ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-30 19:57:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

print hipOccupancyDefined Grid size for best in ckProfiler

Browse Source
This commit is contained in:
ozturkosu
2025-06-22 19:29:34 -04:00
parent b0e4ef4f96
commit 256aec203e
4 changed files with 37 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
include/ck/tensor_operation/gpu/device/device_base.hpp Executable file → Normal file
View File

@@ -56,17 +56,10 @@ struct BaseArgument
virtual ~BaseArgument() {}
void* p_workspace_ = nullptr;
virtual dim3 GetLaunchGridDims() const
{
return dim3{0, 0, 0};
}
virtual bool HasLaunchGridDims() const
{
return false;
}
virtual dim3 GetLaunchGridDims() const { return dim3{0, 0, 0}; }
virtual bool HasLaunchGridDims() const { return false; }
};
struct BaseInvoker

17
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp Executable file → Normal file
View File

@@ -168,17 +168,18 @@ struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout
hip_check_error(hipGetDevice(&dev));
hip_check_error(hipGetDeviceProperties(&dev_prop, dev));
num_cu = dev_prop.multiProcessorCount;
// arg.Grid_size = num_cu * occupancy;
arg.Grid_size = num_cu * occupancy;
// grid_dim = arg.Grid_size;
grid_dim.x = num_cu * occupancy; // Set the x-dimension
// TODO: Set grid_dim.y and grid_dim.z appropriately if they are not 1.
// This often comes from the block_2_ctile_map.CalculateGridSize(...)
// For now, assuming they might be 1 or derived from block_2_ctile_map elsewhere if needed.
// If block_2_ctile_map.CalculateGridSize gives (N0, M0, k_split), then
// grid_dim might be (N0, M0, k_split) or (total_blocks, 1, 1)
// The current code sets grid_dim = arg.Grid_size (if positive) or occupancy-based (if negative)
// which implies a 1D grid of blocks. We'll stick to that interpretation for grid_dim.x
// For now, assuming they might be 1 or derived from block_2_ctile_map elsewhere
// if needed. If block_2_ctile_map.CalculateGridSize gives (N0, M0, k_split),
// then grid_dim might be (N0, M0, k_split) or (total_blocks, 1, 1) The current
// code sets grid_dim = arg.Grid_size (if positive) or occupancy-based (if
// negative) which implies a 1D grid of blocks. We'll stick to that
// interpretation for grid_dim.x
grid_dim.y = 1;
grid_dim.z = 1;
}
@@ -190,7 +191,7 @@ struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout
grid_dim.y = 1;
grid_dim.z = 1;
}
arg.SetLaunchGridDims(grid_dim); // Store the determined launch grid dimensions
if(stream_config.flush_cache)
@@ -819,7 +820,7 @@ struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
<< "BlkGemmPipelinePrefetchStages: "
<< GridwiseGemm::BlockwiseGemmPipe::PrefetchStages ;
// clang-format on
return str.str();

15
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_streamk_v3.hpp Executable file → Normal file
View File

@@ -631,19 +631,10 @@ struct GridwiseGemm_xdl_cshuffle_streamk_v3
mutable dim3 launch_grid_dims_;
void SetLaunchGridDims(dim3 dims) const
{
launch_grid_dims_ = dims;
}
void SetLaunchGridDims(dim3 dims) const { launch_grid_dims_ = dims; }
dim3 GetLaunchGridDims() const override
{
return launch_grid_dims_;
}
bool HasLaunchGridDims() const override
{
return launch_grid_dims_.x > 0;
}
dim3 GetLaunchGridDims() const override { return launch_grid_dims_; }
bool HasLaunchGridDims() const override { return launch_grid_dims_.x > 0; }
};
struct SplitKBatchOffset