truffle: src/gpu/ptx/vm/gpu_ptx.cpp comparison

comparison src/gpu/ptx/vm/gpu_ptx.cpp @ 13819:49db2c1e3bee

added support for co-existing GPU backends (JBS:GRAAL-1)

author	Doug Simon <doug.simon@oracle.com>
date	Thu, 30 Jan 2014 00:52:33 +0100
parents	80cd5c3b8827
children	5c8a3c09397b ab370d74a8eb

comparison

equal deleted inserted replaced

-:d2f520f46180
+:49db2c1e3bee
 #include "utilities/globalDefinitions.hpp"
 #include "utilities/ostream.hpp"
 #include "memory/allocation.hpp"
 #include "memory/allocation.inline.hpp"
 #include "runtime/interfaceSupport.hpp"
+#include "graal/graalEnv.hpp"
+#include "graal/graalCompiler.hpp"
 #include "ptxKernelArguments.hpp"
+// Entry to GPU native method implementation that transitions current thread to '_thread_in_vm'.
+#define GPU_VMENTRY(result_type, name, signature) \
+JNIEXPORT result_type JNICALL name signature { \
+GRAAL_VM_ENTRY_MARK; \
+// Entry to GPU native method implementation that calls a JNI function
+// and hence cannot transition current thread to '_thread_in_vm'.
+#define GPU_ENTRY(result_type, name, signature) \
+JNIEXPORT result_type JNICALL name signature { \
+#define GPU_END }
+#define CC (char*)  /*cast a literal from (const char*)*/
+#define FN_PTR(f) CAST_FROM_FN_PTR(void*, &(f))
+#define STRING                "Ljava/lang/String;"
+JNINativeMethod gpu::Ptx::PTX_methods[] = {
+{CC"initialize",              CC"()Z",               FN_PTR(gpu::Ptx::initialize)},
+{CC"generateKernel",          CC"([B" STRING ")J",   FN_PTR(gpu::Ptx::generate_kernel)},
+{CC"getLaunchKernelAddress",  CC"()J",               FN_PTR(gpu::Ptx::get_execute_kernel_from_vm_address)},
+{CC"getAvailableProcessors0", CC"()I",               FN_PTR(gpu::Ptx::get_total_cores)},
+};
 void * gpu::Ptx::_device_context;
 int    gpu::Ptx::_cu_device = 0;
 gpu::Ptx::cuda_cu_init_func_t gpu::Ptx::_cuda_cu_init;
 #define LOOKUP_CUDA_FUNCTION(name, alias)  \
 _##alias =                               \
 CAST_TO_FN_PTR(alias##_func_t, os::dll_lookup(handle, STRINGIFY(name))); \
 if (_##alias == NULL) {      \
 tty->print_cr("[CUDA] ***** Error: Failed to lookup %s", STRINGIFY(name)); \
-return 0; \
+return false; \
 } \
 #define LOOKUP_CUDA_V2_FUNCTION(name, alias)  LOOKUP_CUDA_FUNCTION(name##_v2, alias)
 /*
 * see http://en.wikipedia.org/wiki/CUDA#Supported_GPUs
 */
-int ncores(int major, int minor) {
+int gpu::Ptx::ncores(int major, int minor) {
 int device_type = (major << 4) + minor;
 switch (device_type) {
 case 0x10: return 8;
 case 0x11: return 8;
 tty->print_cr("[CUDA] Warning: Unhandled device %x", device_type);
 return 0;
 }
 }
-bool gpu::Ptx::initialize_gpu() {
+bool gpu::Ptx::register_natives(JNIEnv* env) {
+jclass klass = env->FindClass("com/oracle/graal/hotspot/ptx/PTXHotSpotBackend");
+if (klass == NULL) {
+if (TraceGPUInteraction) {
+tty->print_cr("PTXHotSpotBackend class not found");
+}
+return false;
+}
+jint status = env->RegisterNatives(klass, PTX_methods, sizeof(PTX_methods) / sizeof(JNINativeMethod));
+if (status != JNI_OK) {
+if (TraceGPUInteraction) {
+tty->print_cr("Error registering natives for PTXHotSpotBackend: %d", status);
+}
+return false;
+}
+return true;
+}
+GPU_ENTRY(jboolean, gpu::Ptx::initialize, (JNIEnv *env, jclass))
+if (!link()) {
+return false;
+}
 /* Initialize CUDA driver API */
 int status = _cuda_cu_init(0);
 if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("Failed to initialize CUDA device");
+if (TraceGPUInteraction) {
