[flang-commits] [flang] [flang][cuda] Add entry point to launch global function with cluster_dims (PR #113958)
Valentin Clement バレンタイン クレメン via flang-commits
flang-commits at lists.llvm.org
Mon Oct 28 12:59:57 PDT 2024
https://github.com/clementval created https://github.com/llvm/llvm-project/pull/113958
None
>From 8e02c902fcb522510d0e8c860a191be238c3e657 Mon Sep 17 00:00:00 2001
From: Valentin Clement <clementval at gmail.com>
Date: Fri, 25 Oct 2024 11:16:14 -0700
Subject: [PATCH] [flang][cuda] Add entry point to launch global function with
cluster_dims
---
flang/include/flang/Runtime/CUDA/kernel.h | 8 +++++++-
flang/runtime/CUDA/kernel.cpp | 25 ++++++++++++++++++++++-
2 files changed, 31 insertions(+), 2 deletions(-)
diff --git a/flang/include/flang/Runtime/CUDA/kernel.h b/flang/include/flang/Runtime/CUDA/kernel.h
index cf07d874a082c0..85afda09e347ae 100644
--- a/flang/include/flang/Runtime/CUDA/kernel.h
+++ b/flang/include/flang/Runtime/CUDA/kernel.h
@@ -15,13 +15,19 @@
extern "C" {
-// This function uses intptr_t instead of CUDA's unsigned int to match
+// These functions use intptr_t instead of CUDA's unsigned int to match
// the type of MLIR's index type. This avoids the need for casts in the
// generated MLIR code.
+
void RTDEF(CUFLaunchKernel)(const void *kernelName, intptr_t gridX,
intptr_t gridY, intptr_t gridZ, intptr_t blockX, intptr_t blockY,
intptr_t blockZ, int32_t smem, void **params, void **extra);
+void RTDEF(CUFLaunchClusterKernel)(const void *kernelName, intptr_t clusterX,
+ intptr_t clusterY, intptr_t clusterZ, intptr_t gridX, intptr_t gridY,
+ intptr_t gridZ, intptr_t blockX, intptr_t blockY, intptr_t blockZ,
+ int32_t smem, void **params, void **extra);
+
} // extern "C"
#endif // FORTRAN_RUNTIME_CUDA_KERNEL_H_
diff --git a/flang/runtime/CUDA/kernel.cpp b/flang/runtime/CUDA/kernel.cpp
index f81153a1af4bc7..abb7ebb72e5923 100644
--- a/flang/runtime/CUDA/kernel.cpp
+++ b/flang/runtime/CUDA/kernel.cpp
@@ -25,9 +25,32 @@ void RTDEF(CUFLaunchKernel)(const void *kernel, intptr_t gridX, intptr_t gridY,
blockDim.x = blockX;
blockDim.y = blockY;
blockDim.z = blockZ;
- cudaStream_t stream = 0;
+ cudaStream_t stream = 0; // TODO stream managment
CUDA_REPORT_IF_ERROR(
cudaLaunchKernel(kernel, gridDim, blockDim, params, smem, stream));
}
+void RTDEF(CUFLaunchClusterKernel)(const void *kernel, intptr_t clusterX,
+ intptr_t clusterY, intptr_t clusterZ, intptr_t gridX, intptr_t gridY,
+ intptr_t gridZ, intptr_t blockX, intptr_t blockY, intptr_t blockZ,
+ int32_t smem, void **params, void **extra) {
+ cudaLaunchConfig_t config;
+ config.gridDim.x = gridX;
+ config.gridDim.y = gridY;
+ config.gridDim.z = gridZ;
+ config.blockDim.x = blockX;
+ config.blockDim.y = blockY;
+ config.blockDim.z = blockZ;
+ config.dynamicSmemBytes = smem;
+ config.stream = 0; // TODO stream managment
+ cudaLaunchAttribute launchAttr[1];
+ launchAttr[0].id = cudaLaunchAttributeClusterDimension;
+ launchAttr[0].val.clusterDim.x = clusterX;
+ launchAttr[0].val.clusterDim.y = clusterY;
+ launchAttr[0].val.clusterDim.z = clusterZ;
+ config.numAttrs = 1;
+ config.attrs = launchAttr;
+ CUDA_REPORT_IF_ERROR(cudaLaunchKernelExC(&config, kernel, params));
+}
+
} // extern "C"
More information about the flang-commits
mailing list