this.x = x;
+ this.y = y;
+ this.z = z;
+ }
+
+ CLFFTDim3(int[] size) {
+ x = size[0];
+ y = size.length > 1 ? size[1] : 1;
+ z = size.length > 2 ? size[2] : 1;
+ }
+ }
+
+ private class WorkDimensions {
+
+ int batchSize;
+ long gWorkItems;
+ long lWorkItems;
+
+ public WorkDimensions(int batchSize, long gWorkItems, long lWorkItems) {
+ this.batchSize = batchSize;
+ this.gWorkItems = gWorkItems;
+ this.lWorkItems = lWorkItems;
+ }
+ }
+
+ private class fftPadding {
+
+ int lMemSize;
+ int offset;
+ int midPad;
+
+ public fftPadding(int lMemSize, int offset, int midPad) {
+ this.lMemSize = lMemSize;
+ this.offset = offset;
+ this.midPad = midPad;
+ }
+ }
+
+ class CLFFTKernelInfo {
+
+ CLKernel kernel;
+ String kernel_name;
+ int lmem_size;
+ int num_workgroups;
+ int num_xforms_per_workgroup;
+ int num_workitems_per_workgroup;
+ CLFFTKernelDir dir;
+ boolean in_place_possible;
+ };
+
+ public enum CLFFTDirection {
+
+ Forward {
+
+ int value() {
+ return -1;
+ }
+ },
+ Inverse {
+
+ int value() {
+ return 1;
+ }
+ };
+
+ abstract int value();
+ };
+
+ enum CLFFTKernelDir {
+ X,
+ Y,
+ Z
+ };
+
+ public enum CLFFTDataFormat {
+ SplitComplexFormat,
+ InterleavedComplexFormat,
+ }
+
+ // context in which fft resources are created and kernels are executed
+ CLContext context;
+ // size of signal
+ CLFFTDim3 size;
+ // dimension of transform ... must be either 1, 2 or 3
+ int dim;
+ // data format ... must be either interleaved or plannar
+ CLFFTDataFormat format;
+ // string containing kernel source. Generated at runtime based on
+ // size, dim, format and other parameters
+ StringBuilder kernel_string;
+ // CL program containing source and kernel this particular
+ // size, dim, data format
+ CLProgram program;
+ // linked list of kernels which needs to be executed for this fft
+ LinkedList<CLFFTKernelInfo> kernel_list;
+ // twist kernel for virtualizing fft of very large sizes that do not
+ // fit in GPU global memory
+ CLKernel twist_kernel;
+ // flag indicating if temporary intermediate buffer is needed or not.
+ // this depends on fft kernels being executed and if transform is
+ // in-place or out-of-place. e.g. Local memory fft (say 1D 1024 ...
+ // one that does not require global transpose do not need temporary buffer)
+ // 2D 1024x1024 out-of-place fft however do require intermediate buffer.
+ // If temp buffer is needed, its allocation is lazy i.e. its not allocated
+ // until its needed
+ boolean temp_buffer_needed;
+ // Batch size is runtime parameter and size of temporary buffer (if needed)
+ // depends on batch size. Allocation of temporary buffer is lazy i.e. its
+ // only created when needed. Once its created at first call of clFFT_Executexxx
+ // it is not allocated next time if next time clFFT_Executexxx is called with
+ // batch size different than the first call. last_batch_size caches the last
+ // batch size with which this plan is used so that we dont keep allocating/deallocating
+ // temp buffer if same batch size is used again and again.
+ int last_batch_size;
+ // temporary buffer for interleaved plan
+ CLMemory tempmemobj;
+ // temporary buffer for planner plan. Only one of tempmemobj or
+ // (tempmemobj_real, tempmemobj_imag) pair is valid (allocated) depending
+ // data format of plan (plannar or interleaved)
+ CLMemory tempmemobj_real, tempmemobj_imag;
+ // Maximum size of signal for which local memory transposed based
+ // fft is sufficient i.e. no global mem transpose (communication)
+ // is needed
+ int max_localmem_fft_size;
+ // Maximum work items per work group allowed. This, along with max_radix below controls
+ // maximum local memory being used by fft kernels of this plan. Set to 256 by default
+ int max_work_item_per_workgroup;
+ // Maximum base radix for local memory fft ... this controls the maximum register
+ // space used by work items. Currently defaults to 16
+ int max_radix;
+ // Device depended parameter that tells how many work-items need to be read consecutive
+ // values to make sure global memory access by work-items of a work-group result in
+ // coalesced memory access to utilize full bandwidth e.g. on NVidia tesla, this is 16
+ int min_mem_coalesce_width;
+ // Number of local memory banks. This is used to geneate kernel with local memory
+ // transposes with appropriate padding to avoid bank conflicts to local memory
+ // e.g. on NVidia it is 16.
+ int num_local_mem_banks;
+
+ public class InvalidContextException extends Exception {
+ }
+
+ /**
+ * Create a new FFT plan.
+ *
+ * Use the matching executeInterleaved() or executePlanar() depending on the dataFormat specified.
+ * @param context
+ * @param sizes Array of sizes for each dimension. The length of array defines how many dimensions there are.
+ * @param dataFormat Data format, InterleavedComplex (array of complex) or SplitComplex (separate planar arrays).
+ * @throws zephyr.cl.CLFFTPlan.InvalidContextException
+ */
+ public CLFFTPlan(CLContext context, int[] sizes, CLFFTDataFormat dataFormat) throws InvalidContextException {
+ int i;
+ int err;
+ boolean isPow2 = true;
+ String kString;
+ int num_devices;
+ boolean gpu_found = false;
+ CLDevice[] devices;
+ int ret_size;
+
+ if (sizes.length < 1 || sizes.length > 3) {
+ throw new IllegalArgumentException("Dimensions must be between 1 and 3");
+ }
+
+ this.size = new CLFFTDim3(sizes);
+
+ isPow2 |= (this.size.x != 0) && (((this.size.x - 1) & this.size.x) == 0);
+ isPow2 |= (this.size.y != 0) && (((this.size.y - 1) & this.size.y) == 0);
+ isPow2 |= (this.size.z != 0) && (((this.size.z - 1) & this.size.z) == 0);
+
+ if (!isPow2) {
+ throw new IllegalArgumentException("Sizes must be power of two");
+ }
+
+ //if( (dim == FFT_1D && (size.y != 1 || size.z != 1)) || (dim == FFT_2D && size.z != 1) )
+ // ERR_MACRO(CL_INVALID_VALUE);
+
+ this.context = context;
+ //clRetainContext(context);
+ //this.size = size;
+ this.dim = sizes.length;
+ this.format = dataFormat;
+ //this.kernel_list = 0;
+ //this.twist_kernel = 0;
+ //this.program = 0;
+ this.temp_buffer_needed = false;
+ this.last_batch_size = 0;
+ //this.tempmemobj = 0;
+ //this.tempmemobj_real = 0;
+ //this.tempmemobj_imag = 0;
+ this.max_localmem_fft_size = 2048;
+ this.max_work_item_per_workgroup = 256;
+ this.max_radix = 16;
+ this.min_mem_coalesce_width = 16;
+ this.num_local_mem_banks = 16;
+
+ boolean done = false;
+
+ // this seems pretty shit, can't it tell this before building it?
+ while (!done) {
+ kernel_list = new LinkedList<CLFFTKernelInfo>();
+
+ this.kernel_string = new StringBuilder();
+ getBlockConfigAndKernelString();
+
+ this.program = context.createProgram(kernel_string.toString());
+
+ devices = context.getDevices();
+ for (i = 0; i < devices.length; i++) {
+ CLDevice dev = devices[i];
+
+ if (dev.getType() == CLDevice.Type.GPU) {
+ gpu_found = true;
+ program.build("-cl-mad-enable", dev);
+ }
+ }
+
+ if (!gpu_found) {
+ throw new InvalidContextException();
+ }
+
+ createKernelList();
+
+ // we created program and kernels based on "some max work group size (default 256)" ... this work group size
+ // may be larger than what kernel may execute with ... if thats the case we need to regenerate the kernel source
+ // setting this as limit i.e max group size and rebuild.
+ if (getPatchingRequired(devices)) {
+ release();
+ this.max_work_item_per_workgroup = (int) getMaxKernelWorkGroupSize(devices);
+ } else {
+ done = true;
+ }
+ }
+ }
+
+ /**
+ * Release system resources.
+ */
+ public void release() {
+ for (CLFFTKernelInfo kInfo : kernel_list) {
+ kInfo.kernel.release();
+ }
+ program.release();
+ }
+
+ void allocateTemporaryBufferInterleaved(int batchSize) {
+ if (temp_buffer_needed && last_batch_size != batchSize) {
+ last_batch_size = batchSize;
+ int tmpLength = size.x * size.y * size.z * batchSize * 2 * 4; // sizeof(float)
+
+ if (tempmemobj != null) {
+ tempmemobj.release();
+ }
+
+ tempmemobj = context.createFloatBuffer(tmpLength, Mem.READ_WRITE);
+ }
+ }
+
+ /**
+ * Calculate FFT on interleaved complex data.
+ * @param queue
+ * @param batchSize How many instances to calculate. Use 1 for a single FFT.
+ * @param dir Direction of calculation, Forward or Inverse.
+ * @param data_in Input buffer.
+ * @param data_out Output buffer. May be the same as data_in for in-place transform.
+ * @param condition Condition to wait for. NOT YET IMPLEMENTED.
+ * @param event Event to wait for completion. NOT YET IMPLEMENTED.
+ */
+ public void executeInterleaved(CLCommandQueue queue, int batchSize, CLFFTDirection dir,
+ CLBuffer<FloatBuffer> data_in, CLBuffer<FloatBuffer> data_out,
+ CLEventList condition, CLEventList event) {
+ int s;
+ if (format != format.InterleavedComplexFormat) {
+ throw new IllegalArgumentException();
+ }
+
+ WorkDimensions wd;
+ boolean inPlaceDone = false;
+
+ boolean isInPlace = data_in == data_out;
+
+ allocateTemporaryBufferInterleaved(batchSize);
+
+ CLMemory[] memObj = new CLMemory[3];
+ memObj[0] = data_in;
+ memObj[1] = data_out;
+ memObj[2] = tempmemobj;
+ int numKernels = kernel_list.size();
+
+ boolean numKernelsOdd = (numKernels & 1) != 0;
+ int currRead = 0;
+ int currWrite = 1;
+
+ // at least one external dram shuffle (transpose) required
+ if (temp_buffer_needed) {
+ // in-place transform
+ if (isInPlace) {
+ inPlaceDone = false;
+ currRead = 1;
+ currWrite = 2;
+ } else {
+ currWrite = (numKernels & 1) == 1 ? 1 : 2;
+ }
+
+ for (CLFFTKernelInfo kernelInfo : kernel_list) {
+ if (isInPlace && numKernelsOdd && !inPlaceDone && kernelInfo.in_place_possible) {
+ currWrite = currRead;
+ inPlaceDone = true;
+ }
+
+ s = batchSize;
+ wd = getKernelWorkDimensions(kernelInfo, s);
+ kernelInfo.kernel.setArg(0, memObj[currRead]);
+ kernelInfo.kernel.setArg(1, memObj[currWrite]);
+ kernelInfo.kernel.setArg(2, dir.value());
+ kernelInfo.kernel.setArg(3, wd.batchSize);
+ queue.put2DRangeKernel(kernelInfo.kernel, 0, 0, wd.gWorkItems, 1, wd.lWorkItems, 1);
+ //queue.put1DRangeKernel(kernelInfo.kernel, 0, wd.gWorkItems, wd.lWorkItems);
+
+ //System.out.printf("execute %s size %d,%d batch %d, dir %d, currread %d currwrite %d\size", kernelInfo.kernel_name, wd.gWorkItems, wd.lWorkItems, wd.batchSize, dir.value(), currRead, currWrite);
+
+ currRead = (currWrite == 1) ? 1 : 2;
+ currWrite = (currWrite == 1) ? 2 : 1;
+ }
+ } else {
+ // no dram shuffle (transpose required) transform
+ // all kernels can execute in-place.
