[TOSA] Usage of 32bit integer for 'index to float' in rfft2d #75098

d-smirnov · 2023-12-11T21:19:14Z

Lowering of rfft2d to linalg now uses index to i32 cast if an output float is of 32bit and cast to i64 otherwise.
@eric-k256 @GeorgeARM

llvmbot · 2023-12-11T21:19:42Z

@llvm/pr-subscribers-mlir-linalg

Author: Dmitriy Smirnov (d-smirnov)

Changes

Lowering of rfft2d to linalg now uses index to i32 cast if an output float is of 32bit and cast to i64 otherwise.
@eric-k256 @GeorgeARM

Full diff: https://github.com/llvm/llvm-project/pull/75098.diff

2 Files Affected:

(modified) mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp (+5-2)
(modified) mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir (+24-24)

diff --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
index 3bf7bf12b5e96f..f0e4a00bdc5179 100644
--- a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
@@ -2284,8 +2284,11 @@ struct RFFT2dConverter final : public OpRewritePattern<RFFT2dOp> {
 
   static Value castIndexToFloat(OpBuilder &builder, Location loc,
                                 FloatType type, Value value) {
-    auto integerVal =
-        builder.create<arith::IndexCastUIOp>(loc, builder.getI64Type(), value);
+    auto integerVal = builder.create<arith::IndexCastUIOp>(
+        loc,
+        32 <= type.getIntOrFloatBitWidth() ? builder.getI32Type()
+                                           : builder.getI64Type(),
+        value);
 
