[Mlir-commits] [mlir] 668c964 - [AMDGPU] [MLIR] Add 96 and 128 bit GatherToLDS for gfx950 (#147496)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Wed Jul 9 08:53:29 PDT 2025


Author: Daniel Hernandez-Juarez
Date: 2025-07-09T11:53:26-04:00
New Revision: 668c964282f0850a07976f46182efe745f77b789

URL: https://github.com/llvm/llvm-project/commit/668c964282f0850a07976f46182efe745f77b789
DIFF: https://github.com/llvm/llvm-project/commit/668c964282f0850a07976f46182efe745f77b789.diff

LOG: [AMDGPU] [MLIR] Add 96 and 128 bit GatherToLDS for gfx950 (#147496)

This PR adds 96 and 128 gather_to_lds support for gfx950. Updating
lowering, verifier and tests.

Added: 
    mlir/test/Conversion/AMDGPUToROCDL/load_lds-gfx950.mlir

Modified: 
    mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
    mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp
    mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir
    mlir/test/Conversion/AMDGPUToROCDL/transpose_load.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
index 910fe1b1d93c1..ef35ee208f002 100644
--- a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
+++ b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
@@ -1196,18 +1196,23 @@ struct GatherToLDSOpLowering : public ConvertOpToLLVMPattern<GatherToLDSOp> {
     // augment it to transfer multiple elements per thread by issuing multiple
     // `global_load_lds` instructions.
     Type transferType = op.getTransferType();
-    size_t loadWidth = [&]() -> size_t {
+    int loadWidth = [&]() -> int {
       if (auto transferVectorType = dyn_cast<VectorType>(transferType)) {
-        return transferVectorType.getNumElements() *
-               (transferVectorType.getElementTypeBitWidth() / 8);
+        return (transferVectorType.getNumElements() *
+                transferVectorType.getElementTypeBitWidth()) /
+               8;
       }
       return transferType.getIntOrFloatBitWidth() / 8;
     }();
 
-    // Currently only 1, 2, and 4 byte loads are supported.
-    if (loadWidth != 1 && loadWidth != 2 && loadWidth != 4)
+    // Currently only 1, 2, 4, 12 and 16 byte loads are supported.
+    if (!llvm::is_contained({1, 2, 4, 12, 16}, loadWidth))
       return op.emitOpError("chipset unsupported element size");
 
+    if (chipset != kGfx950 && llvm::is_contained({12, 16}, loadWidth))
+      return op.emitOpError("Gather to LDS instructions with 12-byte and "
+                            "16-byte load widths are only supported on gfx950");
+
     Value srcPtr =
         getStridedElementPtr(rewriter, loc, srcMemRefType, adaptor.getSrc(),
                              (adaptor.getSrcIndices()));

diff  --git a/mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp b/mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp
index 4613d14461969..acaf6a2f8792a 100644
--- a/mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp
+++ b/mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp
@@ -502,17 +502,18 @@ LogicalResult GatherToLDSOp::verify() {
   if (elemType != dstType.getElementType())
     return emitOpError("source and destination element types must match");
 
-  // copy type sizes should be 1, 2, or 4 bytes.
+  // copy type sizes should be 1, 2, 4, 12 or 16 bytes.
   auto transferType = getTransferType();
-  size_t transferSize;
+  int transferSize;
   if (auto vectorTransfer = dyn_cast<VectorType>(transferType)) {
     transferSize = vectorTransfer.getNumElements() *
                    vectorTransfer.getElementTypeBitWidth();
   } else {
     transferSize = transferType.getIntOrFloatBitWidth();
   }
-  if (transferSize != 8 && transferSize != 16 && transferSize != 32)
-    return emitOpError("Transfering type size must be 8, 16, or 32 bits");
+  if (!llvm::is_contained({8, 16, 32, 96, 128}, transferSize))
+    return emitOpError(
+        "Transfering type size must be 8, 16, 32, 96 or 128 bits");
 
   if (!hasGlobalMemorySpace(srcType.getMemorySpace()) &&
       !hasFatRawBufferMemorySpace(srcType.getMemorySpace()))

