[Mlir-commits] [mlir] [MLIR][GPU] Add gpu.cluster_dim_blocks Op to represent number of blocks per cluster (PR #95245)

[Pradeep Kumar]

https://github.com/schwarzschild-radius updated https://github.com/llvm/llvm-project/pull/95245

>From 4430b67bb571c379c1ec1eade4050e859061e896 Mon Sep 17 00:00:00 2001
From: pradeepku <pradeepku at nvidia.com>
Date: Wed, 12 Jun 2024 12:59:02 +0000
Subject: [PATCH] [MLIR][GPU] Add gpu.cluster_dim_blocks Op to represent number
 of blocks per cluster

This commit adds support for `gpu.cluster_dim_blocks` Op to represent number of blocks per cluster
and updated the description of `gpu.cluster_dim` Op. Also, updated the `cga_cluster.mlir`
test file to use `gpu.cluster_dim_blocks`
---
 mlir/include/mlir/Dialect/GPU/IR/GPUOps.td        | 15 ++++++++++++++-
 mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td       |  6 +++---
 .../Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp |  3 +++
 .../GPU/IR/InferIntRangeInterfaceImpls.cpp        |  5 +++++
 .../Integration/GPU/CUDA/sm90/cga_cluster.mlir    |  6 +++---
 5 files changed, 28 insertions(+), 7 deletions(-)

diff --git a/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td b/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
index eb81b6469746f..e7e55ab42d51f 100644
--- a/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
+++ b/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
@@ -70,7 +70,7 @@ class GPU_IndexOp<string mnemonic, list<Trait> traits = []> :
 
 def GPU_ClusterDimOp : GPU_IndexOp<"cluster_dim"> {
   let description = [{
-    Returns the number of thread blocks in the cluster along
+    Returns the number of cluster identifiers per grid along
     the x, y, or z `dimension`.
 
     Example:
@@ -81,6 +81,19 @@ def GPU_ClusterDimOp : GPU_IndexOp<"cluster_dim"> {
   }];
 }
 
+def GPU_ClusterDimBlocksOp : GPU_IndexOp<"cluster_dim_blocks"> {
+  let description = [{
+    Returns the number of thread blocks in the cluster along
+    the x, y, or z `dimension`.
+
+    Example:
+
+    ```mlir
+    %cDimBlocksX = gpu.cluster_dim_blocks x
+    ```
+  }];
+}
+
 def GPU_ClusterIdOp : GPU_IndexOp<"cluster_id"> {
   let description = [{
     Returns the cluster id, i.e. the index of the current cluster within the
diff --git a/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td b/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
index 4daeeab093863..4d48b3de7a57e 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td
@@ -160,9 +160,9 @@ def NVVM_ClusterDimZOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.nclusterid.z">;
 def NVVM_BlockInClusterIdXOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.ctaid.x">;
 def NVVM_BlockInClusterIdYOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.ctaid.y">;
 def NVVM_BlockInClusterIdZOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.ctaid.z">;
-def NVVM_GridInClusterDimXOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.nctaid.x">;
-def NVVM_GridInClusterDimYOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.nctaid.y">;
-def NVVM_GridInClusterDimZOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.nctaid.z">;
+def NVVM_ClusterDimBlocksXOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.nctaid.x">;
+def NVVM_ClusterDimBlocksYOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.nctaid.y">;
+def NVVM_ClusterDimBlocksZOp : NVVM_SpecialRegisterOp<"read.ptx.sreg.cluster.nctaid.z">;
 
 //===----------------------------------------------------------------------===//
 // CTA index and across Cluster dimensions
diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index b95fba20a00cb..e3d7da782e689 100644
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -344,6 +344,9 @@ void mlir::populateGpuToNVVMConversionPatterns(LLVMTypeConverter &converter,
                                   NVVM::ClusterIdYOp, NVVM::ClusterIdZOp>,
       GPUIndexIntrinsicOpLowering<gpu::ClusterDimOp, NVVM::ClusterDimXOp,
                                   NVVM::ClusterDimYOp, NVVM::ClusterDimZOp>,
+      GPUIndexIntrinsicOpLowering<
+          gpu::ClusterDimBlocksOp, NVVM::ClusterDimBlocksXOp,
+          NVVM::ClusterDimBlocksYOp, NVVM::ClusterDimBlocksZOp>,
       GPUIndexIntrinsicOpLowering<gpu::BlockIdOp, NVVM::BlockIdXOp,
                                   NVVM::BlockIdYOp, NVVM::BlockIdZOp>,
       GPUIndexIntrinsicOpLowering<gpu::GridDimOp, NVVM::GridDimXOp,
diff --git a/mlir/lib/Dialect/GPU/IR/InferIntRangeInterfaceImpls.cpp b/mlir/lib/Dialect/GPU/IR/InferIntRangeInterfaceImpls.cpp
index 69017efb9a0e6..12e2588cb05b4 100644
--- a/mlir/lib/Dialect/GPU/IR/InferIntRangeInterfaceImpls.cpp
+++ b/mlir/lib/Dialect/GPU/IR/InferIntRangeInterfaceImpls.cpp
@@ -86,6 +86,11 @@ static std::optional<uint64_t> getKnownLaunchDim(Op op, LaunchDims type) {
 
 void ClusterDimOp::inferResultRanges(ArrayRef<ConstantIntRanges>,
                                      SetIntRangeFn setResultRange) {
+  setResultRange(getResult(), getIndexRange(1, kMaxDim));
+}
+
+void ClusterDimBlocksOp::inferResultRanges(ArrayRef<ConstantIntRanges>,
+                                           SetIntRangeFn setResultRange) {
   setResultRange(getResult(), getIndexRange(1, kMaxClusterDim));
 }
 
diff --git a/mlir/test/Integration/GPU/CUDA/sm90/cga_cluster.mlir b/mlir/test/Integration/GPU/CUDA/sm90/cga_cluster.mlir
index 025282ec0d688..5c11d80178f72 100644
--- a/mlir/test/Integration/GPU/CUDA/sm90/cga_cluster.mlir
+++ b/mlir/test/Integration/GPU/CUDA/sm90/cga_cluster.mlir
@@ -22,9 +22,9 @@ module attributes {gpu.container_module} {
       %cidX = gpu.cluster_id  x
       %cidY = gpu.cluster_id  y
       %cidZ = gpu.cluster_id  z
-      %cdimX = gpu.cluster_dim  x
-      %cdimY = gpu.cluster_dim  y
-      %cdimZ = gpu.cluster_dim  z
+      %cdimX = gpu.cluster_dim_blocks  x
+      %cdimY = gpu.cluster_dim_blocks  y
+      %cdimZ = gpu.cluster_dim_blocks  z
       %bidX = gpu.block_id  x
       %bidY = gpu.block_id  y
       %bidZ = gpu.block_id  z