diff --git a/hat/backends/ffi/opencl/cpp/opencl_backend.cpp b/hat/backends/ffi/opencl/cpp/opencl_backend.cpp
index 1aa36ab05c3..0816c7bacbf 100644
--- a/hat/backends/ffi/opencl/cpp/opencl_backend.cpp
+++ b/hat/backends/ffi/opencl/cpp/opencl_backend.cpp
@@ -22,6 +22,7 @@
* or visit www.oracle.com if you need additional information or have any
* questions.
*/
+#define opencl_backend_cpp
#include "opencl_backend.h"
OpenCLBackend::OpenCLConfig::OpenCLConfig(int mode):
@@ -29,8 +30,11 @@ OpenCLBackend::OpenCLConfig::OpenCLConfig(int mode):
gpu((mode&GPU_BIT)==GPU_BIT),
cpu((mode&CPU_BIT)==CPU_BIT),
minimizeCopies((mode&MINIMIZE_COPIES_BIT)==MINIMIZE_COPIES_BIT),
+ alwaysCopy(!minimizeCopies),
trace((mode&TRACE_BIT)==TRACE_BIT),
traceCopies((mode&TRACE_COPIES_BIT)==TRACE_COPIES_BIT),
+ traceEnqueues((mode&TRACE_ENQUEUES_BIT)==TRACE_ENQUEUES_BIT),
+ traceCalls((mode&TRACE_CALLS_BIT)==TRACE_CALLS_BIT),
traceSkippedCopies((mode&TRACE_SKIPPED_COPIES_BIT)==TRACE_SKIPPED_COPIES_BIT),
info((mode&INFO_BIT)==INFO_BIT),
showCode((mode&SHOW_CODE_BIT)==SHOW_CODE_BIT),
@@ -41,9 +45,12 @@ OpenCLBackend::OpenCLConfig::OpenCLConfig(int mode):
std::cout << "native gpu " << gpu<<std::endl;
std::cout << "native cpu " << cpu<<std::endl;
std::cout << "native minimizeCopies " << minimizeCopies<<std::endl;
+ std::cout << "native alwaysCopy " << alwaysCopy<<std::endl;
std::cout << "native trace " << trace<<std::endl;
std::cout << "native traceSkippedCopies " << traceSkippedCopies<<std::endl;
+ std::cout << "native traceCalls " << traceCalls<<std::endl;
std::cout << "native traceCopies " << traceCopies<<std::endl;
+ std::cout << "native traceEnqueues " << traceEnqueues<<std::endl;
std::cout << "native profile " << profile<<std::endl;
}
}
@@ -180,6 +187,9 @@ bool OpenCLBackend::getBufferFromDeviceIfDirty(void *memorySegment, long memoryS
std::cout << "from getBufferFromDeviceIfDirty Buffer is device dirty so attempting to get buffer from device from OpenCLBackend "<<std::endl;
// we use static cast because the ptr type is void*
static_cast<OpenCLProgram::OpenCLKernel::OpenCLBuffer *>(bufferState->vendorPtr)->copyFromDevice();
+ if (openclConfig.traceEnqueues | openclConfig.traceCopies){
+ std::cout << "copying buffer from device (from java access) "<< std::endl;
+ }
// if (openclConfig.traceCopies){
// std::cout << "copying buffer from device "<< std::endl;
// bufferState->dump("After copy from device");
diff --git a/hat/backends/ffi/opencl/cpp/opencl_backend_kernel_dispatch.cpp b/hat/backends/ffi/opencl/cpp/opencl_backend_kernel_dispatch.cpp
index ff71974b43a..e69889e66a6 100644
--- a/hat/backends/ffi/opencl/cpp/opencl_backend_kernel_dispatch.cpp
+++ b/hat/backends/ffi/opencl/cpp/opencl_backend_kernel_dispatch.cpp
@@ -27,7 +27,7 @@
void dispatchKernel(Kernel kernel, KernelContext kc, Arg ... args) {
for (int argn = 0; argn<args.length; argn++){
Arg arg = args[argn];
- if (!minimizingBuffers || (((arg.flags &JavaDirty)==JavaDirty) && kernel.readsFrom(arg))) {
+ if (alwaysCopyBuffers || (((arg.flags &JavaDirty)==JavaDirty) && kernel.readsFrom(arg))) {
enqueueCopyToDevice(arg);
}
}
@@ -36,7 +36,7 @@ void dispatchKernel(Kernel kernel, KernelContext kc, Arg ... args) {
for (int argn = 0; argn<args.length; argn++){
Arg arg = args[argn];
- if (!minimizingBuffers){
+ if (alwaysCopyBuffers){
enqueueCopyFromDevice(arg);
arg.flags = 0;
}else{
@@ -50,24 +50,42 @@ void dispatchKernel(Kernel kernel, KernelContext kc, Arg ... args) {
}
*/
-bool shouldCopyToDevice(BufferState_s *bufferState, Arg_s *arg ){
+bool shouldCopyToDevice(BufferState_s *bufferState, Arg_s *arg, bool alwaysCopy, bool showWhy){
bool kernelReadsFromThisArg = (arg->value.buffer.access==RW_BYTE) || (arg->value.buffer.access==RO_BYTE);
bool isHostDirtyOrNew = bufferState->isHostDirty() | bufferState->isHostNew();
bool result= (kernelReadsFromThisArg & isHostDirtyOrNew);
- if (result && bufferState->isDeviceDirty()){
- std::cout << "already still on GPU!"<<std::endl;
- result= false;
- }
- return result;
+
+ if (showWhy){
+ std::cout<<
+ "alwaysCopy="<<alwaysCopy
+ << " | argRW="<<(arg->value.buffer.access==RW_BYTE)
+ << " | argRO="<<(arg->value.buffer.access==RO_BYTE)
+ << " | kernelNeedsToRead="<< kernelReadsFromThisArg
+ << " | hostDirty="<< bufferState->isHostDirty()