diff  --git a/mlir/test/Conversion/AMDGPUToROCDL/load_lds-gfx950.mlir b/mlir/test/Conversion/AMDGPUToROCDL/load_lds-gfx950.mlir
new file mode 100644
index 0000000000000..42fb18006bea4
--- /dev/null
+++ b/mlir/test/Conversion/AMDGPUToROCDL/load_lds-gfx950.mlir
@@ -0,0 +1,90 @@
+// RUN: not mlir-opt %s --split-input-file -convert-amdgpu-to-rocdl=chipset=gfx942 2>&1 | FileCheck %s --check-prefix=GFX942
+// RUN: mlir-opt %s --split-input-file -convert-amdgpu-to-rocdl=chipset=gfx950 | FileCheck %s --check-prefix=GFX950
+
+#gpu_global_addrspace = 1
+#gpu_lds_addrspace = 3
+#amdgpu_fat_buffer_addrspace = 7
+
+// GFX950-LABEL: func @fat_buffer_load_to_rocdl_f96
+// GFX950-SAME: (%[[ARG0:.*]]: memref<128x72xf32, 7>)
+func.func @fat_buffer_load_to_rocdl_f96(%global : memref<128x72xf32, #amdgpu_fat_buffer_addrspace>) {
+  %c0 = arith.constant 0 : index
+  %c12 = arith.constant 12 : index
+  %c32 = arith.constant 32 : index
+  %alloc = memref.alloc() : memref<64x64xf32, #gpu_lds_addrspace>
+  // GFX950: %[[BUFFER_DESC:.*]] = builtin.unrealized_conversion_cast %[[ARG0]]
+
+  // GFX950: %[[C0:.*]] = arith.constant 0 : index
+  // GFX950: %[[IC0:.*]] = builtin.unrealized_conversion_cast %c0 : index to i64
+  // GFX950: %[[C12:.*]] = arith.constant 12 : index
+  // GFX950: %[[IC12:.*]] = builtin.unrealized_conversion_cast %[[C12]]
+  // GFX950: %[[C32:.*]] = arith.constant 32 : index
+  // GFX950: %[[IC32:.*]] = builtin.unrealized_conversion_cast %[[C32]]
+
+  // GFX950: %[[ALLOC:.*]] = memref.alloc()
+  // GFX950: %[[LDS_DESC:.*]] = builtin.unrealized_conversion_cast
+  // GFX950: %[[GLOBAL_BASE:.*]] = llvm.extractvalue %[[BUFFER_DESC]][1]
+
+  // GFX950: %[[C72:.*]] = llvm.mlir.constant(72 : index) : i64
+  // GFX950: %[[MUL:.*]] = llvm.mul %[[IC12]], %[[C72]] : i64
+  // GFX950: %[[SRC_OFFSET:.*]] = llvm.add %[[MUL]], %[[IC0]] : i64
+
+  // GFX950: %[[GLOBAL_PTR:.*]] = llvm.getelementptr %[[GLOBAL_BASE]][%[[SRC_OFFSET]]]
+  // GFX950: %[[LDS_BASE:.*]] = llvm.extractvalue %[[LDS_DESC]][1]
+
+  // GFX950: %[[C64:.*]] = llvm.mlir.constant(64 : index) : i64
+  // GFX950: %[[MUL_2:.*]] = llvm.mul %[[IC32]], %[[C64]] : i64
+  // GFX950: %[[DST_OFFSET:.*]] = llvm.add %[[MUL_2]], %[[IC0]] : i64
+
+  // GFX950: %[[LDS_PTR:.*]] = llvm.getelementptr %[[LDS_BASE]][%[[DST_OFFSET]]]
+  // GFX950: rocdl.load.to.lds %[[GLOBAL_PTR]], %[[LDS_PTR]], 12
+  // GFX942: error: 'amdgpu.gather_to_lds' op Gather to LDS instructions with 12-byte and 16-byte load widths are only supported on gfx950
+  amdgpu.gather_to_lds %global[%c12, %c0], %alloc[%c32, %c0]
+    : vector<16xf6E3M2FN>, memref<128x72xf32, #amdgpu_fat_buffer_addrspace>, memref<64x64xf32, #gpu_lds_addrspace>
+  func.return
+}
+
+// -----
+
+#gpu_global_addrspace = 1
+#gpu_lds_addrspace = 3
+#amdgpu_fat_buffer_addrspace = 7
+
+// GFX950-LABEL: func @fat_buffer_load_to_rocdl_f128
+// GFX950-SAME: (%[[ARG0:.*]]: memref<128x72xf32, 7>)
+func.func @fat_buffer_load_to_rocdl_f128(%global : memref<128x72xf32, #amdgpu_fat_buffer_addrspace>) {
+  %c0 = arith.constant 0 : index
+  %c12 = arith.constant 12 : index
+  %c32 = arith.constant 32 : index
+  %alloc = memref.alloc() : memref<64x64xf32, #gpu_lds_addrspace>
+  // GFX950: %[[BUFFER_DESC:.*]] = builtin.unrealized_conversion_cast %[[ARG0]]
+
+  // GFX950: %[[C0:.*]] = arith.constant 0 : index
+  // GFX950: %[[IC0:.*]] = builtin.unrealized_conversion_cast %c0 : index to i64
+  // GFX950: %[[C12:.*]] = arith.constant 12 : index
+  // GFX950: %[[IC12:.*]] = builtin.unrealized_conversion_cast %[[C12]]
+  // GFX950: %[[C32:.*]] = arith.constant 32 : index
+  // GFX950: %[[IC32:.*]] = builtin.unrealized_conversion_cast %[[C32]]
+
+  // GFX950: %[[ALLOC:.*]] = memref.alloc()
+  // GFX950: %[[LDS_DESC:.*]] = builtin.unrealized_conversion_cast
+  // GFX950: %[[GLOBAL_BASE:.*]] = llvm.extractvalue %[[BUFFER_DESC]][1]
+
+  // GFX950: %[[C72:.*]] = llvm.mlir.constant(72 : index) : i64
+  // GFX950: %[[MUL:.*]] = llvm.mul %[[IC12]], %[[C72]] : i64
+  // GFX950: %[[SRC_OFFSET:.*]] = llvm.add %[[MUL]], %[[IC0]] : i64
+
+  // GFX950: %[[GLOBAL_PTR:.*]] = llvm.getelementptr %[[GLOBAL_BASE]][%[[SRC_OFFSET]]]
+  // GFX950: %[[LDS_BASE:.*]] = llvm.extractvalue %[[LDS_DESC]][1]
+
+  // GFX950: %[[C64:.*]] = llvm.mlir.constant(64 : index) : i64
+  // GFX950: %[[MUL_2:.*]] = llvm.mul %[[IC32]], %[[C64]] : i64
+  // GFX950: %[[DST_OFFSET:.*]] = llvm.add %[[MUL_2]], %[[IC0]] : i64
+
+  // GFX950: %[[LDS_PTR:.*]] = llvm.getelementptr %[[LDS_BASE]][%[[DST_OFFSET]]]
+  // GFX950: rocdl.load.to.lds %[[GLOBAL_PTR]], %[[LDS_PTR]], 16
+  // GFX942: error: 'amdgpu.gather_to_lds' op Gather to LDS instructions with 12-byte and 16-byte load widths are only supported on gfx950
+  amdgpu.gather_to_lds %global[%c12, %c0], %alloc[%c32, %c0]
+    : f128, memref<128x72xf32, #amdgpu_fat_buffer_addrspace>, memref<64x64xf32, #gpu_lds_addrspace>
+  func.return
+}

