[Mlir-commits] [mlir] b92ff6b - [ROCDL] Added `rocdl.cvt.scale.pk8` ops (#161411)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Thu Oct 2 07:14:23 PDT 2025
Author: ravil-mobile
Date: 2025-10-02T10:14:19-04:00
New Revision: b92ff6b2099fa5103a0350028c6752ad953b9b8d
URL: https://github.com/llvm/llvm-project/commit/b92ff6b2099fa5103a0350028c6752ad953b9b8d
DIFF: https://github.com/llvm/llvm-project/commit/b92ff6b2099fa5103a0350028c6752ad953b9b8d.diff
LOG: [ROCDL] Added `rocdl.cvt.scale.pk8` ops (#161411)
This patch introduces some missing FP conversion instructions in the
ROCDL dialect
Specifically:
- Downscaling 8x packed F16, Bf16, Fp32 values to Fp8, Bf8, Fp4
Tests:
- Added lit-tests to check MLIR -> LLVM lowering
Added:
Modified:
mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td
mlir/test/Dialect/LLVMIR/rocdl.mlir
mlir/test/Target/LLVMIR/rocdl.mlir
Removed:
################################################################################
diff --git a/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td b/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td
index 8b687a7f29bef..29001e26eaaaf 100644
--- a/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td
+++ b/mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td
@@ -985,7 +985,6 @@ class ScaleArgInfo<TypeConstraint argTyVal, string typeName> {
//===---------------------------------------------------------------------===//
// Scaled {fp4,bf8,fp8} to {bf16,f16,f32} conversion intrinsics
//===---------------------------------------------------------------------===//
-
foreach smallT = [
ScaleArgInfo<I32, "Fp4">,
ScaleArgInfo<ROCDL_V2I32Type, "Fp8">,
@@ -996,6 +995,8 @@ foreach smallT = [
ScaleArgInfo<ROCDL_V8BF16Type, "Bf16">,
ScaleArgInfo<ROCDL_V8F32Type, "F32">,
] in {
+
+ // Up-scaling
def ROCDL_CvtPkScalePk8 # largeT.nameForOp # smallT.nameForOp # Op :
ROCDL_ConcreteNonMemIntrOp<"cvt.scale.pk8." # largeT.name # "." # smallT.name,
[Pure], 1, [2], ["scaleSel"]>,
@@ -1010,13 +1011,30 @@ foreach smallT = [
attr-dict $src `,` $scale `[` $scaleSel `]` `:` type($res)
}];
}
+
+ // Down-scaling
+ def ROCDL_CvtScaleF32Pk8 # smallT.nameForOp # largeT.nameForOp # Op :
+ ROCDL_ConcreteNonMemIntrOp<"cvt.scalef32.pk8." # smallT.name # "." # largeT.name,
+ [Pure], 1>,
+ Arguments<(ins largeT.type:$src, F32:$scale)> {
+ let results = (outs smallT.type:$res);
+ let summary = "Scale and convert packed "
+ # largeT.name # " to packed " # smallT.name ;
+ let description = [{
+ Convert 8 packed }] # largeT.name # [{ values to packed }]
+ # smallT.name # [{, multiplying by the exponent part of `scale`
+ before doing so. This op is for gfx1250+ arch.
+ }];
+ let assemblyFormat = [{
+ attr-dict $src `,` $scale `:` type($res)
+ }];
+ }
} // foreach largeT
} // foreach smallTOp
//===---------------------------------------------------------------------===//
// Scaled {bf6,fp6} to {bf16,f16,f32} conversion intrinsics
//===---------------------------------------------------------------------===//
-
foreach smallT = [
ScaleArgInfo<ROCDL_V3I32Type, "Fp6">,
ScaleArgInfo<ROCDL_V3I32Type, "Bf6">
diff --git a/mlir/test/Dialect/LLVMIR/rocdl.mlir b/mlir/test/Dialect/LLVMIR/rocdl.mlir
index 0bad151570029..6134695e9ced6 100644
--- a/mlir/test/Dialect/LLVMIR/rocdl.mlir
+++ b/mlir/test/Dialect/LLVMIR/rocdl.mlir
@@ -1068,6 +1068,38 @@ llvm.func @rocdl.cvt.scale.pk8(%i32: i32, %v2xi32: vector<2xi32>, %scale: i32) {
// -----
+// CHECK-LABEL: rocdl.cvt.scalef32.pk8
+llvm.func @rocdl.cvt.scalef32.pk8(%v8xf32: vector<8xf32>,
+ %v8xf16: vector<8xf16>,
+ %v8xbf16: vector<8xbf16>,
+ %scale: f32) {
+