+ for (CLFFTKernelInfo kernelInfo : kernel_list) {
+ {
+ s = batchSize;
+ wd = getKernelWorkDimensions(kernelInfo, s);
+
+ kernelInfo.kernel.setArg(0, memObj[currRead]);
+ kernelInfo.kernel.setArg(1, memObj[currWrite]);
+ kernelInfo.kernel.setArg(2, dir.value());
+ kernelInfo.kernel.setArg(3, wd.batchSize);
+ queue.put2DRangeKernel(kernelInfo.kernel, 0, 0, wd.gWorkItems, 1, wd.lWorkItems, 1);
+
+ //System.out.printf("execute %s size %d,%d batch %d, currread %d currwrite %d\size", kernelInfo.kernel_name, wd.gWorkItems, wd.lWorkItems, wd.batchSize, currRead, currWrite);
+
+ currRead = 1;
+ currWrite = 1;
+ }
+ }
+ }
+ }
+
+ void allocateTemporaryBufferPlanar(int batchSize) {
+ if (temp_buffer_needed && last_batch_size != batchSize) {
+ last_batch_size = batchSize;
+ int tmpLength = size.x * size.y * size.z * batchSize * 4; //sizeof(cl_float);
+
+ if (tempmemobj_real != null) {
+ tempmemobj_real.release();
+ }
+
+ if (tempmemobj_imag != null) {
+ tempmemobj_imag.release();
+ }
+
+ tempmemobj_real = context.createFloatBuffer(tmpLength, Mem.READ_WRITE);
+ tempmemobj_imag = context.createFloatBuffer(tmpLength, Mem.READ_WRITE);
+ }
+ }
+
+ /**
+ * Calculate FFT of planar data.
+ * @param queue
+ * @param batchSize
+ * @param dir
+ * @param data_in_real
+ * @param data_in_imag
+ * @param data_out_real
+ * @param data_out_imag
+ * @param contition
+ * @param event
+ */
+ public void executePlanar(CLCommandQueue queue, int batchSize, CLFFTDirection dir,
+ CLBuffer<FloatBuffer> data_in_real, CLBuffer<FloatBuffer> data_in_imag, CLBuffer<FloatBuffer> data_out_real, CLBuffer<FloatBuffer> data_out_imag,
+ CLEventList contition, CLEventList event) {
+ int s;
+
+ if (format != format.SplitComplexFormat) {
+ throw new IllegalArgumentException();
+ }
+
+ int err;
+ WorkDimensions wd;
+ boolean inPlaceDone = false;
+
+ boolean isInPlace = ((data_in_real == data_out_real) && (data_in_imag == data_out_imag));
+
+ allocateTemporaryBufferPlanar(batchSize);
+
+ CLMemory[] memObj_real = new CLMemory[3];
+ CLMemory[] memObj_imag = new CLMemory[3];
+ memObj_real[0] = data_in_real;
+ memObj_real[1] = data_out_real;
+ memObj_real[2] = tempmemobj_real;
+ memObj_imag[0] = data_in_imag;
+ memObj_imag[1] = data_out_imag;
+ memObj_imag[2] = tempmemobj_imag;
+
+ int numKernels = kernel_list.size();
+
+ boolean numKernelsOdd = (numKernels & 1) == 1;
+ int currRead = 0;
+ int currWrite = 1;
+
+ // at least one external dram shuffle (transpose) required
+ if (temp_buffer_needed) {
+ // in-place transform
+ if (isInPlace) {
+ inPlaceDone = false;
+ currRead = 1;
+ currWrite = 2;
+ } else {
+ currWrite = (numKernels & 1) == 1 ? 1 : 2;
+ }
+
+ for (CLFFTKernelInfo kernelInfo : kernel_list) {
+ if (isInPlace && numKernelsOdd && !inPlaceDone && kernelInfo.in_place_possible) {
+ currWrite = currRead;
+ inPlaceDone = true;
+ }
+
+ s = batchSize;
+ wd = getKernelWorkDimensions(kernelInfo, s);
+
+ kernelInfo.kernel.setArg(0, memObj_real[currRead]);
+ kernelInfo.kernel.setArg(1, memObj_imag[currRead]);
+ kernelInfo.kernel.setArg(2, memObj_real[currWrite]);
+ kernelInfo.kernel.setArg(3, memObj_imag[currWrite]);
+ kernelInfo.kernel.setArg(4, dir.value());
+ kernelInfo.kernel.setArg(5, wd.batchSize);
+
+ queue.put1DRangeKernel(kernelInfo.kernel, 0, wd.gWorkItems, wd.lWorkItems);
+
+
+ currRead = (currWrite == 1) ? 1 : 2;
+ currWrite = (currWrite == 1) ? 2 : 1;
+
+ }
+ } // no dram shuffle (transpose required) transform
+ else {
+
+ for (CLFFTKernelInfo kernelInfo : kernel_list) {
+ s = batchSize;
+ wd = getKernelWorkDimensions(kernelInfo, s);
+
+ kernelInfo.kernel.setArg(0, memObj_real[currRead]);
+ kernelInfo.kernel.setArg(1, memObj_imag[currRead]);
+ kernelInfo.kernel.setArg(2, memObj_real[currWrite]);
+ kernelInfo.kernel.setArg(3, memObj_imag[currWrite]);
+ kernelInfo.kernel.setArg(4, dir.value());
+ kernelInfo.kernel.setArg(5, wd.batchSize);
+
+ queue.put1DRangeKernel(kernelInfo.kernel, 0, wd.gWorkItems, wd.lWorkItems);
+ currRead = 1;
+ currWrite = 1;
+ }
+ }
+ }
+
+ /**
+ * Dump the planner result to the output stream.
+ * @param os if null, System.out is used.
+ */
+ public void dumpPlan(OutputStream os) {
+ PrintStream out = os == null ? System.out : new PrintStream(os);
+
+ for (CLFFTKernelInfo kInfo : kernel_list) {
+ int s = 1;
+ WorkDimensions wd = getKernelWorkDimensions(kInfo, s);
+ out.printf("Run kernel %s with global dim = {%d*BatchSize}, local dim={%d}\n", kInfo.kernel_name, wd.gWorkItems, wd.lWorkItems);
+ }
+ out.printf("%s\n", kernel_string.toString());
+ }
+
+ WorkDimensions getKernelWorkDimensions(CLFFTKernelInfo kernelInfo, int batchSize) {
+ int lWorkItems = kernelInfo.num_workitems_per_workgroup;
+ int numWorkGroups = kernelInfo.num_workgroups;
+ int numXFormsPerWG = kernelInfo.num_xforms_per_workgroup;
+
+ switch (kernelInfo.dir) {
+ case X:
+ batchSize *= (size.y * size.z);
+ numWorkGroups = ((batchSize % numXFormsPerWG) != 0) ? (batchSize / numXFormsPerWG + 1) : (batchSize / numXFormsPerWG);
+ numWorkGroups *= kernelInfo.num_workgroups;
+ break;
+ case Y:
+ batchSize *= size.z;
+ numWorkGroups *= batchSize;
+ break;
+ case Z:
+ numWorkGroups *= batchSize;
+ break;
+ }
+
+ return new WorkDimensions(batchSize, numWorkGroups * lWorkItems, lWorkItems);
+ }
+
+ /*
+ *
+ * Kernel building/customisation code follows
+ *
+ */
+ private void getBlockConfigAndKernelString() {
+ this.temp_buffer_needed = false;
+ this.kernel_string.append(baseKernels);
+
+ if (this.format == CLFFTDataFormat.SplitComplexFormat) {
+ this.kernel_string.append(twistKernelPlannar);
+ } else {
+ this.kernel_string.append(twistKernelInterleaved);
+ }
+
+ switch (this.dim) {
+ case 1:
+ FFT1D(CLFFTKernelDir.X);
+ break;
+
+ case 2:
+ FFT1D(CLFFTKernelDir.X);
+ FFT1D(CLFFTKernelDir.Y);
+ break;
+
+ case 3:
+ FFT1D(CLFFTKernelDir.X);
+ FFT1D(CLFFTKernelDir.Y);
+ FFT1D(CLFFTKernelDir.Z);
+ break;
+
+ default:
+ return;
+ }
+
+ this.temp_buffer_needed = false;
+ for (CLFFTKernelInfo kInfo : this.kernel_list) {
+ this.temp_buffer_needed |= !kInfo.in_place_possible;
+ }
+ }
+
+ private void createKernelList() {
+ CLFFTKernelInfo kern;
+ for (CLFFTKernelInfo kinfo : this.kernel_list) {
+ kinfo.kernel = program.createCLKernel(kinfo.kernel_name);
+ }
+
+ if (format == format.SplitComplexFormat) {
+ twist_kernel = program.createCLKernel("clFFT_1DTwistSplit");
+ } else {
+ twist_kernel = program.createCLKernel("clFFT_1DTwistInterleaved");
+ }
+ }
+
+ private boolean getPatchingRequired(CLDevice[] devices) {
+ int i;
+ for (i = 0; i < devices.length; i++) {
+ for (CLFFTKernelInfo kInfo : kernel_list) {
+ if (kInfo.kernel.getWorkGroupSize(devices[i]) < kInfo.num_workitems_per_workgroup) {
+ return true;
+ }
+ }
+ }
+ return false;
+ }
+
+ long getMaxKernelWorkGroupSize(CLDevice[] devices) {
+ long max_wg_size = Integer.MAX_VALUE;
+ int i;
+
+ for (i = 0; i < devices.length; i++) {
+ for (CLFFTKernelInfo kInfo : kernel_list) {
+ long wg_size = kInfo.kernel.getWorkGroupSize(devices[i]);
+
+ if (max_wg_size > wg_size) {
+ max_wg_size = wg_size;
+ }
+ }
+ }
+
+ return max_wg_size;
+ }
+
+ int log2(int x) {
+ return 32 - Integer.numberOfLeadingZeros(x - 1);
+ }
+
+// For any size, this function decomposes size into factors for loacal memory tranpose
+// based fft. Factors (radices) are sorted such that the first one (radixArray[0])
+// is the largest. This base radix determines the number of registers used by each
+// work item and product of remaining radices determine the size of work group needed.
+// To make things concrete with and example, suppose size = 1024. It is decomposed into
+// 1024 = 16 x 16 x 4. Hence kernel uses float2 a[16], for local in-register fft and
+// needs 16 x 4 = 64 work items per work group. So kernel first performance 64 length
+// 16 ffts (64 work items working in parallel) following by transpose using local
+// memory followed by again 64 length 16 ffts followed by transpose using local memory
+// followed by 256 length 4 ffts. For the last step since with size of work group is
+// 64 and each work item can array for 16 values, 64 work items can compute 256 length
+// 4 ffts by each work item computing 4 length 4 ffts.
