[Mlir-commits] [mlir] [mlir] Add extra value types for gpu::ShuffleOp (PR #104605)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Fri Aug 16 08:28:08 PDT 2024


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir-gpu

Author: Finlay (FMarno)

<details>
<summary>Changes</summary>

Adds the i8, i16, and f16 types.
Also update the gpu-to-llvm-spv pass to support that.
I looked at the other gpu-to-* passes and it looks like they should already support these types, or will return failure.
* gpu-to-vulkan: currently doesn't lower the shuffle op
* gpu-to-rocdl: only supports 32-bit anyway - https://github.com/llvm/llvm-project/blob/99696b35bc8a0054e0b0c1a26e8dd5049fa8c41b/mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp#L143-L144
* gpu-to-spirv: directly passes the value on and supports "8/16/32/64-bit integer or 16/32/64-bit float or bool or vector of bool or 8/16/32/64-bit integer or 16/32/64-bit float values of length 2/3/4/8/16"
* gpu-to-nvvm: directly passes the value on and supports "LLVM dialect-compatible type(s)"

---
Full diff: https://github.com/llvm/llvm-project/pull/104605.diff


4 Files Affected:

- (modified) mlir/include/mlir/Dialect/GPU/IR/GPUOps.td (+7-7) 
- (modified) mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp (+5) 
- (modified) mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir (+17-16) 
- (modified) mlir/test/Dialect/GPU/invalid.mlir (+1-1) 


``````````diff
diff --git a/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td b/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
index a024c3018eb8d3..47c0a8039d942a 100644
--- a/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
+++ b/mlir/include/mlir/Dialect/GPU/IR/GPUOps.td
@@ -1250,17 +1250,17 @@ def GPU_ShuffleMode : I32EnumAttr<"ShuffleMode",
 def GPU_ShuffleModeAttr : EnumAttr<GPU_Dialect, GPU_ShuffleMode,
                                    "shuffle_mode">;
 
-def I32I64F32OrF64 : TypeConstraint<Or<[I32.predicate,
-                                        I64.predicate,
-                                        F32.predicate,
-                                        F64.predicate]>,
-                                       "i32, i64, f32 or f64">;
+def GPUShuffleTypes : TypeConstraint<Or<[
+  SignlessIntOfWidths<[8, 16, 32, 64]>.predicate,
+  FloatOfWidths<[16, 32, 64]>.predicate
+  ]>,
+  "i8, i16, i32, i64, f16, f32 or f64">;
 
 def GPU_ShuffleOp : GPU_Op<
     "shuffle", [Pure, AllTypesMatch<["value", "shuffleResult"]>]>,
-    Arguments<(ins I32I64F32OrF64:$value, I32:$offset, I32:$width,
+    Arguments<(ins GPUShuffleTypes:$value, I32:$offset, I32:$width,
                GPU_ShuffleModeAttr:$mode)>,
-    Results<(outs I32I64F32OrF64:$shuffleResult, I1:$valid)> {
+    Results<(outs GPUShuffleTypes:$shuffleResult, I1:$valid)> {
   let summary = "Shuffles values within a subgroup.";
   let description = [{
     The "shuffle" op moves values to a across lanes (a.k.a., invocations,
diff --git a/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp b/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
index ced4236402923a..62776cae834eb8 100644
--- a/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
+++ b/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
@@ -243,10 +243,15 @@ struct GPUShuffleConversion final : ConvertOpToLLVMPattern<gpu::ShuffleOp> {
 
   static StringRef getTypeMangling(Type type) {
     return TypeSwitch<Type, StringRef>(type)
+        .Case<Float16Type>([](auto) { return "Dhj"; })
         .Case<Float32Type>([](auto) { return "fj"; })
         .Case<Float64Type>([](auto) { return "dj"; })
         .Case<IntegerType>([](auto intTy) {
           switch (intTy.getWidth()) {
+          case 8:
+            return "cj";
+          case 16:
+            return "sj";
           case 32:
             return "ij";
           case 64:
diff --git a/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir b/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
index ec4f4a304d5073..eb0da3a95ba4b7 100644
--- a/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
+++ b/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
@@ -311,19 +311,19 @@ gpu.module @shuffles attributes {
   // CHECK-SAME-DAG:  will_return
   // CHECK-NOT:       memory_effects = #llvm.memory_effects
   // CHECK-SAME:      }
-  // CHECK:           llvm.func spir_funccc @_Z20sub_group_shuffle_upfj(f32, i32) -> f32 attributes {