+ << " | hostNew="<< bufferState->isHostNew()
+ << " | deviceDirty="<< bufferState->isDeviceDirty()
+ <<" so "
+ ;
+ }
+ if (result && bufferState->isDeviceDirty()){
+ result= false;
+ }
+ return alwaysCopy |result;
}
-bool shouldCopyFromDevice( BufferState_s *bufferState, Arg_s *arg ){
+bool shouldCopyFromDevice( BufferState_s *bufferState, Arg_s *arg, bool alwaysCopy, bool showWhy ){
bool kernelWroteToThisArg = (arg->value.buffer.access==WO_BYTE) | (arg->value.buffer.access==RW_BYTE);
bool result = kernelWroteToThisArg;
- //if (!result){
- // std::cout << "shouldCopyFromDevice false"<<std::endl;
- // }
- return result;
+ if (showWhy){
+ std::cout<<
+ "alwaysCopy="<<alwaysCopy
+ << " | argWO="<<(arg->value.buffer.access==WO_BYTE)
+ << " | argRW="<<(arg->value.buffer.access==RW_BYTE)
+ << " | kernelWroteToThisArg="<< kernelWroteToThisArg
+ <<" so "
+ ;
+ }
+ return alwaysCopy;
}
long OpenCLBackend::OpenCLProgram::OpenCLKernel::ndrange(void *argArray) {
@@ -78,6 +96,9 @@ long OpenCLBackend::OpenCLProgram::OpenCLKernel::ndrange(void *argArray) {
// std::cout << "Kernel name '"<< (dynamic_cast<Backend::Program::Kernel*>(this))->name<<"'"<<std::endl;
openclBackend->openclQueue.marker(openclBackend->openclQueue.EnterKernelDispatchBits,
(dynamic_cast<Backend::Program::Kernel*>(this))->name);
+ if (openclBackend->openclConfig.traceCalls){
+ std::cout << "ndrange(\"" << (dynamic_cast<Backend::Program::Kernel*>(this))->name<< "\"){"<<std::endl;
+ }
if (openclBackend->openclConfig.trace){
Sled::show(std::cout, argArray);
}
@@ -113,20 +134,23 @@ long OpenCLBackend::OpenCLProgram::OpenCLKernel::ndrange(void *argArray) {
}
openclBuffer= static_cast<OpenCLBuffer*>(bufferState->vendorPtr);
}
- if (!openclBackend->openclConfig.minimizeCopies
- || shouldCopyToDevice(bufferState, arg)){
+ if (shouldCopyToDevice(bufferState, arg,openclBackend->openclConfig.alwaysCopy,
+ (openclBackend->openclConfig.traceCopies|openclBackend->openclConfig.traceEnqueues))){
if (openclBackend->openclConfig.traceCopies){
- // std::cout << "We are not minimising copies OR (HOST is JAVA dirty and the kernel is READS this arg) so copying arg " << arg->idx <<" to device "<< std::endl;
+ std::cout << "We are always cloying OR (HOST is JAVA dirty and the kernel is READS this arg) so copying arg " << arg->idx <<" to device "<< std::endl;
}
bufferState->clearHostDirty();
+ if (openclBackend->openclConfig.traceEnqueues){
+ std::cout << "copying arg " << arg->idx <<" to device "<< std::endl;
+ }
openclBuffer->copyToDevice();
}else{
if (openclBackend->openclConfig.traceSkippedCopies){
- std::cout << "NOT copying arg " << arg->idx <<" to device "<< std::endl;
+ std::cout << "NOT copying arg " << arg->idx <<" to device "<< std::endl;
// bufferState->dump("After copy from device");
- }
+ }
}
cl_int status = clSetKernelArg(kernel, arg->idx, sizeof(cl_mem), &openclBuffer->clMem);
@@ -184,8 +208,9 @@ long OpenCLBackend::OpenCLProgram::OpenCLKernel::ndrange(void *argArray) {
std::cerr << OpenCLBackend::errorMsg(status) << std::endl;
exit(1);
}
- if (openclBackend->openclConfig.trace){
- std::cout << "enqueued kernel dispatch globalSize=" << globalSize << std::endl;
+ if (openclBackend->openclConfig.trace | openclBackend->openclConfig.traceEnqueues){
+ std::cout << "enqueued kernel dispatch \"" << (dynamic_cast<Backend::Program::Kernel*>(this))->name <<
+ "\" globalSize=" << globalSize << std::endl;
}
@@ -193,12 +218,17 @@ long OpenCLBackend::OpenCLProgram::OpenCLKernel::ndrange(void *argArray) {
Arg_s *arg = argSled.arg(i);
if (arg->variant == '&') {
BufferState_s * bufferState = BufferState_s::of(arg );
- if (!openclBackend->openclConfig.minimizeCopies || shouldCopyFromDevice(bufferState,arg)){
+ if (shouldCopyFromDevice(bufferState,arg,
+ openclBackend->openclConfig.alwaysCopy,
+ openclBackend->openclConfig.traceEnqueues)){
static_cast<OpenCLBuffer *>(bufferState->vendorPtr)->copyFromDevice();
//if (openclBackend->openclConfig.traceCopies){
//std::cout << "copying arg " << arg->idx <<" from device "<< std::endl;
// bufferState->dump("After copy from device");
//}
+ if (openclBackend->openclConfig.traceEnqueues){
+ std::cout << "copying arg " << arg->idx <<" from device "<< std::endl;
+ }
bufferState->setDeviceDirty();
}else{