+// Similarly for size = 2048 = 8 x 8 x 8 x 4, each work group has 8 x 8 x 4 = 256 work
+// iterms which each computes 256 (in-parallel) length 8 ffts in-register, followed
+// by transpose using local memory, followed by 256 length 8 in-register ffts, followed
+// by transpose using local memory, followed by 256 length 8 in-register ffts, followed
+// by transpose using local memory, followed by 512 length 4 in-register ffts. Again,
+// for the last step, each work item computes two length 4 in-register ffts and thus
+// 256 work items are needed to compute all 512 ffts.
+// For size = 32 = 8 x 4, 4 work items first compute 4 in-register
+// lenth 8 ffts, followed by transpose using local memory followed by 8 in-register
+// length 4 ffts, where each work item computes two length 4 ffts thus 4 work items
+// can compute 8 length 4 ffts. However if work group size of say 64 is choosen,
+// each work group can compute 64/ 4 = 16 size 32 ffts (batched transform).
+// Users can play with these parameters to figure what gives best performance on
+// their particular device i.e. some device have less register space thus using
+// smaller base radix can avoid spilling ... some has small local memory thus
+// using smaller work group size may be required etc
+ int getRadixArray(int n, int[] radixArray, int maxRadix) {
+ if (maxRadix > 1) {
+ maxRadix = Math.min(n, maxRadix);
+ int cnt = 0;
+ while (n > maxRadix) {
+ radixArray[cnt++] = maxRadix;
+ n /= maxRadix;
+ }
+ radixArray[cnt++] = n;
+ return cnt;
+ }
+
+ switch (n) {
+ case 2:
+ radixArray[0] = 2;
+ return 1;
+
+ case 4:
+ radixArray[0] = 4;
+ return 1;
+
+ case 8:
+ radixArray[0] = 8;
+ return 1;
+
+ case 16:
+ radixArray[0] = 8;
+ radixArray[1] = 2;
+ return 2;
+
+ case 32:
+ radixArray[0] = 8;
+ radixArray[1] = 4;
+ return 2;
+
+ case 64:
+ radixArray[0] = 8;
+ radixArray[1] = 8;
+ return 2;
+
+ case 128:
+ radixArray[0] = 8;
+ radixArray[1] = 4;
+ radixArray[2] = 4;
+ return 3;
+
+ case 256:
+ radixArray[0] = 4;
+ radixArray[1] = 4;
+ radixArray[2] = 4;
+ radixArray[3] = 4;
+ return 4;
+
+ case 512:
+ radixArray[0] = 8;
+ radixArray[1] = 8;
+ radixArray[2] = 8;
+ return 3;
+
+ case 1024:
+ radixArray[0] = 16;
+ radixArray[1] = 16;
+ radixArray[2] = 4;
+ return 3;
+ case 2048:
+ radixArray[0] = 8;
+ radixArray[1] = 8;
+ radixArray[2] = 8;
+ radixArray[3] = 4;
+ return 4;
+ default:
+ return 0;
+ }
+ }
+
+ void insertHeader(StringBuilder kernelString, String kernelName, CLFFTDataFormat dataFormat) {
+ if (dataFormat == CLFFTPlan.CLFFTDataFormat.SplitComplexFormat) {
+ kernelString.append("__kernel void ").append(kernelName).append("(__global float *in_real, __global float *in_imag, __global float *out_real, __global float *out_imag, int dir, int S)\n");
+ } else {
+ kernelString.append("__kernel void ").append(kernelName).append("(__global float2 *in, __global float2 *out, int dir, int S)\n");
+ }
+ }
+
+ void insertVariables(StringBuilder kStream, int maxRadix) {
+ kStream.append(" int i, j, r, indexIn, indexOut, index, tid, bNum, xNum, k, l;\n");
+ kStream.append(" int s, ii, jj, offset;\n");
+ kStream.append(" float2 w;\n");
+ kStream.append(" float ang, angf, ang1;\n");
+ kStream.append(" __local float *lMemStore, *lMemLoad;\n");
+ kStream.append(" float2 a[").append(maxRadix).append("];\n");
+ kStream.append(" int lId = get_local_id( 0 );\n");
+ kStream.append(" int groupId = get_group_id( 0 );\n");
+ }
+
+ void formattedLoad(StringBuilder kernelString, int aIndex, int gIndex, CLFFTDataFormat dataFormat) {
+ if (dataFormat == dataFormat.InterleavedComplexFormat) {
+ kernelString.append(" a[").append(aIndex).append("] = in[").append(gIndex).append("];\n");
+ } else {
+ kernelString.append(" a[").append(aIndex).append("].x = in_real[").append(gIndex).append("];\n");
+ kernelString.append(" a[").append(aIndex).append("].y = in_imag[").append(gIndex).append("];\n");
+ }
+ }
+
+ void formattedStore(StringBuilder kernelString, int aIndex, int gIndex, CLFFTDataFormat dataFormat) {
+ if (dataFormat == dataFormat.InterleavedComplexFormat) {
+ kernelString.append(" out[").append(gIndex).append("] = a[").append(aIndex).append("];\n");
+ } else {
+ kernelString.append(" out_real[").append(gIndex).append("] = a[").append(aIndex).append("].x;\n");
+ kernelString.append(" out_imag[").append(gIndex).append("] = a[").append(aIndex).append("].y;\n");
+ }
+ }
+
+ int insertGlobalLoadsAndTranspose(StringBuilder kernelString, int N, int numWorkItemsPerXForm, int numXFormsPerWG, int R0, int mem_coalesce_width, CLFFTDataFormat dataFormat) {
+ int log2NumWorkItemsPerXForm = (int) log2(numWorkItemsPerXForm);
+ int groupSize = numWorkItemsPerXForm * numXFormsPerWG;
+ int i, j;
+ int lMemSize = 0;
+
+ if (numXFormsPerWG > 1) {
+ kernelString.append(" s = S & ").append(numXFormsPerWG - 1).append(";\n");
+ }
+
+ if (numWorkItemsPerXForm >= mem_coalesce_width) {
+ if (numXFormsPerWG > 1) {
+ kernelString.append(" ii = lId & ").append(numWorkItemsPerXForm - 1).append(";\n");
+ kernelString.append(" jj = lId >> ").append(log2NumWorkItemsPerXForm).append(";\n");
+ kernelString.append(" if( !s || (groupId < get_num_groups(0)-1) || (jj < s) ) {\n");
+ kernelString.append(" offset = mad24( mad24(groupId, ").append(numXFormsPerWG).append(", jj), ").append(N).append(", ii );\n");
+ if (dataFormat == dataFormat.InterleavedComplexFormat) {
+ kernelString.append(" in += offset;\n");
+ kernelString.append(" out += offset;\n");
+ } else {
+ kernelString.append(" in_real += offset;\n");
+ kernelString.append(" in_imag += offset;\n");
+ kernelString.append(" out_real += offset;\n");
+ kernelString.append(" out_imag += offset;\n");
+ }
+ for (i = 0; i < R0; i++) {
+ formattedLoad(kernelString, i, i * numWorkItemsPerXForm, dataFormat);
+ }
+ kernelString.append(" }\n");
+ } else {
+ kernelString.append(" ii = lId;\n");
+ kernelString.append(" jj = 0;\n");
+ kernelString.append(" offset = mad24(groupId, ").append(N).append(", ii);\n");
+ if (dataFormat == dataFormat.InterleavedComplexFormat) {
+ kernelString.append(" in += offset;\n");
+ kernelString.append(" out += offset;\n");
+ } else {
+ kernelString.append(" in_real += offset;\n");
+ kernelString.append(" in_imag += offset;\n");
+ kernelString.append(" out_real += offset;\n");
+ kernelString.append(" out_imag += offset;\n");
+ }
+ for (i = 0; i < R0; i++) {
+ formattedLoad(kernelString, i, i * numWorkItemsPerXForm, dataFormat);
+ }
+ }
+ } else if (N >= mem_coalesce_width) {
+ int numInnerIter = N / mem_coalesce_width;
+ int numOuterIter = numXFormsPerWG / (groupSize / mem_coalesce_width);
+
+ kernelString.append(" ii = lId & ").append(mem_coalesce_width - 1).append(";\n");
+ kernelString.append(" jj = lId >> ").append((int) log2(mem_coalesce_width)).append(";\n");
+ kernelString.append(" lMemStore = sMem + mad24( jj, ").append(N + numWorkItemsPerXForm).append(", ii );\n");
+ kernelString.append(" offset = mad24( groupId, ").append(numXFormsPerWG).append(", jj);\n");
+ kernelString.append(" offset = mad24( offset, ").append(N).append(", ii );\n");
+ if (dataFormat == dataFormat.InterleavedComplexFormat) {
+ kernelString.append(" in += offset;\n");
+ kernelString.append(" out += offset;\n");
+ } else {
+ kernelString.append(" in_real += offset;\n");
+ kernelString.append(" in_imag += offset;\n");
+ kernelString.append(" out_real += offset;\n");
+ kernelString.append(" out_imag += offset;\n");
+ }
+
+ kernelString.append("if((groupId == get_num_groups(0)-1) && s) {\n");
+ for (i = 0; i < numOuterIter; i++) {
+ kernelString.append(" if( jj < s ) {\n");
+ for (j = 0; j < numInnerIter; j++) {
+ formattedLoad(kernelString, i * numInnerIter + j, j * mem_coalesce_width + i * (groupSize / mem_coalesce_width) * N, dataFormat);
+ }
+ kernelString.append(" }\n");
+ if (i != numOuterIter - 1) {
+ kernelString.append(" jj += ").append(groupSize / mem_coalesce_width).append(";\n");
+ }
+ }
+ kernelString.append("}\n ");
+ kernelString.append("else {\n");
+ for (i = 0; i < numOuterIter; i++) {
+ for (j = 0; j < numInnerIter; j++) {
+ formattedLoad(kernelString, i * numInnerIter + j, j * mem_coalesce_width + i * (groupSize / mem_coalesce_width) * N, dataFormat);
+ }
+ }
+ kernelString.append("}\n");
+
+ kernelString.append(" ii = lId & ").append(numWorkItemsPerXForm - 1).append(";\n");
+ kernelString.append(" jj = lId >> ").append(log2NumWorkItemsPerXForm).append(";\n");
+ kernelString.append(" lMemLoad = sMem + mad24( jj, ").append(N + numWorkItemsPerXForm).append(", ii);\n");
+
+ for (i = 0; i < numOuterIter; i++) {
+ for (j = 0; j < numInnerIter; j++) {
+ kernelString.append(" lMemStore[").append(j * mem_coalesce_width + i * (groupSize / mem_coalesce_width) * (N + numWorkItemsPerXForm)).append("] = a[").append(i * numInnerIter + j).append("].x;\n");
+ }
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < R0; i++) {
+ kernelString.append(" a[").append(i).append("].x = lMemLoad[").append(i * numWorkItemsPerXForm).append("];\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < numOuterIter; i++) {
+ for (j = 0; j < numInnerIter; j++) {