     return builder.create<arith::UIToFPOp>(loc, type, integerVal);
   }
diff --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
index aa53b366f6da68..6f9597dd399af6 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
@@ -1691,10 +1691,10 @@ func.func @test_static_rfft2d(%arg0: tensor<5x5x8xf32>) -> (tensor<5x5x5xf32>, t
 // CHECK:   %[[EMPTY_1:.*]] = tensor.empty() : tensor<5x5x5xf32>
 // CHECK:   %[[VAR_3:.*]] = linalg.fill ins(%[[CST_ZERO:.*]]: f32) outs(%[[EMPTY_1:.*]] : tensor<5x5x5xf32>) -> tensor<5x5x5xf32>
 // CHECK:   %[[CST_PI:.*]] = arith.constant 6.28318548 : f32
-// CHECK:   %[[VAR_5:.*]] = arith.index_castui %[[CST_5:.*]] : index to i64
-// CHECK:   %[[VAR_6:.*]] = arith.uitofp %[[VAR_5:.*]] : i64 to f32
-// CHECK:   %[[VAR_7:.*]] = arith.index_castui %[[CST_8:.*]] : index to i64
-// CHECK:   %[[VAR_8:.*]] = arith.uitofp %[[VAR_7:.*]] : i64 to f32
+// CHECK:   %[[VAR_5:.*]] = arith.index_castui %[[CST_5:.*]] : index to i32
+// CHECK:   %[[VAR_6:.*]] = arith.uitofp %[[VAR_5:.*]] : i32 to f32
+// CHECK:   %[[VAR_7:.*]] = arith.index_castui %[[CST_8:.*]] : index to i32
+// CHECK:   %[[VAR_8:.*]] = arith.uitofp %[[VAR_7:.*]] : i32 to f32
 // CHECK:   linalg.generic {
 // CHECK:     indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]],
 // CHECK:     iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction"]}
@@ -1702,17 +1702,17 @@ func.func @test_static_rfft2d(%arg0: tensor<5x5x8xf32>) -> (tensor<5x5x5xf32>, t
 // CHECK:     outs(%[[VAR_1]], %[[VAR_3]] : tensor<5x5x5xf32>, tensor<5x5x5xf32>) {
 // CHECK:   ^bb0(%[[IN:.*]]: f32, %[[OUT_0:.*]]: f32, %[[OUT_1:.*]]: f32):
 // CHECK:     %[[INDEX_1:.*]] = linalg.index 1 : index
-// CHECK:     %[[VAR_10:.*]] = arith.index_castui %[[INDEX_1]] : index to i64
-// CHECK:     %[[VAR_11:.*]] = arith.uitofp %[[VAR_10]] : i64 to f32
+// CHECK:     %[[VAR_10:.*]] = arith.index_castui %[[INDEX_1]] : index to i32
+// CHECK:     %[[VAR_11:.*]] = arith.uitofp %[[VAR_10]] : i32 to f32
 // CHECK:     %[[INDEX_2:.*]] = linalg.index 2 : index
-// CHECK:     %[[VAR_13:.*]] = arith.index_castui %[[INDEX_2]] : index to i64
-// CHECK:     %[[VAR_14:.*]] = arith.uitofp %[[VAR_13]] : i64 to f32
+// CHECK:     %[[VAR_13:.*]] = arith.index_castui %[[INDEX_2]] : index to i32
+// CHECK:     %[[VAR_14:.*]] = arith.uitofp %[[VAR_13]] : i32 to f32
 // CHECK:     %[[INDEX_3:.*]] = linalg.index 3 : index
-// CHECK:     %[[VAR_16:.*]] = arith.index_castui %[[INDEX_3]] : index to i64
-// CHECK:     %[[VAR_17:.*]] = arith.uitofp %[[VAR_16]] : i64 to f32
+// CHECK:     %[[VAR_16:.*]] = arith.index_castui %[[INDEX_3]] : index to i32
+// CHECK:     %[[VAR_17:.*]] = arith.uitofp %[[VAR_16]] : i32 to f32
 // CHECK:     %[[INDEX_4:.*]] = linalg.index 4 : index
-// CHECK:     %[[VAR_19:.*]] = arith.index_castui %[[INDEX_4]] : index to i64
-// CHECK:     %[[VAR_20:.*]] = arith.uitofp %[[VAR_19]] : i64 to f32
+// CHECK:     %[[VAR_19:.*]] = arith.index_castui %[[INDEX_4]] : index to i32
+// CHECK:     %[[VAR_20:.*]] = arith.uitofp %[[VAR_19]] : i32 to f32
 // CHECK:     %[[VAR_21:.*]] = arith.mulf %[[VAR_17]], %[[VAR_11]] : f32
