P4 to Git Change 1748674 by cpaquot@cpaquot-ocl-lc-lnx on 2019/02/26 11:31:59

SWDEV-145570 - [HIP] Implemented texture object for ROCm backend
	Needed to implement sampler object and return the getHsaImageObject
	for roc::Image::cpuSrd.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#118 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.hpp#34 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocmemory.hpp#15 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#73 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#22 edit


[ROCm/clr commit: 2732ad0300]

projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp

@@ -27,6 +27,7 @@
 #ifdef WITH_AMDGPU_PRO
 #include "pro/prodriver.hpp"
 #endif
+#include "platform/sampler.hpp"
 #include <cstring>
 #include <fstream>
 #include <sstream>
@@ -841,6 +842,66 @@ hsa_status_t Device::iterateCpuMemoryPoolCallback(hsa_amd_memory_pool_t pool, vo
   return HSA_STATUS_SUCCESS;
 }
 
+bool Device::createSampler(const amd::Sampler& owner, device::Sampler** sampler) const {
+  *sampler = nullptr;
+  Sampler* gpuSampler = new Sampler(*this);
+  if ((nullptr == gpuSampler) || !gpuSampler->create(owner)) {
+    delete gpuSampler;
+    return false;
+  }
+  *sampler = gpuSampler;
+  return true;
+}
+
+void Sampler::fillSampleDescriptor(hsa_ext_sampler_descriptor_t& samplerDescriptor,
+                                   const amd::Sampler& sampler) const {
+  samplerDescriptor.filter_mode = sampler.filterMode() == CL_FILTER_NEAREST
+      ? HSA_EXT_SAMPLER_FILTER_MODE_NEAREST
+      : HSA_EXT_SAMPLER_FILTER_MODE_LINEAR;
+  samplerDescriptor.coordinate_mode = sampler.normalizedCoords()
+      ? HSA_EXT_SAMPLER_COORDINATE_MODE_NORMALIZED
+      : HSA_EXT_SAMPLER_COORDINATE_MODE_UNNORMALIZED;
+  switch (sampler.addressingMode()) {
+    case CL_ADDRESS_CLAMP_TO_EDGE:
+      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_CLAMP_TO_EDGE;
+      break;
+    case CL_ADDRESS_REPEAT:
+      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_REPEAT;
+      break;
+    case CL_ADDRESS_CLAMP:
+      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_CLAMP_TO_BORDER;
+      break;
+    case CL_ADDRESS_MIRRORED_REPEAT:
+      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_MIRRORED_REPEAT;
+      break;
+    case CL_ADDRESS_NONE:
+      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_UNDEFINED;
+      break;
+    default:
+      return;
+  }
+}
+
+bool Sampler::create(const amd::Sampler& owner) {
+  hsa_ext_sampler_descriptor_t samplerDescriptor;
+  fillSampleDescriptor(samplerDescriptor, owner);
+
+  hsa_status_t status = hsa_ext_sampler_create(dev_.getBackendDevice(), &samplerDescriptor, &hsa_sampler);
+
+  if (HSA_STATUS_SUCCESS != status) {
+    return false;
+  }
+
+  hwSrd_ = reinterpret_cast<uint64_t>(hsa_sampler.handle);
+  hwState_ = reinterpret_cast<address>(hsa_sampler.handle);
+
+  return true;
+}
+
+Sampler::~Sampler() {
+  hsa_ext_sampler_destroy(dev_.getBackendDevice(), hsa_sampler);
+}
+
 bool Device::populateOCLDeviceConstants() {
   info_.available_ = true;
 

projects/clr/rocclr/runtime/device/rocm/rocdevice.hpp

@@ -59,6 +59,31 @@ class VirtualDevice;
 class PrintfDbg;
 class IProDevice;
 