+ kernelString.append(" lMemStore[").append(j * mem_coalesce_width + i * (groupSize / mem_coalesce_width) * (N + numWorkItemsPerXForm)).append("] = a[").append(i * numInnerIter + j).append("].y;\n");
+ }
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < R0; i++) {
+ kernelString.append(" a[").append(i).append("].y = lMemLoad[").append(i * numWorkItemsPerXForm).append("];\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ lMemSize = (N + numWorkItemsPerXForm) * numXFormsPerWG;
+ } else {
+ kernelString.append(" offset = mad24( groupId, ").append(N * numXFormsPerWG).append(", lId );\n");
+ if (dataFormat == dataFormat.InterleavedComplexFormat) {
+ kernelString.append(" in += offset;\n");
+ kernelString.append(" out += offset;\n");
+ } else {
+ kernelString.append(" in_real += offset;\n");
+ kernelString.append(" in_imag += offset;\n");
+ kernelString.append(" out_real += offset;\n");
+ kernelString.append(" out_imag += offset;\n");
+ }
+
+ kernelString.append(" ii = lId & ").append(N - 1).append(";\n");
+ kernelString.append(" jj = lId >> ").append((int) log2(N)).append(";\n");
+ kernelString.append(" lMemStore = sMem + mad24( jj, ").append(N + numWorkItemsPerXForm).append(", ii );\n");
+
+ kernelString.append("if((groupId == get_num_groups(0)-1) && s) {\n");
+ for (i = 0; i < R0; i++) {
+ kernelString.append(" if(jj < s )\n");
+ formattedLoad(kernelString, i, i * groupSize, dataFormat);
+ if (i != R0 - 1) {
+ kernelString.append(" jj += ").append(groupSize / N).append(";\n");
+ }
+ }
+ kernelString.append("}\n");
+ kernelString.append("else {\n");
+ for (i = 0; i < R0; i++) {
+ formattedLoad(kernelString, i, i * groupSize, dataFormat);
+ }
+ kernelString.append("}\n");
+
+ if (numWorkItemsPerXForm > 1) {
+ kernelString.append(" ii = lId & ").append(numWorkItemsPerXForm - 1).append(";\n");
+ kernelString.append(" jj = lId >> ").append(log2NumWorkItemsPerXForm).append(";\n");
+ kernelString.append(" lMemLoad = sMem + mad24( jj, ").append(N + numWorkItemsPerXForm).append(", ii );\n");
+ } else {
+ kernelString.append(" ii = 0;\n");
+ kernelString.append(" jj = lId;\n");
+ kernelString.append(" lMemLoad = sMem + mul24( jj, ").append(N + numWorkItemsPerXForm).append(");\n");
+ }
+
+
+ for (i = 0; i < R0; i++) {
+ kernelString.append(" lMemStore[").append(i * (groupSize / N) * (N + numWorkItemsPerXForm)).append("] = a[").append(i).append("].x;\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < R0; i++) {
+ kernelString.append(" a[").append(i).append("].x = lMemLoad[").append(i * numWorkItemsPerXForm).append("];\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < R0; i++) {
+ kernelString.append(" lMemStore[").append(i * (groupSize / N) * (N + numWorkItemsPerXForm)).append("] = a[").append(i).append("].y;\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < R0; i++) {
+ kernelString.append(" a[").append(i).append("].y = lMemLoad[").append(i * numWorkItemsPerXForm).append("];\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ lMemSize = (N + numWorkItemsPerXForm) * numXFormsPerWG;
+ }
+
+ return lMemSize;
+ }
+
+ int insertGlobalStoresAndTranspose(StringBuilder kernelString, int N, int maxRadix, int Nr, int numWorkItemsPerXForm, int numXFormsPerWG, int mem_coalesce_width, CLFFTDataFormat dataFormat) {
+ int groupSize = numWorkItemsPerXForm * numXFormsPerWG;
+ int i, j, k, ind;
+ int lMemSize = 0;
+ int numIter = maxRadix / Nr;
+ String indent = "";
+
+ if (numWorkItemsPerXForm >= mem_coalesce_width) {
+ if (numXFormsPerWG > 1) {
+ kernelString.append(" if( !s || (groupId < get_num_groups(0)-1) || (jj < s) ) {\n");
+ indent = (" ");
+ }
+ for (i = 0; i < maxRadix; i++) {
+ j = i % numIter;
+ k = i / numIter;
+ ind = j * Nr + k;
+ formattedStore(kernelString, ind, i * numWorkItemsPerXForm, dataFormat);
+ }
+ if (numXFormsPerWG > 1) {
+ kernelString.append(" }\n");
+ }
+ } else if (N >= mem_coalesce_width) {
+ int numInnerIter = N / mem_coalesce_width;
+ int numOuterIter = numXFormsPerWG / (groupSize / mem_coalesce_width);
+
+ kernelString.append(" lMemLoad = sMem + mad24( jj, ").append(N + numWorkItemsPerXForm).append(", ii );\n");
+ kernelString.append(" ii = lId & ").append(mem_coalesce_width - 1).append(";\n");
+ kernelString.append(" jj = lId >> ").append((int) log2(mem_coalesce_width)).append(";\n");
+ kernelString.append(" lMemStore = sMem + mad24( jj,").append(N + numWorkItemsPerXForm).append(", ii );\n");
+
+ for (i = 0; i < maxRadix; i++) {
+ j = i % numIter;
+ k = i / numIter;
+ ind = j * Nr + k;
+ kernelString.append(" lMemLoad[").append(i * numWorkItemsPerXForm).append("] = a[").append(ind).append("].x;\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < numOuterIter; i++) {
+ for (j = 0; j < numInnerIter; j++) {
+ kernelString.append(" a[").append(i * numInnerIter + j).append("].x = lMemStore[").append(j * mem_coalesce_width + i * (groupSize / mem_coalesce_width) * (N + numWorkItemsPerXForm)).append("];\n");
+ }
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < maxRadix; i++) {
+ j = i % numIter;
+ k = i / numIter;
+ ind = j * Nr + k;
+ kernelString.append(" lMemLoad[").append(i * numWorkItemsPerXForm).append("] = a[").append(ind).append("].y;\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < numOuterIter; i++) {
+ for (j = 0; j < numInnerIter; j++) {
+ kernelString.append(" a[").append(i * numInnerIter + j).append("].y = lMemStore[").append(j * mem_coalesce_width + i * (groupSize / mem_coalesce_width) * (N + numWorkItemsPerXForm)).append("];\n");
+ }
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ kernelString.append("if((groupId == get_num_groups(0)-1) && s) {\n");
+ for (i = 0; i < numOuterIter; i++) {
+ kernelString.append(" if( jj < s ) {\n");
+ for (j = 0; j < numInnerIter; j++) {
+ formattedStore(kernelString, i * numInnerIter + j, j * mem_coalesce_width + i * (groupSize / mem_coalesce_width) * N, dataFormat);
+ }
+ kernelString.append(" }\n");
+ if (i != numOuterIter - 1) {
+ kernelString.append(" jj += ").append(groupSize / mem_coalesce_width).append(";\n");
+ }
+ }
+ kernelString.append("}\n");
+ kernelString.append("else {\n");
+ for (i = 0; i < numOuterIter; i++) {
+ for (j = 0; j < numInnerIter; j++) {
+ formattedStore(kernelString, i * numInnerIter + j, j * mem_coalesce_width + i * (groupSize / mem_coalesce_width) * N, dataFormat);
+ }
+ }
+ kernelString.append("}\n");
+
+ lMemSize = (N + numWorkItemsPerXForm) * numXFormsPerWG;
+ } else {
+ kernelString.append(" lMemLoad = sMem + mad24( jj,").append(N + numWorkItemsPerXForm).append(", ii );\n");
+
+ kernelString.append(" ii = lId & ").append(N - 1).append(";\n");
+ kernelString.append(" jj = lId >> ").append((int) log2(N)).append(";\n");
+ kernelString.append(" lMemStore = sMem + mad24( jj,").append(N + numWorkItemsPerXForm).append(", ii );\n");
+
+ for (i = 0; i < maxRadix; i++) {
+ j = i % numIter;
+ k = i / numIter;
+ ind = j * Nr + k;
+ kernelString.append(" lMemLoad[").append(i * numWorkItemsPerXForm).append("] = a[").append(ind).append("].x;\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < maxRadix; i++) {
+ kernelString.append(" a[").append(i).append("].x = lMemStore[").append(i * (groupSize / N) * (N + numWorkItemsPerXForm)).append("];\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < maxRadix; i++) {
+ j = i % numIter;
+ k = i / numIter;
+ ind = j * Nr + k;
+ kernelString.append(" lMemLoad[").append(i * numWorkItemsPerXForm).append("] = a[").append(ind).append("].y;\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ for (i = 0; i < maxRadix; i++) {
+ kernelString.append(" a[").append(i).append("].y = lMemStore[").append(i * (groupSize / N) * (N + numWorkItemsPerXForm)).append("];\n");
+ }
+ kernelString.append(" barrier( CLK_LOCAL_MEM_FENCE );\n");
+
+ kernelString.append("if((groupId == get_num_groups(0)-1) && s) {\n");
+ for (i = 0; i < maxRadix; i++) {
+ kernelString.append(" if(jj < s ) {\n");
+ formattedStore(kernelString, i, i * groupSize, dataFormat);
+ kernelString.append(" }\n");
+ if (i != maxRadix - 1) {
+ kernelString.append(" jj +=").append(groupSize / N).append(";\n");
+ }
+ }
+ kernelString.append("}\n");
+ kernelString.append("else {\n");
+ for (i = 0; i < maxRadix; i++) {
+ formattedStore(kernelString, i, i * groupSize, dataFormat);
+ }
+ kernelString.append("}\n");
+
+ lMemSize = (N + numWorkItemsPerXForm) * numXFormsPerWG;
+ }
+
+ return lMemSize;
+ }
+
+ void insertfftKernel(StringBuilder kernelString, int Nr, int numIter) {
+ int i;
+ for (i = 0; i < numIter; i++) {
+ kernelString.append(" fftKernel").append(Nr).append("(a+").append(i * Nr).append(", dir);\n");
+ }
+ }
+
+ void insertTwiddleKernel(StringBuilder kernelString, int Nr, int numIter, int Nprev, int len, int numWorkItemsPerXForm) {
+ int z, k;
+ int logNPrev = log2(Nprev);
+
+ for (z = 0; z < numIter; z++) {
+ if (z == 0) {
+ if (Nprev > 1) {
+ kernelString.append(" angf = (float) (ii >> ").append(logNPrev).append(");\n");
+ } else {
+ kernelString.append(" angf = (float) ii;\n");
+ }
+ } else {
+ if (Nprev > 1) {
+ kernelString.append(" angf = (float) ((").append(z * numWorkItemsPerXForm).append(" + ii) >>").append(logNPrev).append(");\n");
+ } else {
+ kernelString.append(" angf = (float) (").append(z * numWorkItemsPerXForm).append(" + ii);\n");
+ }
+ }
+
+ for (k = 1; k < Nr; k++) {