if (openclBackend->openclConfig.traceSkippedCopies){
@@ -216,5 +246,8 @@ long OpenCLBackend::OpenCLProgram::OpenCLKernel::ndrange(void *argArray) {
);
openclBackend->openclQueue.wait();
openclBackend->openclQueue.release();
+ if (openclBackend->openclConfig.traceCalls){
+ std::cout << "\"" << (dynamic_cast<Backend::Program::Kernel*>(this))->name<< "\"}"<<std::endl;
+ }
return 0;
}
diff --git a/hat/backends/ffi/opencl/include/opencl_backend.h b/hat/backends/ffi/opencl/include/opencl_backend.h
index 8f6f5da2f75..cea91e9a918 100644
--- a/hat/backends/ffi/opencl/include/opencl_backend.h
+++ b/hat/backends/ffi/opencl/include/opencl_backend.h
@@ -45,33 +45,46 @@ class OpenCLBackend : public Backend {
public:
class OpenCLConfig{
public:
- const static int GPU_BIT =1<<1;
- const static int CPU_BIT =1<<2;
- const static int MINIMIZE_COPIES_BIT =1<<3;
- const static int TRACE_BIT =1<<4;
- const static int PROFILE_BIT =1<<5;
- const static int SHOW_CODE_BIT = 1 << 6;
- const static int SHOW_KERNEL_MODEL_BIT = 1 << 7;
- const static int SHOW_COMPUTE_MODEL_BIT = 1 <<8;
- const static int INFO_BIT = 1 <<9;
- const static int TRACE_COPIES_BIT = 1 <<10;
- const static int TRACE_SKIPPED_COPIES_BIT = 1 <<11;
+ // These must sync with hat/backend/ffi/Mode.java
+ // Bits 0-3 select platform id 0..5
+ // Bits 4-7 select device id 0..15
+ const static int START_BIT_IDX = 16;
+ const static int GPU_BIT =1<<START_BIT_IDX;
+ const static int CPU_BIT =1<<17;
+ const static int MINIMIZE_COPIES_BIT =1<<18;
+ const static int TRACE_BIT =1<<19;
+ const static int PROFILE_BIT =1<<20;
+ const static int SHOW_CODE_BIT = 1 << 21;
+ const static int SHOW_KERNEL_MODEL_BIT = 1 << 22;
+ const static int SHOW_COMPUTE_MODEL_BIT = 1 <<23;
+ const static int INFO_BIT = 1<<24;
+ const static int TRACE_COPIES_BIT = 1 <<25;
+ const static int TRACE_SKIPPED_COPIES_BIT = 1 <<26;
+ const static int TRACE_ENQUEUES_BIT = 1 <<27;
+ const static int TRACE_CALLS_BIT = 1 <<28;
+ const static int END_BIT_IDX = 29;
+
+ const static char *bitNames[]; // See below for out of line definition
int mode;
bool gpu;
bool cpu;
bool minimizeCopies;
+ bool alwaysCopy;
bool trace;
bool profile;
bool showCode;
bool info;
bool traceCopies;
- bool traceSkippedCopies;
+ bool traceSkippedCopies;
+ bool traceEnqueues;
+ bool traceCalls;
OpenCLConfig(int mode);
virtual ~OpenCLConfig();
};
class OpenCLQueue {
public:
- static const int CopyToDeviceBits= 1<<20;
+ const static int START_BIT_IDX =20;
+ static const int CopyToDeviceBits= 1<<START_BIT_IDX;
static const int CopyFromDeviceBits= 1<<21;
static const int NDRangeBits =1<<22;
static const int StartComputeBits= 1<<23;
@@ -80,6 +93,7 @@ class OpenCLBackend : public Backend {
static const int LeaveKernelDispatchBits= 1<<26;
static const int HasConstCharPtrArgBits = 1<<27;
static const int hasIntArgBits = 1<<28;
+ const static int END_BIT_IDX = 27;
OpenCLBackend *openclBackend;
size_t eventMax;
cl_event *events;
@@ -166,4 +180,21 @@ class OpenCLBackend : public Backend {
public:
static const char *errorMsg(cl_int status);
};
-extern "C" long getOpenCLBackend(int mode, int platform, int device, int unused);
\ No newline at end of file
+extern "C" long getOpenCLBackend(int mode, int platform, int device, int unused);
+#ifdef opencl_backend_cpp
+const char *OpenCLBackend::OpenCLConfig::bitNames[] = {
+ "GPU",
+ "CPU",
+ "MINIMIZE_COPIES",
+ "TRACE",
+ "PROFILE",
+ "SHOW_CODE",
+ "SHOW_KERNEL_MODEL",
+ "SHOW_COMPUTE_MODEL",
+ "INFO",
+ "TRACE_COPIES",
+ "TRACE_SKIPPED_COPIES",
+ "TRACE_ENQUEUES",
+ "TRACE_CALLS"
+ };
+#endif
\ No newline at end of file
diff --git a/hat/backends/ffi/opencl/src/main/java/hat/backend/ffi/Config.java b/hat/backends/ffi/opencl/src/main/java/hat/backend/ffi/Config.java
new file mode 100644
index 00000000000..c2c05003cfb
--- /dev/null
+++ b/hat/backends/ffi/opencl/src/main/java/hat/backend/ffi/Config.java
@@ -0,0 +1,273 @@
+package hat.backend.ffi;
+
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.List;
+
+public record Config(int bits) {
+ record Bit(int index, String name){};
+ // These must sync with hat/backends/ffi/opencl/include/opencl_backend.h
+ // Bits 0-3 select platform id 0..5
+ // Bits 4-7 select device id 0..15
+ private static final int START_BIT_IDX = 16;
+ private static final int GPU_BIT = 1 << START_BIT_IDX;
