[Mlir-commits] [mlir] Reland #171543 [mlir][amdgpu] Add common gpu mem space conversions to convert-amdgpu-to-rocdl (PR #171599)

[llvmlistbot at llvm.org]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir

Author: Ivan Butygin (Hardcode84)

<details>
<summary>Changes</summary>

Added missing GPU lib `MLIRGPUToGPURuntimeTransforms`.

---
Full diff: https://github.com/llvm/llvm-project/pull/171599.diff


3 Files Affected:

- (modified) mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp (+13) 
- (modified) mlir/lib/Conversion/AMDGPUToROCDL/CMakeLists.txt (+4-3) 
- (modified) mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir (+40-2) 


``````````diff
diff --git a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
index 7584b17075225..e3dacfff16aa2 100644
--- a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
+++ b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
@@ -8,6 +8,7 @@
 
 #include "mlir/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.h"
 
+#include "mlir/Conversion/GPUCommon/GPUCommonPass.h"
 #include "mlir/Conversion/LLVMCommon/ConversionTarget.h"
 #include "mlir/Conversion/LLVMCommon/Pattern.h"
 #include "mlir/Conversion/LLVMCommon/TypeConverter.h"
@@ -2730,6 +2731,18 @@ struct ConvertAMDGPUToROCDLPass
     });
 
     populateAMDGPUToROCDLConversionPatterns(converter, patterns, *maybeChipset);
+    populateGpuMemorySpaceAttributeConversions(
+        converter, [](gpu::AddressSpace space) {
+          switch (space) {
+          case gpu::AddressSpace::Global:
+            return 1;
+          case gpu::AddressSpace::Workgroup:
+            return 3;
+          case gpu::AddressSpace::Private:
+            return 5;
+          }
+          llvm_unreachable("unknown address space enum value");
+        });
     LLVMConversionTarget target(getContext());
     target.addIllegalDialect<::mlir::amdgpu::AMDGPUDialect>();
     target.addLegalDialect<::mlir::LLVM::LLVMDialect>();
diff --git a/mlir/lib/Conversion/AMDGPUToROCDL/CMakeLists.txt b/mlir/lib/Conversion/AMDGPUToROCDL/CMakeLists.txt
index 43cbc0732516d..25cceb9d1d37f 100644
--- a/mlir/lib/Conversion/AMDGPUToROCDL/CMakeLists.txt
+++ b/mlir/lib/Conversion/AMDGPUToROCDL/CMakeLists.txt
@@ -11,11 +11,12 @@ add_mlir_conversion_library(MLIRAMDGPUToROCDL
   Core
 
   LINK_LIBS PUBLIC
-  MLIRLLVMCommonConversion
-  MLIRLLVMDialect
-  MLIRROCDLDialect
   MLIRAMDGPUDialect
   MLIRAMDGPUUtils
+  MLIRGPUToGPURuntimeTransforms
+  MLIRLLVMCommonConversion
+  MLIRLLVMDialect
   MLIRPass
+  MLIRROCDLDialect
   MLIRTransforms
   )
diff --git a/mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir b/mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir
index e48c94195ea56..30578517be1ca 100644
--- a/mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir
+++ b/mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir
@@ -1,5 +1,5 @@
-// RUN: mlir-opt %s -convert-amdgpu-to-rocdl=chipset=gfx942 | FileCheck %s
-// RUN: mlir-opt %s -convert-amdgpu-to-rocdl=chipset=gfx950 | FileCheck %s
+// RUN: mlir-opt %s --convert-amdgpu-to-rocdl=chipset=gfx942 | FileCheck %s
+// RUN: mlir-opt %s --convert-amdgpu-to-rocdl=chipset=gfx950 | FileCheck %s
 
 #gpu_global_addrspace = 1
 #gpu_lds_addrspace = 3
@@ -43,6 +43,44 @@ func.func @global_load_to_rocdl_f32(%global : memref<128x72xf32, #gpu_global_add
   func.return
 }
 
+// CHECK-LABEL: func @global_load_to_rocdl_wg_mem
+// CHECK-SAME: (%[[ARG0:.*]]: memref<128x72xf32>)
+func.func @global_load_to_rocdl_wg_mem(%global : memref<128x72xf32>) {
+  %c0 = arith.constant 0 : index
+  %c12 = arith.constant 12 : index
+  %c32 = arith.constant 32 : index
+  %alloc = memref.alloc() : memref<64x64xf32, #gpu.address_space<workgroup>>
+  // CHECK: %[[GLOBAL_DESC:.*]] = builtin.unrealized_conversion_cast %[[ARG0]]
+
+  // CHECK: %[[C0:.*]] = arith.constant 0 : index
+  // CHECK: %[[IC0:.*]] = builtin.unrealized_conversion_cast %c0 : index to i64
+  // CHECK: %[[C12:.*]] = arith.constant 12 : index
+  // CHECK: %[[IC12:.*]] = builtin.unrealized_conversion_cast %[[C12]]
+  // CHECK: %[[C32:.*]] = arith.constant 32 : index
+  // CHECK: %[[IC32:.*]] = builtin.unrealized_conversion_cast %[[C32]]
+
+  // CHECK: %[[ALLOC:.*]] = memref.alloc()
+  // CHECK: %[[LDS_DESC:.*]] = builtin.unrealized_conversion_cast
+  // CHECK: %[[GLOBAL_BASE:.*]] = llvm.extractvalue %[[GLOBAL_DESC]][1]
+
+  // CHECK: %[[C72:.*]] = llvm.mlir.constant(72 : index) : i64
+  // CHECK: %[[MUL:.*]] = llvm.mul %[[IC12]], %[[C72]] : i64
+  // CHECK: %[[SRC_OFFSET:.*]] = llvm.add %[[MUL]], %[[IC0]] : i64
+
+  // CHECK: %[[GLOBAL_PTR:.*]] = llvm.getelementptr %[[GLOBAL_BASE]][%[[SRC_OFFSET]]]
+  // CHECK: %[[LDS_BASE:.*]] = llvm.extractvalue %[[LDS_DESC]][1]
+
+  // CHECK: %[[C64:.*]] = llvm.mlir.constant(64 : index) : i64
+  // CHECK: %[[MUL_2:.*]] = llvm.mul %[[IC32]], %[[C64]] : i64
+  // CHECK: %[[DST_OFFSET:.*]] = llvm.add %[[MUL_2]], %[[IC0]] : i64
+
+  // CHECK: %[[LDS_PTR:.*]] = llvm.getelementptr %[[LDS_BASE]][%[[DST_OFFSET]]]
+  // CHECK: rocdl.load.to.lds %[[GLOBAL_PTR]], %[[LDS_PTR]], 4
+  amdgpu.gather_to_lds %global[%c12, %c0], %alloc[%c32, %c0]
+    : f32, memref<128x72xf32>, memref<64x64xf32, #gpu.address_space<workgroup>>
+  func.return
+}
+
 // CHECK-LABEL: func @global_load_to_rocdl_i8
 // CHECK-SAME: (%[[ARG0:.*]]: memref<128x72xi8, 1>)
 func.func @global_load_to_rocdl_i8(%global : memref<128x72xi8, #gpu_global_addrspace>) {

``````````

</details>


https://github.com/llvm/llvm-project/pull/171599