+ int ind = z * Nr + k;
+ //float fac = (float) (2.0 * M_PI * (double) k / (double) len);
+ kernelString.append(" ang = dir * ( 2.0f * M_PI * ").append(k).append(".0f / ").append(len).append(".0f )").append(" * angf;\n");
+ kernelString.append(" w = (float2)(native_cos(ang), native_sin(ang));\n");
+ kernelString.append(" a[").append(ind).append("] = complexMul(a[").append(ind).append("], w);\n");
+ }
+ }
+ }
+
+ fftPadding getPadding(int numWorkItemsPerXForm, int Nprev, int numWorkItemsReq, int numXFormsPerWG, int Nr, int numBanks) {
+ int offset, midPad;
+
+ if ((numWorkItemsPerXForm <= Nprev) || (Nprev >= numBanks)) {
+ offset = 0;
+ } else {
+ int numRowsReq = ((numWorkItemsPerXForm < numBanks) ? numWorkItemsPerXForm : numBanks) / Nprev;
+ int numColsReq = 1;
+ if (numRowsReq > Nr) {
+ numColsReq = numRowsReq / Nr;
+ }
+ numColsReq = Nprev * numColsReq;
+ offset = numColsReq;
+ }
+
+ if (numWorkItemsPerXForm >= numBanks || numXFormsPerWG == 1) {
+ midPad = 0;
+ } else {
+ int bankNum = ((numWorkItemsReq + offset) * Nr) & (numBanks - 1);
+ if (bankNum >= numWorkItemsPerXForm) {
+ midPad = 0;
+ } else {
+ midPad = numWorkItemsPerXForm - bankNum;
+ }
+ }
+
+ int lMemSize = (numWorkItemsReq + offset) * Nr * numXFormsPerWG + midPad * (numXFormsPerWG - 1);
+ return new fftPadding(lMemSize, offset, midPad);
+ }
+
+ void insertLocalStores(StringBuilder kernelString, int numIter, int Nr, int numWorkItemsPerXForm, int numWorkItemsReq, int offset, String comp) {
+ int z, k;
+
+ for (z = 0; z < numIter; z++) {
+ for (k = 0; k < Nr; k++) {
+ int index = k * (numWorkItemsReq + offset) + z * numWorkItemsPerXForm;
+ kernelString.append(" lMemStore[").append(index).append("] = a[").append(z * Nr + k).append("].").append(comp).append(";\n");
+ }
+ }
+ kernelString.append(" barrier(CLK_LOCAL_MEM_FENCE);\n");
+ }
+
+ void insertLocalLoads(StringBuilder kernelString, int n, int Nr, int Nrn, int Nprev, int Ncurr, int numWorkItemsPerXForm, int numWorkItemsReq, int offset, String comp) {
+ int numWorkItemsReqN = n / Nrn;
+ int interBlockHNum = Math.max(Nprev / numWorkItemsPerXForm, 1);
+ int interBlockHStride = numWorkItemsPerXForm;
+ int vertWidth = Math.max(numWorkItemsPerXForm / Nprev, 1);
+ vertWidth = Math.min(vertWidth, Nr);
+ int vertNum = Nr / vertWidth;
+ int vertStride = (n / Nr + offset) * vertWidth;
+ int iter = Math.max(numWorkItemsReqN / numWorkItemsPerXForm, 1);
+ int intraBlockHStride = (numWorkItemsPerXForm / (Nprev * Nr)) > 1 ? (numWorkItemsPerXForm / (Nprev * Nr)) : 1;
+ intraBlockHStride *= Nprev;
+
+ int stride = numWorkItemsReq / Nrn;
+ int i;
+ for (i = 0; i < iter; i++) {
+ int ii = i / (interBlockHNum * vertNum);
+ int zz = i % (interBlockHNum * vertNum);
+ int jj = zz % interBlockHNum;
+ int kk = zz / interBlockHNum;
+ int z;
+ for (z = 0; z < Nrn; z++) {
+ int st = kk * vertStride + jj * interBlockHStride + ii * intraBlockHStride + z * stride;
+ kernelString.append(" a[").append(i * Nrn + z).append("].").append(comp).append(" = lMemLoad[").append(st).append("];\n");
+ }
+ }
+ kernelString.append(" barrier(CLK_LOCAL_MEM_FENCE);\n");
+ }
+
+ void insertLocalLoadIndexArithmatic(StringBuilder kernelString, int Nprev, int Nr, int numWorkItemsReq, int numWorkItemsPerXForm, int numXFormsPerWG, int offset, int midPad) {
+ int Ncurr = Nprev * Nr;
+ int logNcurr = log2(Ncurr);
+ int logNprev = log2(Nprev);
+ int incr = (numWorkItemsReq + offset) * Nr + midPad;
+
+ if (Ncurr < numWorkItemsPerXForm) {
+ if (Nprev == 1) {
+ kernelString.append(" j = ii & ").append(Ncurr - 1).append(";\n");
+ } else {
+ kernelString.append(" j = (ii & ").append(Ncurr - 1).append(") >> ").append(logNprev).append(";\n");
+ }
+
+ if (Nprev == 1) {
+ kernelString.append(" i = ii >> ").append(logNcurr).append(";\n");
+ } else {
+ kernelString.append(" i = mad24(ii >> ").append(logNcurr).append(", ").append(Nprev).append(", ii & ").append(Nprev - 1).append(");\n");
+ }
+ } else {
+ if (Nprev == 1) {
+ kernelString.append(" j = ii;\n");
+ } else {
+ kernelString.append(" j = ii >> ").append(logNprev).append(";\n");
+ }
+ if (Nprev == 1) {
+ kernelString.append(" i = 0;\n");
+ } else {
+ kernelString.append(" i = ii & ").append(Nprev - 1).append(";\n");
+ }
+ }
+
+ if (numXFormsPerWG > 1) {
+ kernelString.append(" i = mad24(jj, ").append(incr).append(", i);\n");
+ }
+
+ kernelString.append(" lMemLoad = sMem + mad24(j, ").append(numWorkItemsReq + offset).append(", i);\n");
+ }
+
+ void insertLocalStoreIndexArithmatic(StringBuilder kernelString, int numWorkItemsReq, int numXFormsPerWG, int Nr, int offset, int midPad) {
+ if (numXFormsPerWG == 1) {
+ kernelString.append(" lMemStore = sMem + ii;\n");
+ } else {
+ kernelString.append(" lMemStore = sMem + mad24(jj, ").append((numWorkItemsReq + offset) * Nr + midPad).append(", ii);\n");
+ }
+ }
+
+ void createLocalMemfftKernelString() {
+ int[] radixArray = new int[10];
+ int numRadix;
+
+ int n = this.size.x;
+
+ assert (n <= this.max_work_item_per_workgroup * this.max_radix);
+
+ numRadix = getRadixArray(n, radixArray, 0);
+ assert (numRadix > 0);
+
+ if (n / radixArray[0] > this.max_work_item_per_workgroup) {
+ numRadix = getRadixArray(n, radixArray, this.max_radix);
+ }
+
+ assert (radixArray[0] <= this.max_radix);
+ assert (n / radixArray[0] <= this.max_work_item_per_workgroup);
+
+ int tmpLen = 1;
+ int i;
+ for (i = 0; i < numRadix; i++) {
+ assert ((radixArray[i] != 0) && !(((radixArray[i] - 1) != 0) & (radixArray[i] != 0)));
+ tmpLen *= radixArray[i];
+ }
+ assert (tmpLen == n);
+
+ //int offset, midPad;
+ StringBuilder localString = new StringBuilder();
+ String kernelName;
+
+ CLFFTDataFormat dataFormat = this.format;
+ StringBuilder kernelString = this.kernel_string;
+
+ int kCount = kernel_list.size();
+
+ kernelName = "fft" + (kCount);
+
+ CLFFTKernelInfo kInfo = new CLFFTKernelInfo();
+ kernel_list.add(kInfo);
+ //kInfo.kernel = null;
+ //kInfo.lmem_size = 0;
+ //kInfo.num_workgroups = 0;
+ //kInfo.num_workitems_per_workgroup = 0;
+ kInfo.dir = CLFFTKernelDir.X;
+ kInfo.in_place_possible = true;
+ //kInfo.next = null;
+ kInfo.kernel_name = kernelName;
+
+ int numWorkItemsPerXForm = n / radixArray[0];
+ int numWorkItemsPerWG = numWorkItemsPerXForm <= 64 ? 64 : numWorkItemsPerXForm;
+ assert (numWorkItemsPerWG <= this.max_work_item_per_workgroup);
+ int numXFormsPerWG = numWorkItemsPerWG / numWorkItemsPerXForm;
+ kInfo.num_workgroups = 1;
+ kInfo.num_xforms_per_workgroup = numXFormsPerWG;
+ kInfo.num_workitems_per_workgroup = numWorkItemsPerWG;
+
+ int[] N = radixArray;
+ int maxRadix = N[0];
+ int lMemSize = 0;
+
+ insertVariables(localString, maxRadix);
+
+ lMemSize = insertGlobalLoadsAndTranspose(localString, n, numWorkItemsPerXForm, numXFormsPerWG, maxRadix, this.min_mem_coalesce_width, dataFormat);
+ kInfo.lmem_size = (lMemSize > kInfo.lmem_size) ? lMemSize : kInfo.lmem_size;
+
+ String xcomp = "x";
+ String ycomp = "y";
+
+ int Nprev = 1;
+ int len = n;
+ int r;
+ for (r = 0; r < numRadix; r++) {
+ int numIter = N[0] / N[r];
+ int numWorkItemsReq = n / N[r];
+ int Ncurr = Nprev * N[r];
+ insertfftKernel(localString, N[r], numIter);
+
+ if (r < (numRadix - 1)) {
+ fftPadding pad;
+
+ insertTwiddleKernel(localString, N[r], numIter, Nprev, len, numWorkItemsPerXForm);
+ pad = getPadding(numWorkItemsPerXForm, Nprev, numWorkItemsReq, numXFormsPerWG, N[r], this.num_local_mem_banks);
+ kInfo.lmem_size = (pad.lMemSize > kInfo.lmem_size) ? pad.lMemSize : kInfo.lmem_size;
+ insertLocalStoreIndexArithmatic(localString, numWorkItemsReq, numXFormsPerWG, N[r], pad.offset, pad.midPad);
+ insertLocalLoadIndexArithmatic(localString, Nprev, N[r], numWorkItemsReq, numWorkItemsPerXForm, numXFormsPerWG, pad.offset, pad.midPad);
+ insertLocalStores(localString, numIter, N[r], numWorkItemsPerXForm, numWorkItemsReq, pad.offset, xcomp);
+ insertLocalLoads(localString, n, N[r], N[r + 1], Nprev, Ncurr, numWorkItemsPerXForm, numWorkItemsReq, pad.offset, xcomp);
+ insertLocalStores(localString, numIter, N[r], numWorkItemsPerXForm, numWorkItemsReq, pad.offset, ycomp);
+ insertLocalLoads(localString, n, N[r], N[r + 1], Nprev, Ncurr, numWorkItemsPerXForm, numWorkItemsReq, pad.offset, ycomp);
+ Nprev = Ncurr;
+ len = len / N[r];
+ }
+ }
+
+ lMemSize = insertGlobalStoresAndTranspose(localString, n, maxRadix, N[numRadix - 1], numWorkItemsPerXForm, numXFormsPerWG, this.min_mem_coalesce_width, dataFormat);
+ kInfo.lmem_size = (lMemSize > kInfo.lmem_size) ? lMemSize : kInfo.lmem_size;
+
+ insertHeader(kernelString, kernelName, dataFormat);
+ kernelString.append("{\n");
+ if (kInfo.lmem_size > 0) {
+ kernelString.append(" __local float sMem[").append(kInfo.lmem_size).append("];\n");
+ }
+ kernelString.append(localString);
+ kernelString.append("}\n");
+ }
+
+// For size larger than what can be computed using local memory fft, global transposes
+// multiple kernel launces is needed. For these sizes, size can be decomposed using
+// much larger base radices i.e. say size = 262144 = 128 x 64 x 32. Thus three kernel
+// launches will be needed, first computing 64 x 32, length 128 ffts, second computing
+// 128 x 32 length 64 ffts, and finally a kernel computing 128 x 64 length 32 ffts.