+  // CHECK:           llvm.func spir_funccc @_Z20sub_group_shuffle_upDhj(f16, i32) -> f16 attributes {
   // CHECK-SAME-DAG:  no_unwind
   // CHECK-SAME-DAG:  convergent
   // CHECK-SAME-DAG:  will_return
   // CHECK-NOT:       memory_effects = #llvm.memory_effects
   // CHECK-SAME:      }
-  // CHECK:           llvm.func spir_funccc @_Z21sub_group_shuffle_xorlj(i64, i32) -> i64 attributes {
+  // CHECK:           llvm.func spir_funccc @_Z21sub_group_shuffle_xorsj(i16, i32) -> i16 attributes {
   // CHECK-SAME-DAG:  no_unwind
   // CHECK-SAME-DAG:  convergent
   // CHECK-SAME-DAG:  will_return
   // CHECK-NOT:       memory_effects = #llvm.memory_effects
   // CHECK-SAME:      }
-  // CHECK:           llvm.func spir_funccc @_Z17sub_group_shuffleij(i32, i32) -> i32 attributes {
+  // CHECK:           llvm.func spir_funccc @_Z17sub_group_shufflecj(i8, i32) -> i8 attributes {
   // CHECK-SAME-DAG:  no_unwind
   // CHECK-SAME-DAG:  convergent
   // CHECK-SAME-DAG:  will_return
@@ -331,24 +331,25 @@ gpu.module @shuffles attributes {
   // CHECK-SAME:      }
 
   // CHECK-LABEL: gpu_shuffles
-  // CHECK-SAME:              (%[[VAL_0:.*]]: i32, %[[VAL_1:.*]]: i32, %[[VAL_2:.*]]: i64, %[[VAL_3:.*]]: i32, %[[VAL_4:.*]]: f32, %[[VAL_5:.*]]: i32, %[[VAL_6:.*]]: f64, %[[VAL_7:.*]]: i32)
-  func.func @gpu_shuffles(%val0: i32, %id: i32,
-                          %val1: i64, %mask: i32,
-                          %val2: f32, %delta_up: i32,
-                          %val3: f64, %delta_down: i32) {
+  // CHECK-SAME:              (%[[I8_VAL:.*]]: i8, %[[I16_VAL:.*]]: i16, %[[F16_VAL:.*]]: f16, %[[F64_VAL:.*]]: f64,  %[[OFFSET:.*]]: i32)
+  func.func @gpu_shuffles(%i8_val: i8,
+                          %i16_val: i16,
+                          %f16_val: f16,
+                          %f64_val: f64,
+                          %offset: i32) {
     %width = arith.constant 16 : i32
-    // CHECK:         llvm.call spir_funccc @_Z17sub_group_shuffleij(%[[VAL_0]], %[[VAL_1]])
+    // CHECK:         llvm.call spir_funccc @_Z17sub_group_shufflecj(%[[I8_VAL]], %[[OFFSET]])
     // CHECK:         llvm.mlir.constant(true) : i1
-    // CHECK:         llvm.call spir_funccc @_Z21sub_group_shuffle_xorlj(%[[VAL_2]], %[[VAL_3]])
+    // CHECK:         llvm.call spir_funccc @_Z21sub_group_shuffle_xorsj(%[[I16_VAL]], %[[OFFSET]])
     // CHECK:         llvm.mlir.constant(true) : i1
-    // CHECK:         llvm.call spir_funccc @_Z20sub_group_shuffle_upfj(%[[VAL_4]], %[[VAL_5]])
+    // CHECK:         llvm.call spir_funccc @_Z20sub_group_shuffle_upDhj(%[[F16_VAL]], %[[OFFSET]])
     // CHECK:         llvm.mlir.constant(true) : i1
-    // CHECK:         llvm.call spir_funccc @_Z22sub_group_shuffle_downdj(%[[VAL_6]], %[[VAL_7]])
+    // CHECK:         llvm.call spir_funccc @_Z22sub_group_shuffle_downdj(%[[F64_VAL]], %[[OFFSET]])
     // CHECK:         llvm.mlir.constant(true) : i1
-    %shuffleResult0, %valid0 = gpu.shuffle idx %val0, %id, %width : i32
-    %shuffleResult1, %valid1 = gpu.shuffle xor %val1, %mask, %width : i64
-    %shuffleResult2, %valid2 = gpu.shuffle up %val2, %delta_up, %width : f32
-    %shuffleResult3, %valid3 = gpu.shuffle down %val3, %delta_down, %width : f64
+    %shuffleResult0, %valid0 = gpu.shuffle idx %i8_val, %offset, %width : i8
+    %shuffleResult1, %valid1 = gpu.shuffle xor %i16_val, %offset, %width : i16
+    %shuffleResult2, %valid2 = gpu.shuffle up %f16_val, %offset, %width : f16
+    %shuffleResult3, %valid3 = gpu.shuffle down %f64_val, %offset, %width : f64
     return
   }
 }
diff --git a/mlir/test/Dialect/GPU/invalid.mlir b/mlir/test/Dialect/GPU/invalid.mlir
index e9d8f329be8ede..0831a111d07df6 100644
--- a/mlir/test/Dialect/GPU/invalid.mlir
+++ b/mlir/test/Dialect/GPU/invalid.mlir
@@ -430,7 +430,7 @@ func.func @shuffle_mismatching_type(%arg0 : f32, %arg1 : i32, %arg2 : i32) {
 // -----
 
 func.func @shuffle_unsupported_type(%arg0 : index, %arg1 : i32, %arg2 : i32) {
-  // expected-error at +1 {{operand #0 must be i32, i64, f32 or f64}}
+  // expected-error at +1 {{operand #0 must be i8, i16, i32, i64, f16, f32 or f64}}
   %shfl, %pred = gpu.shuffle xor %arg0, %arg1, %arg2 : index
   return
 }

``````````

</details>


https://github.com/llvm/llvm-project/pull/104605


More information about the Mlir-commits mailing list