+tty->print_cr("Failed to initialize CUDA device: %d", status);
+}
 return false;
 }
 if (TraceGPUInteraction) {
 tty->print_cr("CUDA driver initialization: Success");
 if (TraceGPUInteraction) {
 tty->print_cr("[CUDA] Using %s", device_name);
 }
+gpu::initialized_gpu(device_name);
 return true;
-}
+GPU_END
-unsigned int gpu::Ptx::total_cores() {
+GPU_ENTRY(jint, gpu::Ptx::get_total_cores, (JNIEnv *env, jobject))
 int minor, major, nmp;
 int status = _cuda_cu_device_get_attribute(&minor,
 GRAAL_CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR,
 _cu_device);
 status = _cuda_cu_device_get_attribute(&nmp,
 GRAAL_CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
 _cu_device);
 if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] Failed to get numberof MPs on device: %d", _cu_device);
+tty->print_cr("[CUDA] Failed to get number of MPs on device: %d", _cu_device);
 return 0;
 }
 int total = nmp * ncores(major, minor);
 tty->print_cr("[CUDA] Compatibility version of device %d: %d.%d", _cu_device, major, minor);
 tty->print_cr("[CUDA] Number of cores: %d async engines: %d can map host mem: %d concurrent kernels: %d",
 total, async_engines, can_map_host_memory, concurrent_kernels);
 tty->print_cr("[CUDA] Max threads per block: %d warp size: %d", max_threads_per_block, warp_size);
 }
-return (total);
+return total;
+GPU_END
-}
+GPU_ENTRY(jlong, gpu::Ptx::generate_kernel, (JNIEnv *env, jclass, jbyteArray code_handle, jstring name_handle))
-void *gpu::Ptx::generate_kernel(unsigned char *code, int code_len, const char *name) {
+ResourceMark rm;
+jsize name_len = env->GetStringLength(name_handle);
+jsize code_len = env->GetArrayLength(code_handle);
+char* name = NEW_RESOURCE_ARRAY(char, name_len + 1);
+unsigned char *code = NEW_RESOURCE_ARRAY(unsigned char, code_len + 1);
+code[code_len] = 0;
+name[name_len] = 0;
+env->GetByteArrayRegion(code_handle, 0, code_len, (jbyte*) code);
+env->GetStringUTFRegion(name_handle, 0, name_len, name);
 struct CUmod_st * cu_module;
 // Use three JIT compiler options
 const unsigned int jit_num_options = 3;
-int *jit_options = NEW_C_HEAP_ARRAY(int, jit_num_options, mtCompiler);
+int *jit_options = NEW_RESOURCE_ARRAY(int, jit_num_options);
-void **jit_option_values = NEW_C_HEAP_ARRAY(void *, jit_num_options, mtCompiler);
+void **jit_option_values = NEW_RESOURCE_ARRAY(void *, jit_num_options);
 // Set up PTX JIT compiler options
 // 1. set size of compilation log buffer
 int jit_log_buffer_size = 1024;
 jit_options[0] = GRAAL_CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
 jit_option_values[0] = (void *)(size_t)jit_log_buffer_size;
 // 2. set pointer to compilation log buffer
-char *jit_log_buffer = NEW_C_HEAP_ARRAY(char, jit_log_buffer_size, mtCompiler);
+char *jit_log_buffer = NEW_RESOURCE_ARRAY(char, jit_log_buffer_size);
 jit_options[1] = GRAAL_CU_JIT_INFO_LOG_BUFFER;
 jit_option_values[1] = jit_log_buffer;
-// 3. set pointer to set the Maximum # of registers (32) for the kernel
+// 3. set pointer to set the maximum number of registers (32) for the kernel
 int jit_register_count = 32;
 jit_options[2] = GRAAL_CU_JIT_MAX_REGISTERS;
 jit_option_values[2] = (void *)(size_t)jit_register_count;
-/* Create CUDA context to compile and execute the kernel */
+// Create CUDA context to compile and execute the kernel
 int status = _cuda_cu_ctx_create(&_device_context, GRAAL_CU_CTX_MAP_HOST, _cu_device);
 if (status != GRAAL_CUDA_SUCCESS) {