+ private static final int CPU_BIT = 1 << 17;
+ private static final int MINIMIZE_COPIES_BIT = 1 << 18;
+ private static final int TRACE_BIT = 1 << 19;
+ private static final int PROFILE_BIT = 1 << 20;
+ private static final int SHOW_CODE_BIT = 1 << 21;
+ private static final int SHOW_KERNEL_MODEL_BIT = 1 << 22;
+ private static final int SHOW_COMPUTE_MODEL_BIT = 1 << 23;
+ private static final int INFO_BIT = 1 << 24;
+ private static final int TRACE_COPIES_BIT = 1 << 25;
+ private static final int TRACE_SKIPPED_COPIES_BIT = 1 << 26;
+ private static final int TRACE_ENQUEUES_BIT = 1 << 27;
+ private static final int TRACE_CALLS_BIT = 1 << 28;
+ private static final int END_BIT_IDX = 29;
+
+ private static String[] bitNames = {
+ "GPU",
+ "CPU",
+ "MINIMIZE_COPIES",
+ "TRACE",
+ "PROFILE",
+ "SHOW_CODE",
+ "SHOW_KERNEL_MODEL",
+ "SHOW_COMPUTE_MODEL",
+ "INFO",
+ "TRACE_COPIES",
+ "TRACE_SKIPPED_COPIES",
+ "TRACE_ENQUEUES",
+ "TRACE_CALLS"
+ };
+ public static Config of() {
+ if ((((System.getenv("HAT") instanceof String e) ? e : "") +
+ ((System.getProperty("HAT") instanceof String p) ? p : "")) instanceof String opts) {
+ return of(opts);
+ }
+ return of();
+ }
+
+ public static Config of(int bits) {
+ return new Config(bits);
+ }
+
+ public static Config of(List<Config> configs) {
+ int allBits = 0;
+ for (Config config : configs) {
+ allBits |= config.bits;
+ }
+ return new Config(allBits);
+ }
+
+ public static Config of(Config... configs) {
+ return of(List.of(configs));
+ }
+
+ public Config and(Config... configs) {
+ return Config.of(Config.of(List.of(configs)).bits & bits);
+ }
+
+ public Config or(Config... configs) {
+ return Config.of(Config.of(List.of(configs)).bits | bits);
+ }
+
+ public static Config of(String name) {
+ for (int i = 0; i < bitNames.length; i++) {
+ if (bitNames[i].equals(name)) {
+ return new Config(1<<(i+START_BIT_IDX));
+ }
+ }
+
+ if (name.contains(",")) {
+ List<Config> configs = new ArrayList<>();
+ Arrays.stream(name.split(",")).forEach(opt ->
+ configs.add(of(opt))
+ );
+ return of(configs);
+ } else {
+ System.out.println("Unexpected opt '" + name + "'");
+ return Config.of(0);
+ }
+ }
+
+ public static Config TRACE_COPIES() {
+ return new Config(TRACE_COPIES_BIT);
+ }
+ public boolean isTRACE_COPIES() {
+ return (bits & TRACE_COPIES_BIT) == TRACE_COPIES_BIT;
+ }
+ public static Config TRACE_CALLS() {
+ return new Config(TRACE_CALLS_BIT);
+ }
+ public boolean isTRACE_CALLS() {
+ return (bits & TRACE_CALLS_BIT) == TRACE_CALLS_BIT;
+ }
+ public static Config TRACE_ENQUEUES() {
+ return new Config(TRACE_ENQUEUES_BIT);
+ }
+ public boolean isTRACE_ENQUEUES() {
+ return (bits & TRACE_ENQUEUES_BIT) == TRACE_ENQUEUES_BIT;
+ }
+
+
+ public static Config TRACE_SKIPPED_COPIES() {
+ return new Config(TRACE_SKIPPED_COPIES_BIT);
+ }
+ public boolean isTRACE_SKIPPED_COPIES() {
+ return (bits & TRACE_SKIPPED_COPIES_BIT) == TRACE_SKIPPED_COPIES_BIT;
+ }
+
+ public static Config INFO() {
+ return new Config(INFO_BIT);
+ }
+ public boolean isINFO() {
+ return (bits & INFO_BIT) == INFO_BIT;
+ }
+
+ public static Config CPU() {
+ return new Config(CPU_BIT);
+ }
+ public boolean isCPU() {
+ return (bits & CPU_BIT) == CPU_BIT;
+ }
+
+ public static Config GPU() {
+ return new Config(GPU_BIT);
+ }
+ public boolean isGPU() {
+ return (bits & GPU_BIT) == GPU_BIT;
+ }
+
+ public static Config PROFILE() {
+ return new Config(PROFILE_BIT);
+ }
+ public boolean isPROFILE() {
+ return (bits & PROFILE_BIT) == PROFILE_BIT;
+ }
+
+ public static Config TRACE() {
+ return new Config(TRACE_BIT);
+ }
+ public boolean isTRACE() {
+ return (bits & TRACE_BIT) == TRACE_BIT;
+ }
+
+ public static Config MINIMIZE_COPIES() {
+ return new Config(MINIMIZE_COPIES_BIT);
+ }
+ public boolean isMINIMIZE_COPIES() {
+ String hex = Integer.toHexString(bits);
+ return (bits & MINIMIZE_COPIES_BIT) == MINIMIZE_COPIES_BIT;
+ }
+
+ public static Config SHOW_CODE() {
+ return new Config(SHOW_CODE_BIT);
+ }
+ public boolean isSHOW_CODE() {
+ return (bits & SHOW_CODE_BIT) == SHOW_CODE_BIT;
+ }
+
+ public static Config SHOW_KERNEL_MODEL() {
+ return new Config(SHOW_KERNEL_MODEL_BIT);
+ }
+ public boolean isSHOW_KERNEL_MODEL() {
+ return (bits & SHOW_KERNEL_MODEL_BIT) == SHOW_KERNEL_MODEL_BIT;
