P4 to Git Change 1975344 by gandryey@gera-win10 on 2019/07/29 18:38:59

SWDEV-79445 - OCL generic changes and code clean-up
	- Remove buffer allocation for unimplemented image copy optimization

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.cpp#33 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.hpp#13 edit


[ROCm/clr commit: 833d62445b]

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

@@ -695,10 +695,6 @@ KernelBlitManager::KernelBlitManager(VirtualGPU& gpu, Setup setup)
     kernels_[i] = nullptr;
   }
 
-  for (uint i = 0; i < MaxXferBuffers; ++i) {
-    xferBuffers_[i] = nullptr;
-  }
-
   completeOperation_ = false;
 }
 
@@ -721,12 +717,6 @@ KernelBlitManager::~KernelBlitManager() {
     constantBuffer_->release();
   }
 
-  for (uint i = 0; i < MaxXferBuffers; ++i) {
-    if (nullptr != xferBuffers_[i]) {
-      xferBuffers_[i]->release();
-    }
-  }
-
   delete lockXferOps_;
 }
 
@@ -788,34 +778,6 @@ bool KernelBlitManager::createProgram(Device& device) {
     return false;
   }
 
-  if (dev().settings().xferBufSize_ > 0) {
-    xferBufferSize_ = dev().settings().xferBufSize_;
-    for (uint i = 0; i < MaxXferBuffers; ++i) {
-      // Create internal xfer buffers for image copy optimization
-      xferBuffers_[i] = new (*context_) amd::Buffer(*context_, 0, xferBufferSize_);
-
-      // Assign the xfer buffer to the current virtual GPU
-      xferBuffers_[i]->setVirtualDevice(&gpu());
-
-      if ((xferBuffers_[i] != nullptr) && !xferBuffers_[i]->create(nullptr)) {
-        xferBuffers_[i]->release();
-        xferBuffers_[i] = nullptr;
-        return false;
-      } else if (xferBuffers_[i] == nullptr) {
-        return false;
-      }
-
-      //! @note Workaround for conformance allocation test.
-      //! Force GPU mem alloc.
-      //! Unaligned images require xfer optimization,
-      //! but deferred memory allocation can cause
-      //! virtual heap fragmentation for big allocations and
-      //! then fail the following test with 32 bit ISA, because
-      //! runtime runs out of 4GB space.
-      dev().getRocMemory(xferBuffers_[i]);
-    }
-  }
-
   lockXferOps_ = new amd::Monitor("Transfer Ops Lock", true);
   if (nullptr == lockXferOps_) {
     return false;

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

@@ -430,7 +430,6 @@ class KernelBlitManager : public DmaBlitManager {
   amd::Program* program_;                     //!< GPU program obejct
   amd::Kernel* kernels_[BlitTotal];           //!< GPU kernels for blit
   amd::Memory* constantBuffer_;               //!< An internal CB for blits
-  amd::Memory* xferBuffers_[MaxXferBuffers];  //!< Transfer buffers for images
   size_t xferBufferSize_;                     //!< Transfer buffer size
   amd::Monitor* lockXferOps_;                 //!< Lock transfer operation
 };