+class Sampler : public device::Sampler {
+ public:
+  //! Constructor
+    Sampler(const Device& dev) : dev_(dev) {}
+
+  //! Default destructor for the device memory object
+  virtual ~Sampler();
+
+  //! Creates a device sampler from the OCL sampler state
+  bool create(const amd::Sampler& owner  //!< AMD sampler object
+              );
+
+ private:
+  void fillSampleDescriptor(hsa_ext_sampler_descriptor_t& samplerDescriptor,
+                            const amd::Sampler& sampler) const;
+  Sampler& operator=(const Sampler&);
+
+  //! Disable operator=
+  Sampler(const Sampler&);
+
+  const Device& dev_;  //!< Device object associated with the sampler
+
+  hsa_ext_sampler_t hsa_sampler;
+};
+
 // A NULL Device type used only for offline compilation
 // Only functions that are used for compilation will be in this device
 class NullDevice : public amd::Device {
@@ -283,16 +308,7 @@ class Device : public NullDevice {
   //! Sampler object allocation
   virtual bool createSampler(const amd::Sampler& owner,  //!< abstraction layer sampler object
                              device::Sampler** sampler   //!< device sampler object
-                             ) const {
-    //! \todo HSA team has to implement sampler allocation.
-    //! Currently allocate the base device class
-    *sampler = new device::Sampler();
-    if (*sampler == nullptr) {
-      return false;
-    }
-    return true;
-  }
-
+                             ) const;
 
   //! Just returns nullptr for the dummy device
   virtual device::Memory* createView(

projects/clr/rocclr/runtime/device/rocm/rocmemory.hpp

@@ -78,8 +78,6 @@ class Memory : public device::Memory {
 
   virtual uint64_t virtualAddress() const override { return reinterpret_cast<uint64_t>(getDeviceMemory()); }
 
-  virtual const address cpuSrd() const override { return reinterpret_cast<const address>(getDeviceMemory()); }
-
   // Accessors for indirect map memory object
   amd::Memory* mapMemory() const { return mapMemory_; }
 
@@ -176,9 +174,11 @@ class Image : public roc::Memory {
   size_t getDeviceDataSize() { return deviceImageInfo_.size; }
   size_t getDeviceDataAlignment() { return deviceImageInfo_.alignment; }
 
-  hsa_ext_image_t getHsaImageObject() { return hsaImageObject_; }
+  hsa_ext_image_t getHsaImageObject() const { return hsaImageObject_; }
   const hsa_ext_image_descriptor_t& getHsaImageDescriptor() const { return imageDescriptor_; }
 
+  virtual const address cpuSrd() const { return reinterpret_cast<const address>(getHsaImageObject().handle); }
+
  private:
   //! Disable copy constructor
   Image(const Buffer&);

projects/clr/rocclr/runtime/device/rocm/rocvirtual.cpp

@@ -186,35 +186,6 @@ void VirtualGPU::MemoryDependency::clear(bool all) {
   }
 }
 
-static void fillSampleDescriptor(hsa_ext_sampler_descriptor_t& samplerDescriptor,
-                                 const amd::Sampler& sampler) {
-  samplerDescriptor.filter_mode = sampler.filterMode() == CL_FILTER_NEAREST
-      ? HSA_EXT_SAMPLER_FILTER_MODE_NEAREST
-      : HSA_EXT_SAMPLER_FILTER_MODE_LINEAR;
-  samplerDescriptor.coordinate_mode = sampler.normalizedCoords()
-      ? HSA_EXT_SAMPLER_COORDINATE_MODE_NORMALIZED
-      : HSA_EXT_SAMPLER_COORDINATE_MODE_UNNORMALIZED;
-  switch (sampler.addressingMode()) {
-    case CL_ADDRESS_CLAMP_TO_EDGE:
-      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_CLAMP_TO_EDGE;
-      break;
-    case CL_ADDRESS_REPEAT:
-      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_REPEAT;
-      break;
-    case CL_ADDRESS_CLAMP:
-      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_CLAMP_TO_BORDER;
-      break;
-    case CL_ADDRESS_MIRRORED_REPEAT:
-      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_MIRRORED_REPEAT;
-      break;
-    case CL_ADDRESS_NONE:
-      samplerDescriptor.address_mode = HSA_EXT_SAMPLER_ADDRESSING_MODE_UNDEFINED;
-      break;
-    default:
-      return;
-  }
-}
-
 bool VirtualGPU::processMemObjects(const amd::Kernel& kernel, const_address params, size_t& ldsAddress) {
   Kernel& hsaKernel = const_cast<Kernel&>(static_cast<const Kernel&>(*(kernel.getDeviceKernel(dev()))));
   const amd::KernelSignature& signature = kernel.signature();
@@ -395,36 +366,10 @@ bool VirtualGPU::processMemObjects(const amd::Kernel& kernel, const_address para
       const amd::Sampler* sampler = reinterpret_cast<amd::Sampler* const*>(params +
         kernelParams.samplerObjOffset())[index];
 