+ }
+
+ public static Config SHOW_COMPUTE_MODEL() {
+ return new Config(SHOW_COMPUTE_MODEL_BIT);
+ }
+ public boolean isSHOW_COMPUTE_MODEL() {
+ return (bits & SHOW_COMPUTE_MODEL_BIT) == SHOW_COMPUTE_MODEL_BIT;
+ }
+
+ @Override
+ public String toString() {
+ StringBuilder builder = new StringBuilder();
+ for (int bitIdx = START_BIT_IDX; bitIdx < END_BIT_IDX; bitIdx++) {
+ if ((bits&(1<<bitIdx))==(1<<bitIdx)) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append(bitNames[bitIdx-START_BIT_IDX]);
+
+ }
+ }
+ /*
+ if (isTRACE_COPIES()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("TRACE_COPIES");
+ }
+ if (isTRACE_SKIPPED_COPIES()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("TRACE_SKIPPED_COPIES");
+ }
+ if (isINFO()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("INFO");
+ }
+ if (isCPU()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("CPU");
+ }
+ if (isGPU()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("GPU");
+ }
+ if (isTRACE()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("TRACE");
+ }
+ if (isPROFILE()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("PROFILE");
+ }
+ if (isMINIMIZE_COPIES()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("MINIMIZE_COPIES");
+ }
+ if (isSHOW_CODE()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("SHOW_CODE");
+ }
+ if (isSHOW_COMPUTE_MODEL()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("SHOW_COMPUTE_MODEL");
+ }
+ if (isSHOW_KERNEL_MODEL()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("SHOW_KERNEL_MODEL");
+ }
+ if (isMINIMIZE_COPIES()) {
+ if (!builder.isEmpty()) {
+ builder.append("|");
+ }
+ builder.append("MINIMIZE_COPIES");
+ } */
+
+ return builder.toString();
+ }
+}
diff --git a/hat/backends/ffi/opencl/src/main/java/hat/backend/ffi/OpenCLBackend.java b/hat/backends/ffi/opencl/src/main/java/hat/backend/ffi/OpenCLBackend.java
index 65bb520a435..2d5815a26e8 100644
--- a/hat/backends/ffi/opencl/src/main/java/hat/backend/ffi/OpenCLBackend.java
+++ b/hat/backends/ffi/opencl/src/main/java/hat/backend/ffi/OpenCLBackend.java
@@ -32,233 +32,12 @@
import hat.callgraph.KernelCallGraph;
import java.lang.invoke.MethodHandle;
-import java.util.ArrayList;
-import java.util.Arrays;
-import java.util.List;
import static java.lang.foreign.ValueLayout.JAVA_INT;
public class OpenCLBackend extends C99FFIBackend implements BufferTracker {
- public record Mode(int bits) {
- private static final int GPU_BIT = 1 << 1;
- private static final int CPU_BIT = 1 << 2;
- private static final int MINIMIZE_COPIES_BIT = 1 << 3;
- private static final int TRACE_BIT = 1 << 4;
- private static final int PROFILE_BIT = 1 << 5;
- private static final int SHOW_CODE_BIT = 1 << 6;
- private static final int SHOW_KERNEL_MODEL_BIT = 1 << 7;
- private static final int SHOW_COMPUTE_MODEL_BIT = 1 <<8;
- private static final int INFO_BIT = 1 <<9;
- private static final int TRACE_COPIES_BIT = 1 << 10;
- private static final int TRACE_SKIPPED_COPIES_BIT = 1 << 11;
-
- public static Mode of() {
- // List<Mode> modes = new ArrayList<>();
- if (( ((System.getenv("HAT") instanceof String e)?e:"")+
- ((System.getProperty("HAT") instanceof String p)?p:"")) instanceof String opts) {
- // Arrays.stream(opts.split(",")).forEach(opt ->
- return of(opts);
- // );
- }
- return of();
- }
- public static Mode of(int bits) {
-
- return new Mode(bits);
- }
- public static Mode of(List<Mode> modes) {
- int allBits = 0;
- for (Mode mode : modes) {
- allBits |= mode.bits;
- }
- return new Mode(allBits);
- }
- public static Mode of(Mode ...modes) {
- return of(List.of(modes));
- }
- public Mode and(Mode ...modes) {
- return Mode.of(Mode.of(List.of(modes)).bits&bits);
- }
- public Mode or(Mode ...modes) {
- return Mode.of(Mode.of(List.of(modes)).bits|bits);
- }
- public static Mode of(String name) {
- return switch (name){
- case "GPU" -> GPU();
- case "CPU" -> CPU();
- case "MINIMIZE_COPIES" -> MINIMIZE_COPIES();
- case "TRACE" -> TRACE();
- case "TRACE_COPIES" -> TRACE_COPIES();
- case "TRACE_SKIPPED_COPIES" -> TRACE_SKIPPED_COPIES();
- case "SHOW_CODE" -> SHOW_CODE();
- case "SHOW_KERNEL_MODEL" -> SHOW_KERNEL_MODEL();
- case "SHOW_COMPUTE_MODEL" -> SHOW_COMPUTE_MODEL();
- case "PROFILE" -> PROFILE();
- case "INFO" -> INFO();
- default -> {