 // CHECK:     %[[VAR_22:.*]] = arith.mulf %[[VAR_20]], %[[VAR_14]] : f32
 // CHECK:     %[[XCOMP:.*]] = arith.divf %[[VAR_21]], %[[VAR_6]] : f32
@@ -1761,10 +1761,10 @@ func.func @test_dynamic_rfft2d(%arg0: tensor<?x?x?xf32>) -> (tensor<?x?x?xf32>,
 // CHECK:   %[[CST_2:.*]] = arith.constant 2 : index
 // CHECK:   %[[DIM_8:.*]] = tensor.dim %[[ARG_0]], %[[CST_2]] : tensor<?x?x?xf32>
 // CHECK:   %[[CST_9:.*]] = arith.constant 6.28318548 : f32
-// CHECK:   %[[VAR_6:.*]] = arith.index_castui %[[DIM_6]] : index to i64
-// CHECK:   %[[VAR_7:.*]] = arith.uitofp %[[VAR_6]] : i64 to f32
-// CHECK:   %[[VAR_8:.*]] = arith.index_castui %[[DIM_8]] : index to i64
-// CHECK:   %[[VAR_9:.*]] = arith.uitofp %[[VAR_8]] : i64 to f32
+// CHECK:   %[[VAR_6:.*]] = arith.index_castui %[[DIM_6]] : index to i32
+// CHECK:   %[[VAR_7:.*]] = arith.uitofp %[[VAR_6]] : i32 to f32
+// CHECK:   %[[VAR_8:.*]] = arith.index_castui %[[DIM_8]] : index to i32
+// CHECK:   %[[VAR_9:.*]] = arith.uitofp %[[VAR_8]] : i32 to f32
 // CHECK:   linalg.generic {
 // CHECK:     indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]],
 // CHECK:     iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction"]}
@@ -1772,17 +1772,17 @@ func.func @test_dynamic_rfft2d(%arg0: tensor<?x?x?xf32>) -> (tensor<?x?x?xf32>,
 // CHECK:     outs(%[[VAR_3]], %[[VAR_5]] : tensor<?x?x?xf32>, tensor<?x?x?xf32>) {
 // CHECK:   ^bb0(%[[IN:.*]]: f32, %[[OUT_0:.*]]: f32, %[[OUT_1:.*]]: f32):
 // CHECK:     %[[INDEX_1:.*]] = linalg.index 1 : index
-// CHECK:     %[[VAR_12:.*]] = arith.index_castui %[[INDEX_1]] : index to i64
-// CHECK:     %[[VAR_13:.*]] = arith.uitofp %[[VAR_12]] : i64 to f32
+// CHECK:     %[[VAR_12:.*]] = arith.index_castui %[[INDEX_1]] : index to i32
+// CHECK:     %[[VAR_13:.*]] = arith.uitofp %[[VAR_12]] : i32 to f32
 // CHECK:     %[[INDEX_2:.*]] = linalg.index 2 : index
-// CHECK:     %[[VAR_15:.*]] = arith.index_castui %[[INDEX_2]] : index to i64
-// CHECK:     %[[VAR_16:.*]] = arith.uitofp %[[VAR_15]] : i64 to f32
+// CHECK:     %[[VAR_15:.*]] = arith.index_castui %[[INDEX_2]] : index to i32
+// CHECK:     %[[VAR_16:.*]] = arith.uitofp %[[VAR_15]] : i32 to f32
 // CHECK:     %[[INDEX_3:.*]] = linalg.index 3 : index
-// CHECK:     %[[VAR_18:.*]] = arith.index_castui %[[INDEX_3]] : index to i64
-// CHECK:     %[[VAR_19:.*]] = arith.uitofp %[[VAR_18]] : i64 to f32
+// CHECK:     %[[VAR_18:.*]] = arith.index_castui %[[INDEX_3]] : index to i32
+// CHECK:     %[[VAR_19:.*]] = arith.uitofp %[[VAR_18]] : i32 to f32
 // CHECK:     %[[INDEX_4:.*]] = linalg.index 4 : index
-// CHECK:     %[[VAR_21:.*]] = arith.index_castui %[[INDEX_4]] : index to i64
-// CHECK:     %[[VAR_22:.*]] = arith.uitofp %[[VAR_21]] : i64 to f32
+// CHECK:     %[[VAR_21:.*]] = arith.index_castui %[[INDEX_4]] : index to i32
+// CHECK:     %[[VAR_22:.*]] = arith.uitofp %[[VAR_21]] : i32 to f32
 // CHECK:     %[[VAR_23:.*]] = arith.mulf %[[VAR_19]], %[[VAR_13]] : f32
 // CHECK:     %[[VAR_24:.*]] = arith.mulf %[[VAR_22]], %[[VAR_16]] : f32
 // CHECK:     %[[XCOMP:.*]] = arith.divf %[[VAR_23]], %[[VAR_7]] : f32

