P4 to Git Change 1531138 by cpaquot@cpaquot-ocl-lc-lnx on 2018/03/23 00:10:40

SWDEV-145570 - [HIP] Module
	Check for correct device id in hipDeviceGetAttribute
	Implement hipModuleLoad
	Handle kernelParams in hipModuleLaunchKernel

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/hip/hip_device.cpp#8 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_device_runtime.cpp#3 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_internal.hpp#5 edit
... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#3 edit

hipamd/api/hip/hip_device.cpp

@@ -89,6 +89,10 @@ hipError_t hipDeviceComputeCapability(int *major, int *minor, hipDevice_t device
 hipError_t hipDeviceGetCount(int* count) {
   HIP_INIT_API(count);
 
+  return ihipDeviceGetCount(count);
+}
+
+hipError_t ihipDeviceGetCount(int* count) {
   if (count == nullptr) {
     return hipErrorInvalidValue;
   }

hipamd/api/hip/hip_device_runtime.cpp

@@ -35,7 +35,7 @@ hipError_t hipChooseDevice(int* device, const hipDeviceProp_t* properties) {
   *device = 0;
   cl_uint maxMatchedCount = 0;
   int count = 0;
-  hipDeviceGetCount(&count);
+  ihipDeviceGetCount(&count);
 
   for (cl_int i = 0; i< count; ++i) {
     hipDeviceProp_t currentProp = {0};
@@ -146,10 +146,11 @@ hipError_t hipDeviceGetAttribute(int* pi, hipDeviceAttribute_t attr, int device)
     return hipErrorInvalidValue;
   }
 
-  //if (unsigned(device) >= g_context->devices().size()) {
-  //  return hipErrorInvalidDevice;
-  //}
-  //auto* deviceHandle = g_context->devices()[device];
+  int count = 0;
+  ihipDeviceGetCount(&count);
+  if (device < 0 || device >= count) {
+    return hipErrorInvalidDevice;
+  }
 
   //FIXME: should we cache the props, or just select from deviceHandle->info_?
   hipDeviceProp_t prop = {0};
@@ -253,7 +254,7 @@ hipError_t hipDeviceGetByPCIBusId(int* device, const char*pciBusIdstr) {
 
   if (sscanf (pciBusIdstr, "%04x:%02x:%02x", &pciDomainID, &pciBusID, &pciDeviceID) == 0x3) {
     int count = 0;
-    hipDeviceGetCount(&count);
+    ihipDeviceGetCount(&count);
     for (cl_int i = 0; i < count; i++) {
       int pi = 0;
       hipDevice_t dev;
@@ -312,7 +313,7 @@ hipError_t hipDeviceGetPCIBusId ( char* pciBusId, int  len, int  device ) {
   HIP_INIT_API((void*)pciBusId, len, device);
 
   int count;
-  hipDeviceGetCount(&count);
+  ihipDeviceGetCount(&count);
   if (device < 0 || device > count) {
     return hipErrorInvalidDevice;
   }
@@ -393,7 +394,7 @@ hipError_t hipGetDevice ( int* deviceId ) {
 hipError_t hipGetDeviceCount ( int* count ) {
   HIP_INIT_API(count);
 
-  return hipDeviceGetCount(count);
+  return ihipDeviceGetCount(count);
 }
 
 hipError_t hipGetDeviceFlags ( unsigned int* flags ) {

hipamd/api/hip/hip_internal.hpp

@@ -39,6 +39,6 @@ THE SOFTWARE.
 extern thread_local amd::Context* g_context;
 extern std::vector<amd::Context*> g_devices;
 
-hipError_t hipDeviceGetCount(int* count);
+hipError_t ihipDeviceGetCount(int* count);
 
 #endif // HIP_SRC_HIP_INTERNAL_H

hipamd/api/hip/hip_module.cpp

@@ -22,10 +22,13 @@ THE SOFTWARE.
 
 #include <hip/hip_runtime.h>
 #include <libelf.h>
+#include <fstream>
 
 #include "hip_internal.hpp"
 #include "platform/program.hpp"
 
+hipError_t ihipModuleLoadData(hipModule_t *module, const void *image);
+
 static uint64_t ElfSize(const void *emi)
 {
   const Elf64_Ehdr *ehdr = (const Elf64_Ehdr*)emi;
@@ -51,9 +54,19 @@ hipError_t hipModuleLoad(hipModule_t *module, const char *fname)
 {
   HIP_INIT_API(module, fname);
 
-  assert(0 && "Unimplemented");
+  if (!fname) {
+    return hipErrorInvalidValue;
+  }
 
-  return hipErrorUnknown;
+  std::ifstream file{fname};
+
+  if (!file.is_open()) {
+    return hipErrorFileNotFound;
+  }
+
+  std::vector<char> tmp{std::istreambuf_iterator<char>{file}, std::istreambuf_iterator<char>{}};
+
+  return ihipModuleLoadData(module, tmp.data());
 }
 
 
@@ -61,15 +74,26 @@ hipError_t hipModuleUnload(hipModule_t hmod)
 {
   HIP_INIT_API(hmod);
 
-  assert(0 && "Unimplemented");
+  if (hmod == nullptr) {
+    return hipErrorUnknown;
+  }
 
-  return hipErrorUnknown;
+  amd::Program* program = as_amd(reinterpret_cast<cl_program>(hmod));
+
+  program->release();
+
+  return hipSuccess;
 }
 
 hipError_t hipModuleLoadData(hipModule_t *module, const void *image)
 {
   HIP_INIT_API(module, image);
 
+  return ihipModuleLoadData(module, image);
+}
+
+hipError_t ihipModuleLoadData(hipModule_t *module, const void *image)
+{
   amd::Program* program = new amd::Program(*g_context);
   if (program == NULL) {
     return hipErrorOutOfMemory;
@@ -133,11 +157,16 @@ hipError_t hipModuleLaunchKernel(hipFunction_t f,
   amd::NDRangeContainer ndrange(3, globalWorkOffset, globalWorkSize, localWorkSize);
   amd::Command::EventWaitList waitList;
 
-  assert(!kernelParams && extra && "check this code");
   const amd::KernelSignature& signature = kernel->signature();
   for (size_t i = 0; i < signature.numParameters(); ++i) {
     const amd::KernelParameterDescriptor& desc = signature.at(i);
-    kernel->parameters().set(i, desc.size_, reinterpret_cast<address>(extra[1]) + desc.offset_);
+    if (kernelParams == nullptr) {
+      assert(extra);
+      kernel->parameters().set(i, desc.size_, reinterpret_cast<address>(extra[1]) + desc.offset_);
+    } else {
+      assert(!extra);
+      kernel->parameters().set(i, desc.size_, kernelParams[i]);
+    }
   }
 
   amd::NDRangeKernelCommand* command = new amd::NDRangeKernelCommand(*queue, waitList, *kernel, ndrange);