- if (name.contains(",")) {
- List<Mode> modes = new ArrayList<>();
- Arrays.stream(name.split(",")).forEach(opt ->
- modes.add(of(opt))
- );
- yield of(modes);
- } else {
- System.out.println("Unexpected opt '" + name + "'");
- yield Mode.of(0);
- }
- }
- };
- }
- public static Mode TRACE_COPIES() {
- return new Mode(TRACE_COPIES_BIT);
- }
- public boolean isTRACE_COPIES() {
- return (bits&TRACE_COPIES_BIT)==TRACE_COPIES_BIT;
- }
- public static Mode TRACE_SKIPPED_COPIES() {
- return new Mode(TRACE_SKIPPED_COPIES_BIT);
- }
- public boolean isTRACE_SKIPPED_COPIES() {
- return (bits&TRACE_SKIPPED_COPIES_BIT)==TRACE_SKIPPED_COPIES_BIT;
- }
- public static Mode INFO() {
- return new Mode(INFO_BIT);
- }
- public boolean isINFO() {
- return (bits&INFO_BIT)==INFO_BIT;
- }
- public static Mode CPU() {
- return new Mode(CPU_BIT);
- }
- public boolean isCPU() {
- return (bits&CPU_BIT)==CPU_BIT;
- }
- public static Mode GPU() {
- return new Mode(GPU_BIT);
- }
- public boolean isGPU() {
- return (bits&GPU_BIT)==GPU_BIT;
- }
- public static Mode PROFILE() {
- return new Mode(PROFILE_BIT);
- }
- public boolean isPROFILE() {
- return (bits&PROFILE_BIT)==PROFILE_BIT;
- }
- public static Mode TRACE() {
- return new Mode(TRACE_BIT);
- }
- public boolean isTRACE() {
- return (bits&TRACE_BIT)==TRACE_BIT;
- }
- public static Mode MINIMIZE_COPIES() {
- return new Mode(MINIMIZE_COPIES_BIT);
- }
- public boolean isMINIMIZE_COPIES() {
- String hex = Integer.toHexString(bits);
- return (bits&MINIMIZE_COPIES_BIT)==MINIMIZE_COPIES_BIT;
- }
- public static Mode SHOW_CODE() {
- return new Mode(SHOW_CODE_BIT);
- }
- public boolean isSHOW_CODE() {
- return (bits&SHOW_CODE_BIT)==SHOW_CODE_BIT;
- }
- public static Mode SHOW_KERNEL_MODEL() {
- return new Mode(SHOW_KERNEL_MODEL_BIT);
- }
- public boolean isSHOW_KERNEL_MODEL() {
- return (bits&SHOW_KERNEL_MODEL_BIT)==SHOW_KERNEL_MODEL_BIT;
- }
- public static Mode SHOW_COMPUTE_MODEL() {
- return new Mode(SHOW_COMPUTE_MODEL_BIT);
- }
- public boolean isSHOW_COMPUTE_MODEL() {
- return (bits&SHOW_COMPUTE_MODEL_BIT)==SHOW_COMPUTE_MODEL_BIT;
- }
-
- @Override
- public String toString() {
- StringBuilder builder = new StringBuilder();
- if (isTRACE_COPIES()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("TRACE_COPIES");
- }
- if (isTRACE_SKIPPED_COPIES()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("TRACE_SKIPPED_COPIES");
- }
- if (isINFO()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("INFO");
- }
- if (isCPU()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("CPU");
- }
- if (isGPU()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("GPU");
- }
- if (isTRACE()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("TRACE");
- }
- if (isPROFILE()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("PROFILE");
- }
- if (isMINIMIZE_COPIES()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("MINIMIZE_COPIES");
- }
- if (isSHOW_CODE()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("SHOW_CODE");
- }
- if (isSHOW_COMPUTE_MODEL()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("SHOW_COMPUTE_MODEL");
- }
- if (isSHOW_KERNEL_MODEL()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("SHOW_KERNEL_MODEL");
- }
- if (isMINIMIZE_COPIES()) {
- if (!builder.isEmpty()){
- builder.append("|");
- }
- builder.append("MINIMIZE_COPIES");
- }
-
- return builder.toString();
- }
- }
-
- final Mode mode;
+ final Config config;
final MethodHandle getBackend_MH;
public long getBackend(int mode, int platform, int device, int unused) {
@@ -269,38 +48,38 @@ public long getBackend(int mode, int platform, int device, int unused) {
}
return backendHandle;
}
- public OpenCLBackend(String spec) {
- this(Mode.of(spec));
+ public OpenCLBackend(String configSpec) {
+ this(Config.of(configSpec));
}
- public OpenCLBackend(Mode mode) {
+ public OpenCLBackend(Config config) {
super("opencl_backend");
- this.mode = mode;
+ this.config = config;
getBackend_MH = nativeLibrary.longFunc("getOpenCLBackend",JAVA_INT,JAVA_INT, JAVA_INT, JAVA_INT);
- getBackend(mode.bits,0, 0, 0 );
- if (mode.isINFO()) {
- System.out.println(mode);
+ getBackend(config.bits(),0, 0, 0 );
+ if (config.isINFO()) {
+ System.out.println("CONFIG = "+config);
info();
}
}
public OpenCLBackend() {