llvmbot · 2023-12-11T21:19:42Z

@llvm/pr-subscribers-mlir

Author: Dmitriy Smirnov (d-smirnov)

Changes

Lowering of rfft2d to linalg now uses index to i32 cast if an output float is of 32bit and cast to i64 otherwise.
@eric-k256 @GeorgeARM

Full diff: https://github.com/llvm/llvm-project/pull/75098.diff

2 Files Affected:

(modified) mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp (+5-2)
(modified) mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir (+24-24)

diff --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
index 3bf7bf12b5e96f..f0e4a00bdc5179 100644
--- a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
@@ -2284,8 +2284,11 @@ struct RFFT2dConverter final : public OpRewritePattern<RFFT2dOp> {
 
   static Value castIndexToFloat(OpBuilder &builder, Location loc,
                                 FloatType type, Value value) {
-    auto integerVal =
-        builder.create<arith::IndexCastUIOp>(loc, builder.getI64Type(), value);
+    auto integerVal = builder.create<arith::IndexCastUIOp>(
+        loc,
+        32 <= type.getIntOrFloatBitWidth() ? builder.getI32Type()
+                                           : builder.getI64Type(),
+        value);
 
     return builder.create<arith::UIToFPOp>(loc, type, integerVal);
   }
diff --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
index aa53b366f6da68..6f9597dd399af6 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
@@ -1691,10 +1691,10 @@ func.func @test_static_rfft2d(%arg0: tensor<5x5x8xf32>) -> (tensor<5x5x5xf32>, t
 // CHECK:   %[[EMPTY_1:.*]] = tensor.empty() : tensor<5x5x5xf32>
 // CHECK:   %[[VAR_3:.*]] = linalg.fill ins(%[[CST_ZERO:.*]]: f32) outs(%[[EMPTY_1:.*]] : tensor<5x5x5xf32>) -> tensor<5x5x5xf32>
 // CHECK:   %[[CST_PI:.*]] = arith.constant 6.28318548 : f32
-// CHECK:   %[[VAR_5:.*]] = arith.index_castui %[[CST_5:.*]] : index to i64
-// CHECK:   %[[VAR_6:.*]] = arith.uitofp %[[VAR_5:.*]] : i64 to f32
-// CHECK:   %[[VAR_7:.*]] = arith.index_castui %[[CST_8:.*]] : index to i64
-// CHECK:   %[[VAR_8:.*]] = arith.uitofp %[[VAR_7:.*]] : i64 to f32
+// CHECK:   %[[VAR_5:.*]] = arith.index_castui %[[CST_5:.*]] : index to i32
+// CHECK:   %[[VAR_6:.*]] = arith.uitofp %[[VAR_5:.*]] : i32 to f32
+// CHECK:   %[[VAR_7:.*]] = arith.index_castui %[[CST_8:.*]] : index to i32
+// CHECK:   %[[VAR_8:.*]] = arith.uitofp %[[VAR_7:.*]] : i32 to f32
 // CHECK:   linalg.generic {
 // CHECK:     indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]],
 // CHECK:     iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction"]}
@@ -1702,17 +1702,17 @@ func.func @test_static_rfft2d(%arg0: tensor<5x5x8xf32>) -> (tensor<5x5x5xf32>, t
 // CHECK:     outs(%[[VAR_1]], %[[VAR_3]] : tensor<5x5x5xf32>, tensor<5x5x5xf32>) {
 // CHECK:   ^bb0(%[[IN:.*]]: f32, %[[OUT_0:.*]]: f32, %[[OUT_1:.*]]: f32):
 // CHECK:     %[[INDEX_1:.*]] = linalg.index 1 : index
-// CHECK:     %[[VAR_10:.*]] = arith.index_castui %[[INDEX_1]] : index to i64
-// CHECK:     %[[VAR_11:.*]] = arith.uitofp %[[VAR_10]] : i64 to f32
+// CHECK:     %[[VAR_10:.*]] = arith.index_castui %[[INDEX_1]] : index to i32
+// CHECK:     %[[VAR_11:.*]] = arith.uitofp %[[VAR_10]] : i32 to f32
 // CHECK:     %[[INDEX_2:.*]] = linalg.index 2 : index
-// CHECK:     %[[VAR_13:.*]] = arith.index_castui %[[INDEX_2]] : index to i64
-// CHECK:     %[[VAR_14:.*]] = arith.uitofp %[[VAR_13]] : i64 to f32
+// CHECK:     %[[VAR_13:.*]] = arith.index_castui %[[INDEX_2]] : index to i32
+// CHECK:     %[[VAR_14:.*]] = arith.uitofp %[[VAR_13]] : i32 to f32
 // CHECK:     %[[INDEX_3:.*]] = linalg.index 3 : index
-// CHECK:     %[[VAR_16:.*]] = arith.index_castui %[[INDEX_3]] : index to i64
-// CHECK:     %[[VAR_17:.*]] = arith.uitofp %[[VAR_16]] : i64 to f32