+// Each of these base radices can futher be divided into factors so that each of these
+// base ffts can be computed within one kernel launch using in-register ffts and local
+// memory transposes i.e for the first kernel above which computes 64 x 32 ffts on length
+// 128, 128 can be decomposed into 128 = 16 x 8 i.e. 8 work items can compute 8 length
+// 16 ffts followed by transpose using local memory followed by each of these eight
+// work items computing 2 length 8 ffts thus computing 16 length 8 ffts in total. This
+// means only 8 work items are needed for computing one length 128 fft. If we choose
+// work group size of say 64, we can compute 64/8 = 8 length 128 ffts within one
+// work group. Since we need to compute 64 x 32 length 128 ffts in first kernel, this
+// means we need to launch 64 x 32 / 8 = 256 work groups with 64 work items in each
+// work group where each work group is computing 8 length 128 ffts where each length
+// 128 fft is computed by 8 work items. Same logic can be applied to other two kernels
+// in this example. Users can play with difference base radices and difference
+// decompositions of base radices to generates different kernels and see which gives
+// best performance. Following function is just fixed to use 128 as base radix
+ int getGlobalRadixInfo(int n, int[] radix, int[] R1, int[] R2) {
+ int baseRadix = Math.min(n, 128);
+
+ int numR = 0;
+ int N = n;
+ while (N > baseRadix) {
+ N /= baseRadix;
+ numR++;
+ }
+
+ for (int i = 0; i < numR; i++) {
+ radix[i] = baseRadix;
+ }
+
+ radix[numR] = N;
+ numR++;
+
+ for (int i = 0; i < numR; i++) {
+ int B = radix[i];
+ if (B <= 8) {
+ R1[i] = B;
+ R2[i] = 1;
+ continue;
+ }
+
+ int r1 = 2;
+ int r2 = B / r1;
+ while (r2 > r1) {
+ r1 *= 2;
+ r2 = B / r1;
+ }
+ R1[i] = r1;
+ R2[i] = r2;
+ }
+ return numR;
+ }
+
+ void createGlobalFFTKernelString(int n, int BS, CLFFTKernelDir dir, int vertBS) {
+ int i, j, k, t;
+ int[] radixArr = new int[10];
+ int[] R1Arr = new int[10];
+ int[] R2Arr = new int[10];
+ int radix, R1, R2;
+ int numRadices;
+
+ int maxThreadsPerBlock = this.max_work_item_per_workgroup;
+ int maxArrayLen = this.max_radix;
+ int batchSize = this.min_mem_coalesce_width;
+ CLFFTDataFormat dataFormat = this.format;
+ boolean vertical = (dir == dir.X) ? false : true;
+
+ numRadices = getGlobalRadixInfo(n, radixArr, R1Arr, R2Arr);
+
+ int numPasses = numRadices;
+
+ StringBuilder localString = new StringBuilder();
+ String kernelName;
+ StringBuilder kernelString = this.kernel_string;
+
+ int kCount = kernel_list.size();
+ //cl_fft_kernel_info **kInfo = &this.kernel_list;
+ //int kCount = 0;
+
+ //while(*kInfo)
+ //{
+ // kInfo = &kInfo.next;
+ // kCount++;
+ //}
+
+ int N = n;
+ int m = (int) log2(n);
+ int Rinit = vertical ? BS : 1;
+ batchSize = vertical ? Math.min(BS, batchSize) : batchSize;
+ int passNum;
+
+ for (passNum = 0; passNum < numPasses; passNum++) {
+
+ localString.setLength(0);
+ //kernelName.clear();
+
+ radix = radixArr[passNum];
+ R1 = R1Arr[passNum];
+ R2 = R2Arr[passNum];
+
+ int strideI = Rinit;
+ for (i = 0; i < numPasses; i++) {
+ if (i != passNum) {
+ strideI *= radixArr[i];
+ }
+ }
+
+ int strideO = Rinit;
+ for (i = 0; i < passNum; i++) {
+ strideO *= radixArr[i];
+ }
+
+ int threadsPerXForm = R2;
+ batchSize = R2 == 1 ? this.max_work_item_per_workgroup : batchSize;
+ batchSize = Math.min(batchSize, strideI);
+ int threadsPerBlock = batchSize * threadsPerXForm;
+ threadsPerBlock = Math.min(threadsPerBlock, maxThreadsPerBlock);
+ batchSize = threadsPerBlock / threadsPerXForm;
+ assert (R2 <= R1);
+ assert (R1 * R2 == radix);
+ assert (R1 <= maxArrayLen);
+ assert (threadsPerBlock <= maxThreadsPerBlock);
+
+ int numIter = R1 / R2;
+ int gInInc = threadsPerBlock / batchSize;
+
+
+ int lgStrideO = log2(strideO);
+ int numBlocksPerXForm = strideI / batchSize;
+ int numBlocks = numBlocksPerXForm;
+ if (!vertical) {
+ numBlocks *= BS;
+ } else {
+ numBlocks *= vertBS;
+ }
+
+ kernelName = "fft" + (kCount);
+ CLFFTKernelInfo kInfo = new CLFFTKernelInfo();
+ if (R2 == 1) {
+ kInfo.lmem_size = 0;
+ } else {
+ if (strideO == 1) {
+ kInfo.lmem_size = (radix + 1) * batchSize;
+ } else {
+ kInfo.lmem_size = threadsPerBlock * R1;
+ }
+ }
+ kInfo.num_workgroups = numBlocks;
+ kInfo.num_xforms_per_workgroup = 1;
+ kInfo.num_workitems_per_workgroup = threadsPerBlock;
+ kInfo.dir = dir;
+ kInfo.in_place_possible = ((passNum == (numPasses - 1)) && ((numPasses & 1) != 0));
+ //kInfo.next = NULL;
+ kInfo.kernel_name = kernelName;
+
+ insertVariables(localString, R1);
+
+ if (vertical) {
+ localString.append("xNum = groupId >> ").append((int) log2(numBlocksPerXForm)).append(";\n");
+ localString.append("groupId = groupId & ").append(numBlocksPerXForm - 1).append(";\n");
+ localString.append("indexIn = mad24(groupId, ").append(batchSize).append(", xNum << ").append((int) log2(n * BS)).append(");\n");
+ localString.append("tid = mul24(groupId, ").append(batchSize).append(");\n");
+ localString.append("i = tid >> ").append(lgStrideO).append(";\n");
+ localString.append("j = tid & ").append(strideO - 1).append(";\n");
+ int stride = radix * Rinit;
+ for (i = 0; i < passNum; i++) {
+ stride *= radixArr[i];
+ }
+ localString.append("indexOut = mad24(i, ").append(stride).append(", j + ").append("(xNum << ").append((int) log2(n * BS)).append("));\n");
+ localString.append("bNum = groupId;\n");
+ } else {
+ int lgNumBlocksPerXForm = log2(numBlocksPerXForm);
+ localString.append("bNum = groupId & ").append(numBlocksPerXForm - 1).append(";\n");
+ localString.append("xNum = groupId >> ").append(lgNumBlocksPerXForm).append(";\n");
+ localString.append("indexIn = mul24(bNum, ").append(batchSize).append(");\n");
+ localString.append("tid = indexIn;\n");
+ localString.append("i = tid >> ").append(lgStrideO).append(";\n");
+ localString.append("j = tid & ").append(strideO - 1).append(";\n");
+ int stride = radix * Rinit;
+ for (i = 0; i < passNum; i++) {
+ stride *= radixArr[i];
+ }
+ localString.append("indexOut = mad24(i, ").append(stride).append(", j);\n");
+ localString.append("indexIn += (xNum << ").append(m).append(");\n");
+ localString.append("indexOut += (xNum << ").append(m).append(");\n");
+ }
+
+ // Load Data
+ int lgBatchSize = log2(batchSize);
+ localString.append("tid = lId;\n");
+ localString.append("i = tid & ").append(batchSize - 1).append(";\n");
+ localString.append("j = tid >> ").append(lgBatchSize).append(";\n");
+ localString.append("indexIn += mad24(j, ").append(strideI).append(", i);\n");
+
+ if (dataFormat == dataFormat.SplitComplexFormat) {
+ localString.append("in_real += indexIn;\n");
+ localString.append("in_imag += indexIn;\n");
+ for (j = 0; j < R1; j++) {
+ localString.append("a[").append(j).append("].x = in_real[").append(j * gInInc * strideI).append("];\n");
+ }
+ for (j = 0; j < R1; j++) {
+ localString.append("a[").append(j).append("].y = in_imag[").append(j * gInInc * strideI).append("];\n");
+ }
+ } else {
+ localString.append("in += indexIn;\n");
+ for (j = 0; j < R1; j++) {
+ localString.append("a[").append(j).append("] = in[").append(j * gInInc * strideI).append("];\n");
+ }
+ }
+
+ localString.append("fftKernel").append(R1).append("(a, dir);\n");
+
+ if (R2 > 1) {
+ // twiddle
+ for (k = 1; k < R1; k++) {
+ localString.append("ang = dir*(2.0f*M_PI*").append(k).append("/").append(radix).append(")*j;\n");
+ localString.append("w = (float2)(native_cos(ang), native_sin(ang));\n");
+ localString.append("a[").append(k).append("] = complexMul(a[").append(k).append("], w);\n");
+ }
+
+ // shuffle
+ numIter = R1 / R2;
+ localString.append("indexIn = mad24(j, ").append(threadsPerBlock * numIter).append(", i);\n");
+ localString.append("lMemStore = sMem + tid;\n");
+ localString.append("lMemLoad = sMem + indexIn;\n");
+ for (k = 0; k < R1; k++) {