- this(Mode.of().or(Mode.GPU()));
+ this(Config.of().or(Config.GPU()));
}
@Override
public void computeContextHandoff(ComputeContext computeContext) {
//System.out.println("OpenCL backend received computeContext");
- injectBufferTracking(computeContext.computeCallGraph.entrypoint, mode.isSHOW_COMPUTE_MODEL());
+ injectBufferTracking(computeContext.computeCallGraph.entrypoint, config.isSHOW_COMPUTE_MODEL());
}
@Override
public void dispatchKernel(KernelCallGraph kernelCallGraph, NDRange ndRange, Object... args) {
//System.out.println("OpenCL backend dispatching kernel " + kernelCallGraph.entrypoint.method);
CompiledKernel compiledKernel = kernelCallGraphCompiledCodeMap.computeIfAbsent(kernelCallGraph, (_) -> {
- String code = createCode(kernelCallGraph, new OpenCLHatKernelBuilder(), args, mode.isSHOW_KERNEL_MODEL());
- if (mode.isSHOW_CODE()) {
+ String code = createCode(kernelCallGraph, new OpenCLHatKernelBuilder(), args, config.isSHOW_KERNEL_MODEL());
+ if (config.isSHOW_CODE()) {
System.out.println(code);
}
long programHandle = compileProgram(code);
@@ -317,7 +96,7 @@ public void dispatchKernel(KernelCallGraph kernelCallGraph, NDRange ndRange, Obj
@Override
public void preMutate(Buffer b) {
- if (mode.isMINIMIZE_COPIES()) {
+ if (config.isMINIMIZE_COPIES()) {
if (b.isDeviceDirty()) {
if (!b.isHostChecked()) {
getBufferFromDeviceIfDirty(b);// calls through FFI and might block when fetching from device
@@ -330,14 +109,14 @@ public void preMutate(Buffer b) {
@Override
public void postMutate(Buffer b) {
- if (mode.isMINIMIZE_COPIES()) {
+ if (config.isMINIMIZE_COPIES()) {
b.setHostDirty();
}
}
@Override
public void preAccess(Buffer b) {
- if (mode.isMINIMIZE_COPIES()) {
+ if (config.isMINIMIZE_COPIES()) {
if (b.isDeviceDirty() && !b.isHostChecked()) {
getBufferFromDeviceIfDirty(b); // calls through FFI and might block when fetching from device
// We don't call clearDeviceDirty() if we did then 'just reading on the host' would force copy in next dispatch
@@ -354,7 +133,7 @@ public void postAccess(Buffer b) {
@Override
public void preEscape(Buffer b) {
- if (mode.isMINIMIZE_COPIES()) {
+ if (config.isMINIMIZE_COPIES()) {
if (b.isDeviceDirty()) {
if (!b.isHostChecked()) {
getBufferFromDeviceIfDirty(b);
@@ -367,7 +146,7 @@ public void preEscape(Buffer b) {
@Override
public void postEscape(Buffer b) {
- if (mode.isMINIMIZE_COPIES()) {
+ if (config.isMINIMIZE_COPIES()) {
b.setHostDirty(); // We have no choice but to assume escapee was modified by the call
}
}
diff --git a/hat/examples/experiments/src/main/java/experiments/Mesh.java b/hat/examples/experiments/src/main/java/experiments/Mesh.java
index 22347ed746b..ecde0c9c825 100644
--- a/hat/examples/experiments/src/main/java/experiments/Mesh.java
+++ b/hat/examples/experiments/src/main/java/experiments/Mesh.java
@@ -28,12 +28,10 @@
import hat.ComputeContext;
import hat.KernelContext;
import hat.backend.ffi.OpenCLBackend;
-import static hat.backend.ffi.OpenCLBackend.Mode.*;
+import static hat.backend.ffi.Config.*;
import hat.ifacemapper.BoundSchema;
import hat.ifacemapper.Schema;
-import hat.backend.DebugBackend;
import hat.buffer.Buffer;
-import hat.buffer.BufferAllocator;
import java.lang.foreign.GroupLayout;
import java.lang.foreign.MemoryLayout;
diff --git a/hat/examples/experiments/src/main/java/experiments/MinBufferTest.java b/hat/examples/experiments/src/main/java/experiments/MinBufferTest.java
index c6b256f1a8b..31de41465a2 100644
--- a/hat/examples/experiments/src/main/java/experiments/MinBufferTest.java
+++ b/hat/examples/experiments/src/main/java/experiments/MinBufferTest.java
@@ -28,14 +28,13 @@
import hat.ComputeContext;
import hat.KernelContext;
import hat.backend.ffi.OpenCLBackend;
-import hat.buffer.Buffer;
import hat.buffer.S32Array;
import static hat.ifacemapper.MappableIface.*;
import jdk.incubator.code.CodeReflection;
import java.lang.invoke.MethodHandles;
-import static hat.backend.ffi.OpenCLBackend.Mode.*;
+import static hat.backend.ffi.Config.*;
public class MinBufferTest {
diff --git a/hat/examples/life/src/main/java/life/Viewer.java b/hat/examples/life/src/main/java/life/Viewer.java
index 76c3eaad533..ef875c66f1b 100644
--- a/hat/examples/life/src/main/java/life/Viewer.java
+++ b/hat/examples/life/src/main/java/life/Viewer.java
