GPU TPC: Decoding: Removed GPU specific thread count tuning parameter…

…s for Decompression Kernels and added memory scratch flag to decompression context
AliceO2Group · Mar 4, 2024 · b4a04fd · b4a04fd
1 parent eb39c87
commit b4a04fd
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 15 deletions.
diff --git a/GPU/GPUTracking/DataCompression/GPUTPCDecompression.cxx b/GPU/GPUTracking/DataCompression/GPUTPCDecompression.cxx
@@ -86,10 +86,10 @@ void* GPUTPCDecompression::SetPointersTmpNativeBuffersInput(void* mem)
 void GPUTPCDecompression::RegisterMemoryAllocation()
 {
   AllocateAndInitializeLate();
-  mMemoryResInputGPU = mRec->RegisterMemoryAllocation(this, &GPUTPCDecompression::SetPointersInputGPU, GPUMemoryResource::MEMORY_INPUT_FLAG | GPUMemoryResource::MEMORY_GPU | GPUMemoryResource::MEMORY_EXTERNAL, "TPCDecompressionInput");
+  mMemoryResInputGPU = mRec->RegisterMemoryAllocation(this, &GPUTPCDecompression::SetPointersInputGPU, GPUMemoryResource::MEMORY_INPUT_FLAG | GPUMemoryResource::MEMORY_GPU | GPUMemoryResource::MEMORY_EXTERNAL | GPUMemoryResource::MEMORY_SCRATCH, "TPCDecompressionInput");
   mRec->RegisterMemoryAllocation(this, &GPUTPCDecompression::SetPointersTmpNativeBuffersGPU, GPUMemoryResource::MEMORY_SCRATCH, "TPCDecompressionTmpBuffersGPU");
-  mResourceTmpIndexes = mRec->RegisterMemoryAllocation(this, &GPUTPCDecompression::SetPointersTmpNativeBuffersOutput, GPUMemoryResource::MEMORY_OUTPUT, "TPCDecompressionTmpBuffersOutput");
-  mResourceTmpClustersOffsets = mRec->RegisterMemoryAllocation(this, &GPUTPCDecompression::SetPointersTmpNativeBuffersInput, GPUMemoryResource::MEMORY_INPUT, "TPCDecompressionTmpBuffersInput");
+  mResourceTmpIndexes = mRec->RegisterMemoryAllocation(this, &GPUTPCDecompression::SetPointersTmpNativeBuffersOutput, GPUMemoryResource::MEMORY_OUTPUT | GPUMemoryResource::MEMORY_SCRATCH, "TPCDecompressionTmpBuffersOutput");
+  mResourceTmpClustersOffsets = mRec->RegisterMemoryAllocation(this, &GPUTPCDecompression::SetPointersTmpNativeBuffersInput, GPUMemoryResource::MEMORY_INPUT | GPUMemoryResource::MEMORY_SCRATCH, "TPCDecompressionTmpBuffersInput");
 }
 
 void GPUTPCDecompression::SetMaxData(const GPUTrackingInOutPointers& io)

diff --git a/GPU/GPUTracking/Definitions/GPUDefGPUParameters.h b/GPU/GPUTracking/Definitions/GPUDefGPUParameters.h
@@ -71,9 +71,6 @@
   #define GPUCA_LB_GPUTPCGMMergerFinalize_2 256
   #define GPUCA_LB_GPUTPCCompressionKernels_step0attached 64, 2
   #define GPUCA_LB_GPUTPCCompressionKernels_step1unattached 512, 2
-  #define GPUCA_LB_GPUTPCDecompressionKernels_step0attached 64, 2
-  #define GPUCA_LB_GPUTPCDecompressionKernels_step1unattached 512, 2
-  #define GPUCA_LB_GPUTPCDecompressionUtilKernels_sortPerSectorRow 512, 2
   #define GPUCA_LB_GPUTPCCFCheckPadBaseline 64
   #define GPUCA_LB_GPUTPCCFChargeMapFiller_fillIndexMap 512
   #define GPUCA_LB_GPUTPCCFChargeMapFiller_fillFromDigits 512
@@ -138,9 +135,6 @@
   #define GPUCA_LB_GPUTPCGMMergerFinalize_2 256
   #define GPUCA_LB_GPUTPCCompressionKernels_step0attached 192, 2
   #define GPUCA_LB_GPUTPCCompressionKernels_step1unattached 512, 2
-  #define GPUCA_LB_GPUTPCDecompressionKernels_step0attached 192, 2
-  #define GPUCA_LB_GPUTPCDecompressionKernels_step1unattached 512, 2
-  #define GPUCA_LB_GPUTPCDecompressionUtilKernels_sortPerSectorRow 512, 2
   #define GPUCA_LB_GPUTPCCFCheckPadBaseline 64
   #define GPUCA_LB_GPUTPCCFChargeMapFiller_fillIndexMap 512
   #define GPUCA_LB_GPUTPCCFChargeMapFiller_fillFromDigits 512
@@ -204,10 +198,7 @@
   #define GPUCA_LB_GPUTPCGMMergerFinalize_1 256
   #define GPUCA_LB_GPUTPCGMMergerFinalize_2 256
   #define GPUCA_LB_GPUTPCCompressionKernels_step0attached 64, 2
-  #define GPUCA_LB_GPUTPCDecompressionKernels_step0attached 64, 2
   #define GPUCA_LB_GPUTPCCompressionKernels_step1unattached 512, 3
-  #define GPUCA_LB_GPUTPCDecompressionKernels_step1unattached 512, 3
-  #define GPUCA_LB_GPUTPCDecompressionUtilKernels_sortPerSectorRow 512, 3
   #define GPUCA_LB_GPUTPCCFCheckPadBaseline 64,8
   #define GPUCA_LB_GPUTPCCFChargeMapFiller_fillIndexMap 448
   #define GPUCA_LB_GPUTPCCFChargeMapFiller_fillFromDigits 448
@@ -271,10 +262,7 @@
   #define GPUCA_LB_GPUTPCGMMergerFinalize_1 256
   #define GPUCA_LB_GPUTPCGMMergerFinalize_2 256
   #define GPUCA_LB_GPUTPCCompressionKernels_step0attached 128
-  #define GPUCA_LB_GPUTPCDecompressionKernels_step0attached 128
   #define GPUCA_LB_GPUTPCCompressionKernels_step1unattached 512, 2
-  #define GPUCA_LB_GPUTPCDecompressionKernels_step1unattached 512, 2
-  #define GPUCA_LB_GPUTPCDecompressionUtilKernels_sortPerSectorRow 512, 2
   #define GPUCA_LB_COMPRESSION_GATHER 1024
   #define GPUCA_NEIGHBOURS_FINDER_MAX_NNEIGHUP 4
   #define GPUCA_TRACKLET_SELECTOR_HITS_REG_SIZE 20