+ localString.append("lMemStore[").append(k * threadsPerBlock).append("] = a[").append(k).append("].x;\n");
+ }
+ localString.append("barrier(CLK_LOCAL_MEM_FENCE);\n");
+ for (k = 0; k < numIter; k++) {
+ for (t = 0; t < R2; t++) {
+ localString.append("a[").append(k * R2 + t).append("].x = lMemLoad[").append(t * batchSize + k * threadsPerBlock).append("];\n");
+ }
+ }
+ localString.append("barrier(CLK_LOCAL_MEM_FENCE);\n");
+ for (k = 0; k < R1; k++) {
+ localString.append("lMemStore[").append(k * threadsPerBlock).append("] = a[").append(k).append("].y;\n");
+ }
+ localString.append("barrier(CLK_LOCAL_MEM_FENCE);\n");
+ for (k = 0; k < numIter; k++) {
+ for (t = 0; t < R2; t++) {
+ localString.append("a[").append(k * R2 + t).append("].y = lMemLoad[").append(t * batchSize + k * threadsPerBlock).append("];\n");
+ }
+ }
+ localString.append("barrier(CLK_LOCAL_MEM_FENCE);\n");
+
+ for (j = 0; j < numIter; j++) {
+ localString.append("fftKernel").append(R2).append("(a + ").append(j * R2).append(", dir);\n");
+ }
+ }
+
+ // twiddle
+ if (passNum < (numPasses - 1)) {
+ localString.append("l = ((bNum << ").append(lgBatchSize).append(") + i) >> ").append(lgStrideO).append(";\n");
+ localString.append("k = j << ").append((int) log2(R1 / R2)).append(";\n");
+ localString.append("ang1 = dir*(2.0f*M_PI/").append(N).append(")*l;\n");
+ for (t = 0; t < R1; t++) {
+ localString.append("ang = ang1*(k + ").append((t % R2) * R1 + (t / R2)).append(");\n");
+ localString.append("w = (float2)(native_cos(ang), native_sin(ang));\n");
+ localString.append("a[").append(t).append("] = complexMul(a[").append(t).append("], w);\n");
+ }
+ }
+
+ // Store Data
+ if (strideO == 1) {
+
+ localString.append("lMemStore = sMem + mad24(i, ").append(radix + 1).append(", j << ").append((int) log2(R1 / R2)).append(");\n");
+ localString.append("lMemLoad = sMem + mad24(tid >> ").append((int) log2(radix)).append(", ").append(radix + 1).append(", tid & ").append(radix - 1).append(");\n");
+
+ for (i = 0; i < R1 / R2; i++) {
+ for (j = 0; j < R2; j++) {
+ localString.append("lMemStore[ ").append(i + j * R1).append("] = a[").append(i * R2 + j).append("].x;\n");
+ }
+ }
+ localString.append("barrier(CLK_LOCAL_MEM_FENCE);\n");
+ if (threadsPerBlock >= radix) {
+ for (i = 0; i < R1; i++) {
+ localString.append("a[").append(i).append("].x = lMemLoad[").append(i * (radix + 1) * (threadsPerBlock / radix)).append("];\n");
+ }
+ } else {
+ int innerIter = radix / threadsPerBlock;
+ int outerIter = R1 / innerIter;
+ for (i = 0; i < outerIter; i++) {
+ for (j = 0; j < innerIter; j++) {
+ localString.append("a[").append(i * innerIter + j).append("].x = lMemLoad[").append(j * threadsPerBlock + i * (radix + 1)).append("];\n");
+ }
+ }
+ }
+ localString.append("barrier(CLK_LOCAL_MEM_FENCE);\n");
+
+ for (i = 0; i < R1 / R2; i++) {
+ for (j = 0; j < R2; j++) {
+ localString.append("lMemStore[ ").append(i + j * R1).append("] = a[").append(i * R2 + j).append("].y;\n");
+ }
+ }
+ localString.append("barrier(CLK_LOCAL_MEM_FENCE);\n");
+ if (threadsPerBlock >= radix) {
+ for (i = 0; i < R1; i++) {
+ localString.append("a[").append(i).append("].y = lMemLoad[").append(i * (radix + 1) * (threadsPerBlock / radix)).append("];\n");
+ }
+ } else {
+ int innerIter = radix / threadsPerBlock;
+ int outerIter = R1 / innerIter;
+ for (i = 0; i < outerIter; i++) {
+ for (j = 0; j < innerIter; j++) {
+ localString.append("a[").append(i * innerIter + j).append("].y = lMemLoad[").append(j * threadsPerBlock + i * (radix + 1)).append("];\n");
+ }
+ }
+ }
+ localString.append("barrier(CLK_LOCAL_MEM_FENCE);\n");
+
+ localString.append("indexOut += tid;\n");
+ if (dataFormat == dataFormat.SplitComplexFormat) {
+ localString.append("out_real += indexOut;\n");
+ localString.append("out_imag += indexOut;\n");
+ for (k = 0; k < R1; k++) {
+ localString.append("out_real[").append(k * threadsPerBlock).append("] = a[").append(k).append("].x;\n");
+ }
+ for (k = 0; k < R1; k++) {
+ localString.append("out_imag[").append(k * threadsPerBlock).append("] = a[").append(k).append("].y;\n");
+ }
+ } else {
+ localString.append("out += indexOut;\n");
+ for (k = 0; k < R1; k++) {
+ localString.append("out[").append(k * threadsPerBlock).append("] = a[").append(k).append("];\n");
+ }
+ }
+
+ } else {
+ localString.append("indexOut += mad24(j, ").append(numIter * strideO).append(", i);\n");
+ if (dataFormat == dataFormat.SplitComplexFormat) {
+ localString.append("out_real += indexOut;\n");
+ localString.append("out_imag += indexOut;\n");
+ for (k = 0; k < R1; k++) {
+ localString.append("out_real[").append(((k % R2) * R1 + (k / R2)) * strideO).append("] = a[").append(k).append("].x;\n");
+ }
+ for (k = 0; k < R1; k++) {
+ localString.append("out_imag[").append(((k % R2) * R1 + (k / R2)) * strideO).append("] = a[").append(k).append("].y;\n");
+ }
+ } else {
+ localString.append("out += indexOut;\n");
+ for (k = 0; k < R1; k++) {
+ localString.append("out[").append(((k % R2) * R1 + (k / R2)) * strideO).append("] = a[").append(k).append("];\n");
+ }
+ }
+ }
+
+ insertHeader(kernelString, kernelName, dataFormat);
+ kernelString.append("{\n");
+ if (kInfo.lmem_size > 0) {
+ kernelString.append(" __local float sMem[").append(kInfo.lmem_size).append("];\n");
+ }
+ kernelString.append(localString);
+ kernelString.append("}\n");
+
+ N /= radix;
+ kernel_list.add(kInfo);
+ kCount++;
+ }
+ }
+
+ void FFT1D(CLFFTKernelDir dir) {
+ int[] radixArray = new int[10];
+
+ switch (dir) {
+ case X:
+ if (this.size.x > this.max_localmem_fft_size) {
+ createGlobalFFTKernelString(this.size.x, 1, dir, 1);
+ } else if (this.size.x > 1) {
+ getRadixArray(this.size.x, radixArray, 0);
+ if (this.size.x / radixArray[0] <= this.max_work_item_per_workgroup) {
+ createLocalMemfftKernelString();
+ } else {
+ getRadixArray(this.size.x, radixArray, this.max_radix);
+ if (this.size.x / radixArray[0] <= this.max_work_item_per_workgroup) {
+ createLocalMemfftKernelString();
+ } else {
+ createGlobalFFTKernelString(this.size.x, 1, dir, 1);
+ }
+ }
+ }
+ break;
+
+ case Y:
+ if (this.size.y > 1) {
+ createGlobalFFTKernelString(this.size.y, this.size.x, dir, 1);
+ }
+ break;
+
+ case Z:
+ if (this.size.z > 1) {
+ createGlobalFFTKernelString(this.size.z, this.size.x * this.size.y, dir, 1);
+ }
+ default:
+ return;
+ }
+ }
+
+ /*
+ *
+ * Pre-defined kernel parts
+ *
+ */
+ static String baseKernels =
+ "#ifndef M_PI\n"
+ + "#define M_PI 0x1.921fb54442d18p+1\n"
+ + "#endif\n"
+ + "#define complexMul(a,b) ((float2)(mad(-(a).y, (b).y, (a).x * (b).x), mad((a).y, (b).x, (a).x * (b).y)))\n"
+ + "#define conj(a) ((float2)((a).x, -(a).y))\n"
+ + "#define conjTransp(a) ((float2)(-(a).y, (a).x))\n"
+ + "\n"
+ + "#define fftKernel2(a,dir) \\\n"
+ + "{ \\\n"
+ + " float2 c = (a)[0]; \\\n"
+ + " (a)[0] = c + (a)[1]; \\\n"
+ + " (a)[1] = c - (a)[1]; \\\n"
+ + "}\n"
+ + "\n"
+ + "#define fftKernel2S(d1,d2,dir) \\\n"
+ + "{ \\\n"
+ + " float2 c = (d1); \\\n"
+ + " (d1) = c + (d2); \\\n"
+ + " (d2) = c - (d2); \\\n"
+ + "}\n"
+ + "\n"
+ + "#define fftKernel4(a,dir) \\\n"
+ + "{ \\\n"
+ + " fftKernel2S((a)[0], (a)[2], dir); \\\n"
+ + " fftKernel2S((a)[1], (a)[3], dir); \\\n"
+ + " fftKernel2S((a)[0], (a)[1], dir); \\\n"
+ + " (a)[3] = (float2)(dir)*(conjTransp((a)[3])); \\\n"
+ + " fftKernel2S((a)[2], (a)[3], dir); \\\n"
+ + " float2 c = (a)[1]; \\\n"
+ + " (a)[1] = (a)[2]; \\\n"
+ + " (a)[2] = c; \\\n"
+ + "}\n"
+ + "\n"
+ + "#define fftKernel4s(a0,a1,a2,a3,dir) \\\n"
+ + "{ \\\n"
+ + " fftKernel2S((a0), (a2), dir); \\\n"
+ + " fftKernel2S((a1), (a3), dir); \\\n"
+ + " fftKernel2S((a0), (a1), dir); \\\n"
+ + " (a3) = (float2)(dir)*(conjTransp((a3))); \\\n"
+ + " fftKernel2S((a2), (a3), dir); \\\n"
+ + " float2 c = (a1); \\\n"
+ + " (a1) = (a2); \\\n"
+ + " (a2) = c; \\\n"
+ + "}\n"
+ + "\n"
+ + "#define bitreverse8(a) \\\n"
+ + "{ \\\n"
+ + " float2 c; \\\n"
+ + " c = (a)[1]; \\\n"