+// CHECK:     %[[VAR_16:.*]] = arith.index_castui %[[INDEX_3]] : index to i32
+// CHECK:     %[[VAR_17:.*]] = arith.uitofp %[[VAR_16]] : i32 to f32
 // CHECK:     %[[INDEX_4:.*]] = linalg.index 4 : index
-// CHECK:     %[[VAR_19:.*]] = arith.index_castui %[[INDEX_4]] : index to i64
-// CHECK:     %[[VAR_20:.*]] = arith.uitofp %[[VAR_19]] : i64 to f32
+// CHECK:     %[[VAR_19:.*]] = arith.index_castui %[[INDEX_4]] : index to i32
+// CHECK:     %[[VAR_20:.*]] = arith.uitofp %[[VAR_19]] : i32 to f32
 // CHECK:     %[[VAR_21:.*]] = arith.mulf %[[VAR_17]], %[[VAR_11]] : f32
 // CHECK:     %[[VAR_22:.*]] = arith.mulf %[[VAR_20]], %[[VAR_14]] : f32
 // CHECK:     %[[XCOMP:.*]] = arith.divf %[[VAR_21]], %[[VAR_6]] : f32
@@ -1761,10 +1761,10 @@ func.func @test_dynamic_rfft2d(%arg0: tensor<?x?x?xf32>) -> (tensor<?x?x?xf32>,
 // CHECK:   %[[CST_2:.*]] = arith.constant 2 : index
 // CHECK:   %[[DIM_8:.*]] = tensor.dim %[[ARG_0]], %[[CST_2]] : tensor<?x?x?xf32>
 // CHECK:   %[[CST_9:.*]] = arith.constant 6.28318548 : f32
-// CHECK:   %[[VAR_6:.*]] = arith.index_castui %[[DIM_6]] : index to i64
-// CHECK:   %[[VAR_7:.*]] = arith.uitofp %[[VAR_6]] : i64 to f32
-// CHECK:   %[[VAR_8:.*]] = arith.index_castui %[[DIM_8]] : index to i64
-// CHECK:   %[[VAR_9:.*]] = arith.uitofp %[[VAR_8]] : i64 to f32
+// CHECK:   %[[VAR_6:.*]] = arith.index_castui %[[DIM_6]] : index to i32
+// CHECK:   %[[VAR_7:.*]] = arith.uitofp %[[VAR_6]] : i32 to f32
+// CHECK:   %[[VAR_8:.*]] = arith.index_castui %[[DIM_8]] : index to i32
+// CHECK:   %[[VAR_9:.*]] = arith.uitofp %[[VAR_8]] : i32 to f32
 // CHECK:   linalg.generic {
 // CHECK:     indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]],
 // CHECK:     iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction"]}
@@ -1772,17 +1772,17 @@ func.func @test_dynamic_rfft2d(%arg0: tensor<?x?x?xf32>) -> (tensor<?x?x?xf32>,
 // CHECK:     outs(%[[VAR_3]], %[[VAR_5]] : tensor<?x?x?xf32>, tensor<?x?x?xf32>) {
 // CHECK:   ^bb0(%[[IN:.*]]: f32, %[[OUT_0:.*]]: f32, %[[OUT_1:.*]]: f32):
 // CHECK:     %[[INDEX_1:.*]] = linalg.index 1 : index
-// CHECK:     %[[VAR_12:.*]] = arith.index_castui %[[INDEX_1]] : index to i64
-// CHECK:     %[[VAR_13:.*]] = arith.uitofp %[[VAR_12]] : i64 to f32
+// CHECK:     %[[VAR_12:.*]] = arith.index_castui %[[INDEX_1]] : index to i32
+// CHECK:     %[[VAR_13:.*]] = arith.uitofp %[[VAR_12]] : i32 to f32
 // CHECK:     %[[INDEX_2:.*]] = linalg.index 2 : index
-// CHECK:     %[[VAR_15:.*]] = arith.index_castui %[[INDEX_2]] : index to i64
-// CHECK:     %[[VAR_16:.*]] = arith.uitofp %[[VAR_15]] : i64 to f32
+// CHECK:     %[[VAR_15:.*]] = arith.index_castui %[[INDEX_2]] : index to i32
+// CHECK:     %[[VAR_16:.*]] = arith.uitofp %[[VAR_15]] : i32 to f32
 // CHECK:     %[[INDEX_3:.*]] = linalg.index 3 : index
-// CHECK:     %[[VAR_18:.*]] = arith.index_castui %[[INDEX_3]] : index to i64
-// CHECK:     %[[VAR_19:.*]] = arith.uitofp %[[VAR_18]] : i64 to f32
+// CHECK:     %[[VAR_18:.*]] = arith.index_castui %[[INDEX_3]] : index to i32
+// CHECK:     %[[VAR_19:.*]] = arith.uitofp %[[VAR_18]] : i32 to f32
 // CHECK:     %[[INDEX_4:.*]] = linalg.index 4 : index
-// CHECK:     %[[VAR_21:.*]] = arith.index_castui %[[INDEX_4]] : index to i64
-// CHECK:     %[[VAR_22:.*]] = arith.uitofp %[[VAR_21]] : i64 to f32
+// CHECK:     %[[VAR_21:.*]] = arith.index_castui %[[INDEX_4]] : index to i32
+// CHECK:     %[[VAR_22:.*]] = arith.uitofp %[[VAR_21]] : i32 to f32
 // CHECK:     %[[VAR_23:.*]] = arith.mulf %[[VAR_19]], %[[VAR_13]] : f32
 // CHECK:     %[[VAR_24:.*]] = arith.mulf %[[VAR_22]], %[[VAR_16]] : f32
 // CHECK:     %[[XCOMP:.*]] = arith.divf %[[VAR_23]], %[[VAR_7]] : f32