 tty->print_cr("[CUDA] Failed to create CUDA context for device(%d): %d", _cu_device, status);
-return NULL;
+return 0L;
 }
 if (TraceGPUInteraction) {
 tty->print_cr("[CUDA] Success: Created context for device: %d", _cu_device);
 }
 status = _cuda_cu_ctx_set_current(_device_context);
 if (status != GRAAL_CUDA_SUCCESS) {
 tty->print_cr("[CUDA] Failed to set current context for device: %d", _cu_device);
-return NULL;
+return 0L;
 }
 if (TraceGPUInteraction) {
 tty->print_cr("[CUDA] Success: Set current context for device: %d", _cu_device);
-}
-if (TraceGPUInteraction) {
 tty->print_cr("[CUDA] PTX Kernel\n%s", code);
 tty->print_cr("[CUDA] Function name : %s", name);
 }
 /* Load module's data with compiler options */
 status = _cuda_cu_module_load_data_ex(&cu_module, (void*) code, jit_num_options,
 jit_options, (void **)jit_option_values);
 if (status != GRAAL_CUDA_SUCCESS) {
 if (status == GRAAL_CUDA_ERROR_NO_BINARY_FOR_GPU) {
 tty->print_cr("[CUDA] Check for malformed PTX kernel or incorrect PTX compilation options");
 }
 tty->print_cr("[CUDA] *** Error (%d) Failed to load module data with online compiler options for method %s",
 status, name);
-return NULL;
+return 0L;
 }
 if (TraceGPUInteraction) {
 tty->print_cr("[CUDA] Loaded data for PTX Kernel");
 }
-struct CUfunc_st * cu_function;
+struct CUfunc_st* cu_function;
 status = _cuda_cu_module_get_function(&cu_function, cu_module, name);
 if (status != GRAAL_CUDA_SUCCESS) {
 tty->print_cr("[CUDA] *** Error: Failed to get function %s", name);
-return NULL;
+return 0L;
 }
 if (TraceGPUInteraction) {
 tty->print_cr("[CUDA] Got function handle for %s kernel address %p", name, cu_function);
 }
+return (jlong) cu_function;
-return cu_function;
+GPU_END
-}
 // A PtxCall is used to manage executing a GPU kernel. In addition to launching
 // the kernel, this class releases resources allocated for the execution.
 class PtxCall: StackObj {
 private:
 free_return_value();
 destroy_context();
 }
 };
+GPU_VMENTRY(jlong, gpu::Ptx::get_execute_kernel_from_vm_address, (JNIEnv *env, jclass))
+return (jlong) gpu::Ptx::execute_kernel_from_vm;
+GPU_END
 JRT_ENTRY(jlong, gpu::Ptx::execute_kernel_from_vm(JavaThread* thread, jlong kernel, jint dimX, jint dimY, jint dimZ,
 jlong buffer,
 jint bufferSize,
 jint objectParametersCount,
 if (kernel == 0L) {
 SharedRuntime::throw_and_post_jvmti_exception(thread, vmSymbols::java_lang_NullPointerException(), NULL);
 return 0L;
 }
+#if 0
+Universe::heap()->collect(GCCause::_jvmti_force_gc);
+#endif
 PtxCall call(thread, (address) buffer, bufferSize, (oop*) (address) pinnedObjects, encodedReturnTypeSize);
 #define TRY(action) do { \
 action; \
 if (HAS_PENDING_EXCEPTION) return 0L; \
 return return_val;
 #undef TRY
 JRT_END
-bool gpu::Ptx::execute_kernel(address kernel, PTXKernelArguments &ptxka, JavaValue &ret) {
-return gpu::Ptx::execute_warp(1, 1, 1, kernel, ptxka, ret);
-}
-bool gpu::Ptx::execute_warp(int dimX, int dimY, int dimZ,
-address kernel, PTXKernelArguments &ptxka, JavaValue &ret) {
-// grid dimensionality
-unsigned int gridX = 1;
-unsigned int gridY = 1;
-unsigned int gridZ = 1;
-// thread dimensionality
-unsigned int blockX = dimX;
-unsigned int blockY = dimY;
-unsigned int blockZ = dimZ;
-struct CUfunc_st* cu_function = (struct CUfunc_st*) kernel;
-void * config[5] = {
-GRAAL_CU_LAUNCH_PARAM_BUFFER_POINTER, ptxka._kernelArgBuffer,
-GRAAL_CU_LAUNCH_PARAM_BUFFER_SIZE, &(ptxka._bufferOffset),
-GRAAL_CU_LAUNCH_PARAM_END
-};
-if (kernel == NULL) {
-return false;