+ // CHECK: rocdl.cvt.scalef32.pk8.fp8.f32
+ %0 = rocdl.cvt.scalef32.pk8.fp8.f32 %v8xf32, %scale : vector<2xi32>
+ // CHECK: rocdl.cvt.scalef32.pk8.bf8.f32
+ %1 = rocdl.cvt.scalef32.pk8.bf8.f32 %v8xf32, %scale : vector<2xi32>
+ // CHECK: rocdl.cvt.scalef32.pk8.fp4.f32
+ %2 = rocdl.cvt.scalef32.pk8.fp4.f32 %v8xf32, %scale : i32
+
+ // CHECK: rocdl.cvt.scalef32.pk8.fp8.f16
+ %3 = rocdl.cvt.scalef32.pk8.fp8.f16 %v8xf16, %scale : vector<2xi32>
+ // CHECK: rocdl.cvt.scalef32.pk8.bf8.f16
+ %4 = rocdl.cvt.scalef32.pk8.bf8.f16 %v8xf16, %scale : vector<2xi32>
+ // CHECK: rocdl.cvt.scalef32.pk8.fp4.f16
+ %5 = rocdl.cvt.scalef32.pk8.fp4.f16 %v8xf16, %scale : i32
+
+ // CHECK: rocdl.cvt.scalef32.pk8.fp8.bf16
+ %6 = rocdl.cvt.scalef32.pk8.fp8.bf16 %v8xbf16, %scale : vector<2xi32>
+ // CHECK: rocdl.cvt.scalef32.pk8.bf8.bf16
+ %7 = rocdl.cvt.scalef32.pk8.bf8.bf16 %v8xbf16, %scale : vector<2xi32>
+ // CHECK: rocdl.cvt.scalef32.pk8.fp4.bf16
+ %8 = rocdl.cvt.scalef32.pk8.fp4.bf16 %v8xbf16, %scale : i32
+
+ llvm.return
+}
+
+// -----
+
// CHECK-LABEL: rocdl.cvt.scale.pk16
llvm.func @rocdl.cvt.scale.pk16(%v3xi32: vector<3xi32>, %scale:i32) {
diff --git a/mlir/test/Target/LLVMIR/rocdl.mlir b/mlir/test/Target/LLVMIR/rocdl.mlir
index e043a8c533d05..00ee6b795c43a 100644
--- a/mlir/test/Target/LLVMIR/rocdl.mlir
+++ b/mlir/test/Target/LLVMIR/rocdl.mlir
@@ -1340,6 +1340,34 @@ llvm.func @rocdl.cvt.scale.pk8(%i32: i32, %v2xi32: vector<2xi32>, %scale: i32) {
llvm.return
}
+// CHECK-LABEL: rocdl.cvt.scalef32.pk8
+// CHECK-SAME:(<8 x float> %[[V8F32:.+]], <8 x half> %[[V8F16:.+]], <8 x bfloat> %[[V8BF16:.+]], float %[[SCALE:.+]])
+llvm.func @rocdl.cvt.scalef32.pk8(%v8xf32: vector<8xf32>, %v8xf16: vector<8xf16>, %v8xbf16: vector<8xbf16>, %scale: f32) {
+
+ // CHECK: call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.fp8.f32(<8 x float> %[[V8F32]], float %[[SCALE]])
+ %0 = rocdl.cvt.scalef32.pk8.fp8.f32 %v8xf32, %scale : vector<2xi32>
+ // CHECK: call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.bf8.f32(<8 x float> %[[V8F32]], float %[[SCALE]])
+ %1 = rocdl.cvt.scalef32.pk8.bf8.f32 %v8xf32, %scale : vector<2xi32>
+ // CHECK: call i32 @llvm.amdgcn.cvt.scalef32.pk8.fp4.f32(<8 x float> %[[V8F32]], float %[[SCALE]])
+ %2 = rocdl.cvt.scalef32.pk8.fp4.f32 %v8xf32, %scale : i32
+
+ // CHECK: call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.fp8.f16(<8 x half> %[[V8F16]], float %[[SCALE]])
+ %3 = rocdl.cvt.scalef32.pk8.fp8.f16 %v8xf16, %scale : vector<2xi32>
+ // CHECK: call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.bf8.f16(<8 x half> %[[V8F16]], float %[[SCALE]])
+ %4 = rocdl.cvt.scalef32.pk8.bf8.f16 %v8xf16, %scale : vector<2xi32>
+ // CHECK: call i32 @llvm.amdgcn.cvt.scalef32.pk8.fp4.f16(<8 x half> %[[V8F16]], float %[[SCALE]])
+ %5 = rocdl.cvt.scalef32.pk8.fp4.f16 %v8xf16, %scale : i32
+
+ // CHECK: call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.fp8.bf16(<8 x bfloat> %[[V8BF16]], float %[[SCALE]])
+ %6 = rocdl.cvt.scalef32.pk8.fp8.bf16 %v8xbf16, %scale : vector<2xi32>
+ // CHECK: call <2 x i32> @llvm.amdgcn.cvt.scalef32.pk8.bf8.bf16(<8 x bfloat> %[[V8BF16]], float %[[SCALE]])
+ %7 = rocdl.cvt.scalef32.pk8.bf8.bf16 %v8xbf16, %scale : vector<2xi32>
+ // CHECK: call i32 @llvm.amdgcn.cvt.scalef32.pk8.fp4.bf16(<8 x bfloat> %[[V8BF16]], float %[[SCALE]])
+ %8 = rocdl.cvt.scalef32.pk8.fp4.bf16 %v8xbf16, %scale : i32
+
+ llvm.return
+}
+
// CHECK-LABEL: @rocdl.cvt.scale.pk16
// CHECK-SAME:(<3 x i32> %[[SRC0:.+]], i32 %[[SCALE:.+]])
llvm.func @rocdl.cvt.scale.pk16(%v3xi32: vector<3xi32>, %scale:i32) {
More information about the Mlir-commits
mailing list