+ + " (a)[1] = (a)[4]; \\\n"
+ + " (a)[4] = c; \\\n"
+ + " c = (a)[3]; \\\n"
+ + " (a)[3] = (a)[6]; \\\n"
+ + " (a)[6] = c; \\\n"
+ + "}\n"
+ + "\n"
+ + "#define fftKernel8(a,dir) \\\n"
+ + "{ \\\n"
+ + " const float2 w1 = (float2)(0x1.6a09e6p-1f, dir*0x1.6a09e6p-1f); \\\n"
+ + " const float2 w3 = (float2)(-0x1.6a09e6p-1f, dir*0x1.6a09e6p-1f); \\\n"
+ + " float2 c; \\\n"
+ + " fftKernel2S((a)[0], (a)[4], dir); \\\n"
+ + " fftKernel2S((a)[1], (a)[5], dir); \\\n"
+ + " fftKernel2S((a)[2], (a)[6], dir); \\\n"
+ + " fftKernel2S((a)[3], (a)[7], dir); \\\n"
+ + " (a)[5] = complexMul(w1, (a)[5]); \\\n"
+ + " (a)[6] = (float2)(dir)*(conjTransp((a)[6])); \\\n"
+ + " (a)[7] = complexMul(w3, (a)[7]); \\\n"
+ + " fftKernel2S((a)[0], (a)[2], dir); \\\n"
+ + " fftKernel2S((a)[1], (a)[3], dir); \\\n"
+ + " fftKernel2S((a)[4], (a)[6], dir); \\\n"
+ + " fftKernel2S((a)[5], (a)[7], dir); \\\n"
+ + " (a)[3] = (float2)(dir)*(conjTransp((a)[3])); \\\n"
+ + " (a)[7] = (float2)(dir)*(conjTransp((a)[7])); \\\n"
+ + " fftKernel2S((a)[0], (a)[1], dir); \\\n"
+ + " fftKernel2S((a)[2], (a)[3], dir); \\\n"
+ + " fftKernel2S((a)[4], (a)[5], dir); \\\n"
+ + " fftKernel2S((a)[6], (a)[7], dir); \\\n"
+ + " bitreverse8((a)); \\\n"
+ + "}\n"
+ + "\n"
+ + "#define bitreverse4x4(a) \\\n"
+ + "{ \\\n"
+ + " float2 c; \\\n"
+ + " c = (a)[1]; (a)[1] = (a)[4]; (a)[4] = c; \\\n"
+ + " c = (a)[2]; (a)[2] = (a)[8]; (a)[8] = c; \\\n"
+ + " c = (a)[3]; (a)[3] = (a)[12]; (a)[12] = c; \\\n"
+ + " c = (a)[6]; (a)[6] = (a)[9]; (a)[9] = c; \\\n"
+ + " c = (a)[7]; (a)[7] = (a)[13]; (a)[13] = c; \\\n"
+ + " c = (a)[11]; (a)[11] = (a)[14]; (a)[14] = c; \\\n"
+ + "}\n"
+ + "\n"
+ + "#define fftKernel16(a,dir) \\\n"
+ + "{ \\\n"
+ + " const float w0 = 0x1.d906bcp-1f; \\\n"
+ + " const float w1 = 0x1.87de2ap-2f; \\\n"
+ + " const float w2 = 0x1.6a09e6p-1f; \\\n"
+ + " fftKernel4s((a)[0], (a)[4], (a)[8], (a)[12], dir); \\\n"
+ + " fftKernel4s((a)[1], (a)[5], (a)[9], (a)[13], dir); \\\n"
+ + " fftKernel4s((a)[2], (a)[6], (a)[10], (a)[14], dir); \\\n"
+ + " fftKernel4s((a)[3], (a)[7], (a)[11], (a)[15], dir); \\\n"
+ + " (a)[5] = complexMul((a)[5], (float2)(w0, dir*w1)); \\\n"
+ + " (a)[6] = complexMul((a)[6], (float2)(w2, dir*w2)); \\\n"
+ + " (a)[7] = complexMul((a)[7], (float2)(w1, dir*w0)); \\\n"
+ + " (a)[9] = complexMul((a)[9], (float2)(w2, dir*w2)); \\\n"
+ + " (a)[10] = (float2)(dir)*(conjTransp((a)[10])); \\\n"
+ + " (a)[11] = complexMul((a)[11], (float2)(-w2, dir*w2)); \\\n"
+ + " (a)[13] = complexMul((a)[13], (float2)(w1, dir*w0)); \\\n"
+ + " (a)[14] = complexMul((a)[14], (float2)(-w2, dir*w2)); \\\n"
+ + " (a)[15] = complexMul((a)[15], (float2)(-w0, dir*-w1)); \\\n"
+ + " fftKernel4((a), dir); \\\n"
+ + " fftKernel4((a) + 4, dir); \\\n"
+ + " fftKernel4((a) + 8, dir); \\\n"
+ + " fftKernel4((a) + 12, dir); \\\n"
+ + " bitreverse4x4((a)); \\\n"
+ + "}\n"
+ + "\n"
+ + "#define bitreverse32(a) \\\n"
+ + "{ \\\n"
+ + " float2 c1, c2; \\\n"
+ + " c1 = (a)[2]; (a)[2] = (a)[1]; c2 = (a)[4]; (a)[4] = c1; c1 = (a)[8]; (a)[8] = c2; c2 = (a)[16]; (a)[16] = c1; (a)[1] = c2; \\\n"
+ + " c1 = (a)[6]; (a)[6] = (a)[3]; c2 = (a)[12]; (a)[12] = c1; c1 = (a)[24]; (a)[24] = c2; c2 = (a)[17]; (a)[17] = c1; (a)[3] = c2; \\\n"
+ + " c1 = (a)[10]; (a)[10] = (a)[5]; c2 = (a)[20]; (a)[20] = c1; c1 = (a)[9]; (a)[9] = c2; c2 = (a)[18]; (a)[18] = c1; (a)[5] = c2; \\\n"
+ + " c1 = (a)[14]; (a)[14] = (a)[7]; c2 = (a)[28]; (a)[28] = c1; c1 = (a)[25]; (a)[25] = c2; c2 = (a)[19]; (a)[19] = c1; (a)[7] = c2; \\\n"
+ + " c1 = (a)[22]; (a)[22] = (a)[11]; c2 = (a)[13]; (a)[13] = c1; c1 = (a)[26]; (a)[26] = c2; c2 = (a)[21]; (a)[21] = c1; (a)[11] = c2; \\\n"
+ + " c1 = (a)[30]; (a)[30] = (a)[15]; c2 = (a)[29]; (a)[29] = c1; c1 = (a)[27]; (a)[27] = c2; c2 = (a)[23]; (a)[23] = c1; (a)[15] = c2; \\\n"
+ + "}\n"
+ + "\n"
+ + "#define fftKernel32(a,dir) \\\n"
+ + "{ \\\n"
+ + " fftKernel2S((a)[0], (a)[16], dir); \\\n"
+ + " fftKernel2S((a)[1], (a)[17], dir); \\\n"
+ + " fftKernel2S((a)[2], (a)[18], dir); \\\n"
+ + " fftKernel2S((a)[3], (a)[19], dir); \\\n"
+ + " fftKernel2S((a)[4], (a)[20], dir); \\\n"
+ + " fftKernel2S((a)[5], (a)[21], dir); \\\n"
+ + " fftKernel2S((a)[6], (a)[22], dir); \\\n"
+ + " fftKernel2S((a)[7], (a)[23], dir); \\\n"
+ + " fftKernel2S((a)[8], (a)[24], dir); \\\n"
+ + " fftKernel2S((a)[9], (a)[25], dir); \\\n"
+ + " fftKernel2S((a)[10], (a)[26], dir); \\\n"
+ + " fftKernel2S((a)[11], (a)[27], dir); \\\n"
+ + " fftKernel2S((a)[12], (a)[28], dir); \\\n"
+ + " fftKernel2S((a)[13], (a)[29], dir); \\\n"
+ + " fftKernel2S((a)[14], (a)[30], dir); \\\n"
+ + " fftKernel2S((a)[15], (a)[31], dir); \\\n"
+ + " (a)[17] = complexMul((a)[17], (float2)(0x1.f6297cp-1f, dir*0x1.8f8b84p-3f)); \\\n"
+ + " (a)[18] = complexMul((a)[18], (float2)(0x1.d906bcp-1f, dir*0x1.87de2ap-2f)); \\\n"
+ + " (a)[19] = complexMul((a)[19], (float2)(0x1.a9b662p-1f, dir*0x1.1c73b4p-1f)); \\\n"
+ + " (a)[20] = complexMul((a)[20], (float2)(0x1.6a09e6p-1f, dir*0x1.6a09e6p-1f)); \\\n"
+ + " (a)[21] = complexMul((a)[21], (float2)(0x1.1c73b4p-1f, dir*0x1.a9b662p-1f)); \\\n"
+ + " (a)[22] = complexMul((a)[22], (float2)(0x1.87de2ap-2f, dir*0x1.d906bcp-1f)); \\\n"
+ + " (a)[23] = complexMul((a)[23], (float2)(0x1.8f8b84p-3f, dir*0x1.f6297cp-1f)); \\\n"
+ + " (a)[24] = complexMul((a)[24], (float2)(0x0p+0f, dir*0x1p+0f)); \\\n"
+ + " (a)[25] = complexMul((a)[25], (float2)(-0x1.8f8b84p-3f, dir*0x1.f6297cp-1f)); \\\n"
+ + " (a)[26] = complexMul((a)[26], (float2)(-0x1.87de2ap-2f, dir*0x1.d906bcp-1f)); \\\n"
+ + " (a)[27] = complexMul((a)[27], (float2)(-0x1.1c73b4p-1f, dir*0x1.a9b662p-1f)); \\\n"
+ + " (a)[28] = complexMul((a)[28], (float2)(-0x1.6a09e6p-1f, dir*0x1.6a09e6p-1f)); \\\n"
+ + " (a)[29] = complexMul((a)[29], (float2)(-0x1.a9b662p-1f, dir*0x1.1c73b4p-1f)); \\\n"
+ + " (a)[30] = complexMul((a)[30], (float2)(-0x1.d906bcp-1f, dir*0x1.87de2ap-2f)); \\\n"
+ + " (a)[31] = complexMul((a)[31], (float2)(-0x1.f6297cp-1f, dir*0x1.8f8b84p-3f)); \\\n"
+ + " fftKernel16((a), dir); \\\n"
+ + " fftKernel16((a) + 16, dir); \\\n"
+ + " bitreverse32((a)); \\\n"
+ + "}\n\n";
+ static String twistKernelInterleaved =
+ "__kernel void \\\n"
+ + "clFFT_1DTwistInterleaved(__global float2 *in, unsigned int startRow, unsigned int numCols, unsigned int N, unsigned int numRowsToProcess, int dir) \\\n"
+ + "{ \\\n"
+ + " float2 a, w; \\\n"
+ + " float ang; \\\n"
+ + " unsigned int j; \\\n"
+ + " unsigned int i = get_global_id(0); \\\n"
+ + " unsigned int startIndex = i; \\\n"
+ + " \\\n"
+ + " if(i < numCols) \\\n"
+ + " { \\\n"
+ + " for(j = 0; j < numRowsToProcess; j++) \\\n"
+ + " { \\\n"
+ + " a = in[startIndex]; \\\n"
+ + " ang = 2.0f * M_PI * dir * i * (startRow + j) / N; \\\n"
+ + " w = (float2)(native_cos(ang), native_sin(ang)); \\\n"
+ + " a = complexMul(a, w); \\\n"
+ + " in[startIndex] = a; \\\n"
+ + " startIndex += numCols; \\\n"
+ + " } \\\n"
+ + " } \\\n"
+ + "} \\\n";
+ static String twistKernelPlannar =
+ "__kernel void \\\n"
+ + "clFFT_1DTwistSplit(__global float *in_real, __global float *in_imag , unsigned int startRow, unsigned int numCols, unsigned int N, unsigned int numRowsToProcess, int dir) \\\n"
+ + "{ \\\n"
+ + " float2 a, w; \\\n"
+ + " float ang; \\\n"
+ + " unsigned int j; \\\n"
+ + " unsigned int i = get_global_id(0); \\\n"
+ + " unsigned int startIndex = i; \\\n"
+ + " \\\n"
+ + " if(i < numCols) \\\n"
+ + " { \\\n"
+ + " for(j = 0; j < numRowsToProcess; j++) \\\n"
+ + " { \\\n"
+ + " a = (float2)(in_real[startIndex], in_imag[startIndex]); \\\n"
+ + " ang = 2.0f * M_PI * dir * i * (startRow + j) / N; \\\n"
+ + " w = (float2)(native_cos(ang), native_sin(ang)); \\\n"
+ + " a = complexMul(a, w); \\\n"
+ + " in_real[startIndex] = a.x; \\\n"
+ + " in_imag[startIndex] = a.y; \\\n"
+ + " startIndex += numCols; \\\n"
+ + " } \\\n"
+ + " } \\\n"
+ + "} \\\n";
+}
generated by cgit v1.2.3 (git 2.39.1) at 2024-11-28 17:45:37 +0000