-}
-if (TraceGPUInteraction) {
-tty->print_cr("[CUDA] launching kernel");
-}
-int status = _cuda_cu_launch_kernel(cu_function,
-gridX, gridY, gridZ,
-blockX, blockY, blockZ,
-0, NULL, NULL, (void **) &config);
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] Failed to launch kernel");
-return false;
-}
-if (TraceGPUInteraction) {
-tty->print_cr("[CUDA] Success: Kernel Launch: X: %d Y: %d Z: %d", blockX, blockY, blockZ);
-}
-status = _cuda_cu_ctx_synchronize();
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] Failed to synchronize launched kernel (%d)", status);
-return false;
-}
-if (TraceGPUInteraction) {
-tty->print_cr("[CUDA] Success: Synchronized launch kernel");
-}
-// Get the result. TODO: Move this code to get_return_oop()
-BasicType return_type = ptxka.get_ret_type();
-switch (return_type) {
-case T_INT:
-{
-int return_val;
-status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_INT_BYTE_SIZE);
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status);
-return false;
-}
-ret.set_jint(return_val);
-}
-break;
-case T_BOOLEAN:
-{
-int return_val;
-status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_INT_BYTE_SIZE);
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status);
-return false;
-}
-ret.set_jint(return_val);
-}
-break;
-case T_FLOAT:
-{
-float return_val;
-status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_FLOAT_BYTE_SIZE);
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status);
-return false;
-}
-ret.set_jfloat(return_val);
-}
-break;
-case T_DOUBLE:
-{
-double return_val;
-status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_DOUBLE_BYTE_SIZE);
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status);
-return false;
-}
-ret.set_jdouble(return_val);
-}
-break;
-case T_LONG:
-{
-long return_val;
-status = gpu::Ptx::_cuda_cu_memcpy_dtoh(&return_val, ptxka._dev_return_value, T_LONG_BYTE_SIZE);
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] *** Error (%d) Failed to copy value to device argument", status);
-return false;
-}
-ret.set_jlong(return_val);
-}
-break;
-case T_VOID:
-break;
-default:
-tty->print_cr("[CUDA] TODO *** Unhandled return type: %d", return_type);
-}
-// Free device memory allocated for result
-status = gpu::Ptx::_cuda_cu_memfree(ptxka._dev_return_value);
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] *** Error (%d) Failed to free device memory of return value", status);
-return false;
-}
-if (TraceGPUInteraction) {
-tty->print_cr("[CUDA] Success: Freed device memory of return value");
-}
-// Destroy context
-status = gpu::Ptx::_cuda_cu_ctx_destroy(_device_context);
-if (status != GRAAL_CUDA_SUCCESS) {
-tty->print_cr("[CUDA] *** Error (%d) Failed to destroy context", status);
-return false;
-}
-if (TraceGPUInteraction) {
-tty->print_cr("[CUDA] Success: Destroy context");
-}
-return (status == GRAAL_CUDA_SUCCESS);
-}
 #if defined(LINUX)
 static const char cuda_library_name[] = "libcuda.so";
 #elif defined(__APPLE__)
 static char const cuda_library_name[] = "/usr/local/cuda/lib/libcuda.dylib";
 #else
 static char const cuda_library_name[] = "";
 #endif
-#define STD_BUFFER_SIZE 1024
+bool gpu::Ptx::link() {
+if (cuda_library_name == NULL) {
-bool gpu::Ptx::probe_linkage() {
+if (TraceGPUInteraction) {
-if (cuda_library_name != NULL) {
+tty->print_cr("Failed to find CUDA linkage");
-char *buffer = (char*)malloc(STD_BUFFER_SIZE);
+}
-void *handle = os::dll_load(cuda_library_name, buffer, STD_BUFFER_SIZE);