-      hsa_ext_sampler_descriptor_t samplerDescriptor;
-      fillSampleDescriptor(samplerDescriptor, *sampler);
+      device::Sampler* devSampler = sampler->getDeviceSampler(dev());
 
-      hsa_ext_sampler_t hsa_sampler;
-      hsa_status_t status =
-        hsa_ext_sampler_create(dev().getBackendDevice(), &samplerDescriptor, &hsa_sampler);
-
-      if (status != HSA_STATUS_SUCCESS) {
-        // Wait on a kernel if one is outstanding
-        releaseGpuMemoryFence();
-        // Release the sampler handles allocated for the various
-        // on one or more kernel submissions
-        for (const auto& it: samplerList_) {
-          if (hsa_ext_sampler_destroy(gpu_device_, it) != HSA_STATUS_SUCCESS) {
-              LogWarning("Error destroying device sampler object!");
-          }
-        }
-
-        samplerList_.clear();
-        status = hsa_ext_sampler_create(dev().getBackendDevice(), &samplerDescriptor, &hsa_sampler);
-        if (status != HSA_STATUS_SUCCESS) {
-          LogError("Error creating device sampler object!");
-          return false;
-        }
-      }
-
-      uint64_t sampler_srd = hsa_sampler.handle;
+      uint64_t sampler_srd = devSampler->hwSrd();
       WriteAqlArgAt(const_cast<address>(params), &sampler_srd, sizeof(sampler_srd), desc.offset_);
-      samplerList_.push_back(hsa_sampler);
-      // TODO: destroy sampler.
     }
   }
 
@@ -896,17 +841,6 @@ void VirtualGPU::updateCommandsState(amd::Command* list) {
     current->release();
     current = next;
   }
-
-  // Release the sampler handles allocated for the various
-  // on one or more kernel submissions
-  for (const auto& it: samplerList_) {
-    if (hsa_ext_sampler_destroy(gpu_device_, it) != HSA_STATUS_SUCCESS) {
-      LogWarning("Error destroying device sampler object!");
-    }
-  }
-  samplerList_.clear();
-
-  return;
 }
 
 void VirtualGPU::submitReadMemory(amd::ReadMemoryCommand& cmd) {

projects/clr/rocclr/runtime/device/rocm/rocvirtual.hpp

@@ -291,12 +291,6 @@ class VirtualGPU : public device::VirtualDevice {
   std::vector<Memory*> xferWriteBuffers_;  //!< Stage write buffers
   std::vector<amd::Memory*> pinnedMems_;   //!< Pinned memory list
 
-  /**
-   * @brief Maintains the list of sampler allocated for one or more kernel
-   * submissions.
-   */
-  std::vector<hsa_ext_sampler_t> samplerList_;
-
   /**
    * @brief Indicates if a kernel dispatch is outstanding. This flag is
    * used to synchronized on kernel outputs.