diff  --git a/mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir b/mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir
index 581346e03b893..77103fa5c25f1 100644
--- a/mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir
+++ b/mlir/test/Conversion/AMDGPUToROCDL/load_lds.mlir
@@ -1,4 +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
 
 #gpu_global_addrspace = 1
 #gpu_lds_addrspace = 3
@@ -118,7 +119,6 @@ func.func @global_load_to_rocdl_vec(%global : memref<128x72xi16, #gpu_global_add
   func.return
 }
 
-
 // CHECK-LABEL: func @global_load_to_rocdl_dynamic_indices
 // CHECK-SAME: (%[[ARG0:.*]]: memref<512xi32, 1>, %[[SRC_IDX:.*]]: index, %[[DST_IDX:.*]]: index)
 func.func @global_load_to_rocdl_dynamic_indices(%global : memref<512xi32, #gpu_global_addrspace>, %src_idx : index, %dst_idx : index) {

diff  --git a/mlir/test/Conversion/AMDGPUToROCDL/transpose_load.mlir b/mlir/test/Conversion/AMDGPUToROCDL/transpose_load.mlir
index 68799098f1d36..73481e4bced20 100644
--- a/mlir/test/Conversion/AMDGPUToROCDL/transpose_load.mlir
+++ b/mlir/test/Conversion/AMDGPUToROCDL/transpose_load.mlir
@@ -1,5 +1,5 @@
 // RUN: mlir-opt %s --split-input-file -convert-amdgpu-to-rocdl=chipset=gfx950 | FileCheck %s
-// RUN: not mlir-opt %s --split-input-file -convert-amdgpu-to-rocdl=chipset=gfx945 2>&1 | FileCheck %s --check-prefix=CHECK-OLD 
+// RUN: not mlir-opt %s --split-input-file -convert-amdgpu-to-rocdl=chipset=gfx942 2>&1 | FileCheck %s --check-prefix=CHECK-OLD 
 
 // CHECK-LABEL: func @transpose_load_to_rocdl_4xf16
 func.func @transpose_load_to_rocdl_4xf16(%idx1 : index, %idx2 : index, %wgmem : memref<128x72xf16, 3>) -> vector<4xf16> {


        


More information about the Mlir-commits mailing list