+return false;
-free(buffer);
+}
-if (handle != NULL) {
+char ebuf[O_BUFLEN];
-LOOKUP_CUDA_FUNCTION(cuInit, cuda_cu_init);
+void *handle = os::dll_load(cuda_library_name, ebuf, O_BUFLEN);
-LOOKUP_CUDA_FUNCTION(cuCtxSynchronize, cuda_cu_ctx_synchronize);
+if (handle == NULL) {
-LOOKUP_CUDA_FUNCTION(cuCtxGetCurrent, cuda_cu_ctx_get_current);
+if (TraceGPUInteraction) {
-LOOKUP_CUDA_FUNCTION(cuCtxSetCurrent, cuda_cu_ctx_set_current);
+tty->print_cr("Unsupported CUDA platform: %s", ebuf);
-LOOKUP_CUDA_FUNCTION(cuDeviceGetCount, cuda_cu_device_get_count);
+}
-LOOKUP_CUDA_FUNCTION(cuDeviceGetName, cuda_cu_device_get_name);
+return false;
-LOOKUP_CUDA_FUNCTION(cuDeviceGet, cuda_cu_device_get);
+}
-LOOKUP_CUDA_FUNCTION(cuDeviceComputeCapability, cuda_cu_device_compute_capability);
-LOOKUP_CUDA_FUNCTION(cuDeviceGetAttribute, cuda_cu_device_get_attribute);
+LOOKUP_CUDA_FUNCTION(cuInit, cuda_cu_init);
-LOOKUP_CUDA_FUNCTION(cuModuleGetFunction, cuda_cu_module_get_function);
+LOOKUP_CUDA_FUNCTION(cuCtxSynchronize, cuda_cu_ctx_synchronize);
-LOOKUP_CUDA_FUNCTION(cuModuleLoadDataEx, cuda_cu_module_load_data_ex);
+LOOKUP_CUDA_FUNCTION(cuCtxGetCurrent, cuda_cu_ctx_get_current);
-LOOKUP_CUDA_FUNCTION(cuLaunchKernel, cuda_cu_launch_kernel);
+LOOKUP_CUDA_FUNCTION(cuCtxSetCurrent, cuda_cu_ctx_set_current);
-LOOKUP_CUDA_FUNCTION(cuMemHostRegister, cuda_cu_mem_host_register);
+LOOKUP_CUDA_FUNCTION(cuDeviceGetCount, cuda_cu_device_get_count);
-LOOKUP_CUDA_FUNCTION(cuMemHostUnregister, cuda_cu_mem_host_unregister);
+LOOKUP_CUDA_FUNCTION(cuDeviceGetName, cuda_cu_device_get_name);
+LOOKUP_CUDA_FUNCTION(cuDeviceGet, cuda_cu_device_get);
+LOOKUP_CUDA_FUNCTION(cuDeviceComputeCapability, cuda_cu_device_compute_capability);
+LOOKUP_CUDA_FUNCTION(cuDeviceGetAttribute, cuda_cu_device_get_attribute);
+LOOKUP_CUDA_FUNCTION(cuModuleGetFunction, cuda_cu_module_get_function);
+LOOKUP_CUDA_FUNCTION(cuModuleLoadDataEx, cuda_cu_module_load_data_ex);
+LOOKUP_CUDA_FUNCTION(cuLaunchKernel, cuda_cu_launch_kernel);
+LOOKUP_CUDA_FUNCTION(cuMemHostRegister, cuda_cu_mem_host_register);
+LOOKUP_CUDA_FUNCTION(cuMemHostUnregister, cuda_cu_mem_host_unregister);
 #if defined(__x86_64) || defined(AMD64) || defined(_M_AMD64)
 LOOKUP_CUDA_V2_FUNCTION(cuCtxCreate, cuda_cu_ctx_create);
 LOOKUP_CUDA_V2_FUNCTION(cuCtxDestroy, cuda_cu_ctx_destroy);
 LOOKUP_CUDA_V2_FUNCTION(cuMemAlloc, cuda_cu_memalloc);
 LOOKUP_CUDA_V2_FUNCTION(cuMemFree, cuda_cu_memfree);
 LOOKUP_CUDA_V2_FUNCTION(cuMemcpyHtoD, cuda_cu_memcpy_htod);
 LOOKUP_CUDA_V2_FUNCTION(cuMemcpyDtoH, cuda_cu_memcpy_dtoh);
 LOOKUP_CUDA_V2_FUNCTION(cuMemHostGetDevicePointer, cuda_cu_mem_host_get_device_pointer);
 #else
 LOOKUP_CUDA_FUNCTION(cuCtxCreate, cuda_cu_ctx_create);
 LOOKUP_CUDA_FUNCTION(cuCtxDestroy, cuda_cu_ctx_destroy);
 LOOKUP_CUDA_FUNCTION(cuMemAlloc, cuda_cu_memalloc);
 LOOKUP_CUDA_FUNCTION(cuMemFree, cuda_cu_memfree);
 LOOKUP_CUDA_FUNCTION(cuMemcpyHtoD, cuda_cu_memcpy_htod);
 LOOKUP_CUDA_FUNCTION(cuMemcpyDtoH, cuda_cu_memcpy_dtoh);
 LOOKUP_CUDA_FUNCTION(cuMemHostGetDevicePointer, cuda_cu_mem_host_get_device_pointer);
 #endif
 if (TraceGPUInteraction) {
 tty->print_cr("[CUDA] Success: library linkage");
 }
 return true;
-} else {
-// Unable to dlopen libcuda
-return false;
-}
-} else {
-tty->print_cr("Unsupported CUDA platform");
-return false;
-}
-tty->print_cr("Failed to find CUDA linkage");
-return false;
 }

Mercurial > hg > truffle

comparison src/gpu/ptx/vm/gpu_ptx.cpp @ 13819:49db2c1e3bee