@@ -184,8 +184,8 @@ public void paint(Graphics g) {
}
public static class Controls {
-public final JMenuBar menuBar;
- private final JButton startButton;
+ public final JMenuBar menuBar;
+ private final JButton startButton;
private JToggleButton useGPUToggleButton;
private JToggleButton minimizeCopiesToggleButton;
private SevenSegmentDisplay generationsPerSecondSevenSegment;
@@ -197,7 +197,7 @@ public static class Controls {
this.state = state;
this.menuBar = new JMenuBar();
JPanel panel = new JPanel();
- panel.setLayout(new BoxLayout(panel, BoxLayout.X_AXIS));
+ panel.setLayout(new BoxLayout(panel, BoxLayout.X_AXIS));
((JButton) panel.add(new JButton("Exit"))).addActionListener(_ -> System.exit(0));
this.startButton = (JButton) panel.add(new JButton("Start"));
diff --git a/hat/examples/mandel/src/main/java/mandel/Main.java b/hat/examples/mandel/src/main/java/mandel/Main.java
index 17b7938310a..63b02345e8a 100644
--- a/hat/examples/mandel/src/main/java/mandel/Main.java
+++ b/hat/examples/mandel/src/main/java/mandel/Main.java
@@ -137,9 +137,12 @@ public static void main(String[] args) {
final float fy = y - sign * zoomPoint.y / zoomFrames;
accelerator.compute(cc -> Main.compute(cc, pallette, s32Array2D, fx, fy, fscale));
viewer.imageViewer.syncWithRGB(s32Array2D);
+
}
}
- System.out.println("FPS = " + ((zoomFrames * 2 * 1000) / (System.currentTimeMillis() - startMillis)));
+ var fps = ((zoomFrames * 2 * 1000) / (System.currentTimeMillis() - startMillis));
+ viewer.framesSecondSevenSegment.set((int)fps);
+ // System.out.println("FPS = " +fps);
}
}
}
diff --git a/hat/examples/mandel/src/main/java/mandel/Viewer.java b/hat/examples/mandel/src/main/java/mandel/Viewer.java
index e36727c31b1..9bc427c6b8c 100644
--- a/hat/examples/mandel/src/main/java/mandel/Viewer.java
+++ b/hat/examples/mandel/src/main/java/mandel/Viewer.java
@@ -26,9 +26,16 @@
import hat.buffer.Buffer;
import hat.buffer.S32Array2D;
+import hat.util.ui.SevenSegmentDisplay;
+import javax.swing.Box;
+import javax.swing.BoxLayout;
+import javax.swing.JButton;
import javax.swing.JComponent;
import javax.swing.JFrame;
+import javax.swing.JLabel;
+import javax.swing.JMenuBar;
+import javax.swing.JPanel;
import javax.swing.WindowConstants;
import java.awt.Dimension;
import java.awt.Graphics;
@@ -45,6 +52,8 @@
public class Viewer extends JFrame {
+ public final SevenSegmentDisplay framesSecondSevenSegment;
+
public static class PointF32 {
public final float x;
public final float y;
@@ -124,6 +133,16 @@ public Viewer(String title, S32Array2D s32Array2D) {
super(title);
this.imageViewer = new ImageViewer(new BufferedImage(s32Array2D.width(), s32Array2D.height(), BufferedImage.TYPE_INT_RGB));
+ var menuBar = new JMenuBar();
+ JPanel panel = new JPanel();
+ panel.setLayout(new BoxLayout(panel, BoxLayout.X_AXIS));
+ ((JButton) panel.add(new JButton("Exit"))).addActionListener(_ -> System.exit(0));
+ panel.add(new JLabel("FPS"));
+ this.framesSecondSevenSegment = (SevenSegmentDisplay)
+ panel.add(new SevenSegmentDisplay(3,30,panel.getForeground(),panel.getBackground()));
+ panel.add(Box.createHorizontalStrut(400));
+ menuBar.add(panel);
+ this.setJMenuBar(menuBar);
this.getContentPane().add(this.imageViewer);
this.pack();
this.setLocationRelativeTo(null);
diff --git a/hat/examples/nbody/src/main/java/nbody/opencl/OpenCLNBodyGLWindow.java b/hat/examples/nbody/src/main/java/nbody/opencl/OpenCLNBodyGLWindow.java
index ae08b4208d2..b6e29b33581 100644
--- a/hat/examples/nbody/src/main/java/nbody/opencl/OpenCLNBodyGLWindow.java
+++ b/hat/examples/nbody/src/main/java/nbody/opencl/OpenCLNBodyGLWindow.java
@@ -28,6 +28,7 @@
import hat.Accelerator;
import hat.ComputeContext;
import hat.KernelContext;
+import hat.backend.ffi.Config;
import hat.backend.ffi.OpenCLBackend;
import hat.ifacemapper.SegmentMapper;
import jdk.incubator.code.CodeReflection;
@@ -116,7 +117,7 @@ public OpenCLNBodyGLWindow(Arena arena, int width, int height, GLTexture particl
super(arena, width, height, particle, bodyCount, mode);
final float maxDist = 80f;
accelerator = new Accelerator(MethodHandles.lookup(),
- new OpenCLBackend(OpenCLBackend.Mode.of("GPU"))
+ new OpenCLBackend(Config.of("GPU"))
);
universe = Universe.create(accelerator, bodyCount);
for (int body = 0; body < bodyCount; body++) {