eric-k256

looks okay to me.

GeorgeARM · 2023-12-12T10:04:22Z

Thanks for this PR. Overall it looks fine to me.
Could you please elaborate in your commit-msg what is the reasoning behind this change?

If this is due to conversion targets not being able to handle 64bit integer arithmetic "easily" or at all (e.g. SPIRV) then I think that this might require more discussion. Not sure how we can handle or if we need a more immediate casting operator from index to floating point that can be lowered later on as per target needs.

MaheshRavishankar · 2023-12-12T10:10:31Z

Thanks for the tag @GeorgeARM . I am on vacation right now but @antiagainst or @kuhar can help here

RoboTux

LGTM.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

kuhar · 2023-12-12T15:45:52Z

If this is due to conversion targets not being able to handle 64bit integer arithmetic "easily" or at all (e.g. SPIRV) then I think that this might require more discussion. Not sure how we can handle or if we need a more immediate casting operator from index to floating point that can be lowered later on as per target needs.

FYI, I have patterns designed to help with this set of constraints:

llvm-project/mlir/lib/Dialect/Arith/Transforms/IntNarrowing.cpp

Lines 396 to 423 in 32532c2

    
           //===----------------------------------------------------------------------===// 
        
           // *IToFPOp Patterns 
        
           //===----------------------------------------------------------------------===// 
        
           template <typename IToFPOp, ExtensionKind Extension> 
        
           struct IToFPPattern final : NarrowingPattern<IToFPOp> { 
        
             using NarrowingPattern<IToFPOp>::NarrowingPattern; 
        
             LogicalResult matchAndRewrite(IToFPOp op, 
        
                                           PatternRewriter &rewriter) const override { 
        
               FailureOr<unsigned> narrowestWidth = 
        
                   calculateBitsRequired(op.getIn(), Extension); 
        
               if (failed(narrowestWidth)) 
        
                 return failure(); 
        
               FailureOr<Type> narrowTy = 
        
                   this->getNarrowType(*narrowestWidth, op.getIn().getType()); 
        
               if (failed(narrowTy)) 
        
                 return failure(); 
        
               Value newIn = rewriter.createOrFold<arith::TruncIOp>(op.getLoc(), *narrowTy, 
        
                                                                    op.getIn()); 
        
               rewriter.replaceOpWithNewOp<IToFPOp>(op, op.getType(), newIn); 
        
               return success(); 
        
             } 
        
           }; 
        
           using SIToFPPattern = IToFPPattern<arith::SIToFPOp, ExtensionKind::Sign>; 
        
           using UIToFPPattern = IToFPPattern<arith::UIToFPOp, ExtensionKind::Zero>;

d-smirnov · 2023-12-12T16:12:31Z

Yep, the goal of the patch is to omit `index -> i64` conversion if possible in lowering rfft2d Op as the subsequent transformation to SPIRV not able to "easily'' handle 64bit integer arithmetic (which seems excessive in this particular case anyway).

…

-- Dmitriy Smirnov

On Tue, 12 Dec 2023 at 15:46, Jakub Kuderski ***@***.***> wrote: If this is due to conversion targets not being able to handle 64bit integer arithmetic "easily" or at all (e.g. SPIRV) then I think that this might require more discussion. Not sure how we can handle or if we need a more immediate casting operator from index to floating point that can be lowered later on as per target needs. FYI, I have patterns designed with this set of constraints in mind: https://github.com/llvm/llvm-project/blob/32532c2bbedc6e15eb3eae65cb6049e196cef9ad/mlir/lib/Dialect/Arith/Transforms/IntNarrowing.cpp#L396-L423 — Reply to this email directly, view it on GitHub <#75098 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AGADO34YMXFCTTJMKQNPOXDYJB333AVCNFSM6AAAAABAQNLPVWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNJSGMYDCMZVGE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

github-actions · 2023-12-14T11:54:32Z

✅ With the latest revision this PR passed the C/C++ code formatter.

Lowering of rfft2d to linalg now uses index to i32 cast if an output float is of 32bit and 64bit otherwise.

d-smirnov · 2024-01-03T18:05:14Z

PR was accidentally closed. Reopened now.

GeorgeARM · 2024-01-05T09:49:34Z

LGTM

llvmbot added mlir:linalg mlir labels Dec 11, 2023

eric-k256 approved these changes Dec 12, 2023

View reviewed changes

GeorgeARM requested a review from MaheshRavishankar December 12, 2023 10:04

GeorgeARM requested a review from antiagainst December 12, 2023 10:11

d-smirnov force-pushed the index2float branch from 9d21ac7 to a2e20d0 Compare December 12, 2023 11:56

RoboTux reviewed Dec 12, 2023

View reviewed changes

kuhar reviewed Dec 12, 2023

View reviewed changes

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp Outdated Show resolved Hide resolved

d-smirnov force-pushed the index2float branch from f055337 to a2e20d0 Compare December 14, 2023 11:59

d-smirnov closed this Dec 21, 2023

d-smirnov force-pushed the index2float branch from a2e20d0 to f3dcc23 Compare December 21, 2023 16:20

[TOSA] Usage of 32bit integer for 'index to float' in rfft2d

f960c10

Lowering of rfft2d to linalg now uses index to i32 cast if an output float is of 32bit and 64bit otherwise.

d-smirnov reopened this Jan 3, 2024

d-smirnov mentioned this pull request Jan 4, 2024

[TOSA] FFT2D operator #77005

Merged

GeorgeARM approved these changes Jan 5, 2024

View reviewed changes

GeorgeARM merged commit 2952fb3 into llvm:main Jan 5, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[TOSA] Usage of 32bit integer for 'index to float' in rfft2d #75098

[TOSA] Usage of 32bit integer for 'index to float' in rfft2d #75098

Uh oh!

d-smirnov commented Dec 11, 2023

Uh oh!

llvmbot commented Dec 11, 2023

Uh oh!

llvmbot commented Dec 11, 2023

Uh oh!

eric-k256 left a comment

Uh oh!

GeorgeARM commented Dec 12, 2023 •

edited

Loading

Uh oh!

MaheshRavishankar commented Dec 12, 2023

Uh oh!

RoboTux left a comment

Uh oh!

Uh oh!

kuhar commented Dec 12, 2023 •

edited

Loading

Uh oh!

d-smirnov commented Dec 12, 2023 via email

Uh oh!

github-actions bot commented Dec 14, 2023 •

edited

Loading

Uh oh!

d-smirnov commented Jan 3, 2024

Uh oh!

GeorgeARM commented Jan 5, 2024

Uh oh!

Uh oh!

[TOSA] Usage of 32bit integer for 'index to float' in rfft2d #75098

[TOSA] Usage of 32bit integer for 'index to float' in rfft2d #75098

Uh oh!

Conversation

d-smirnov commented Dec 11, 2023

Uh oh!

llvmbot commented Dec 11, 2023

Uh oh!

llvmbot commented Dec 11, 2023

Uh oh!

eric-k256 left a comment

Choose a reason for hiding this comment

Uh oh!

GeorgeARM commented Dec 12, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MaheshRavishankar commented Dec 12, 2023

Uh oh!

RoboTux left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

kuhar commented Dec 12, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

d-smirnov commented Dec 12, 2023 via email

Uh oh!

github-actions bot commented Dec 14, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

d-smirnov commented Jan 3, 2024

Uh oh!

GeorgeARM commented Jan 5, 2024

Uh oh!

Uh oh!

GeorgeARM commented Dec 12, 2023 •

edited

Loading

kuhar commented Dec 12, 2023 •

edited

Loading

github-actions bot commented Dec 14, 2023 •

edited

Loading