[Mlir-commits] [mlir] [mlir][tosa] Fix tosa.Resize-to-linalg lowering (PR #88514)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Fri Apr 12 06:48:40 PDT 2024
https://github.com/fabrizio-indirli created https://github.com/llvm/llvm-project/pull/88514
Simplify the lowering of tosa.resize to linalg, to match more closely the implementation from the TOSA specification: some internal arithmetic operations are now performed on the integer types (rather than in floating point) to reduce the loss of precision and the num of generated operations.
>From e12435523d8fa2716cee37d6f7097096281d32cb Mon Sep 17 00:00:00 2001
From: Fabrizio Indirli <Fabrizio.Indirli at arm.com>
Date: Fri, 12 Apr 2024 13:15:44 +0100
Subject: [PATCH] [mlir][tosa] Fix tosa.Resize-to-linalg lowering
Simplify the lowering of tosa.resize to linalg, to match
more closely the implementation from the TOSA specification:
some internal arithmetic operations are now performed on
the integer types (rather than in floating point) to reduce
the loss of precision and the num of generated operations.
Signed-off-by: Fabrizio Indirli <Fabrizio.Indirli at arm.com>
---
.../Conversion/TosaToLinalg/TosaToLinalg.cpp | 21 ++--
.../TosaToLinalg/tosa-to-linalg-resize.mlir | 112 ++++++++----------
2 files changed, 60 insertions(+), 73 deletions(-)
diff --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
index 7c477f2e1412be..1a743282adc487 100644
--- a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
@@ -1578,17 +1578,16 @@ class GenericResizeConverter : public OpRewritePattern<tosa::ResizeOp> {
}
// x = x * scale_d + offset;
// ix = floor(x / scale_n)
- // dx = x / scale_n - ix
- Value val = b.create<arith::UIToFPOp>(floatTy, in);
- scaleN = b.create<arith::UIToFPOp>(floatTy, scaleN);
- scaleD = b.create<arith::UIToFPOp>(floatTy, scaleD);
- offset = b.create<arith::SIToFPOp>(floatTy, offset);
- val = b.create<arith::MulFOp>(val, scaleD);
- val = b.create<arith::AddFOp>(val, offset);
- val = b.create<arith::DivFOp>(val, scaleN);
- index = b.create<math::FloorOp>(val);
- delta = b.create<arith::SubFOp>(val, index);
- index = b.create<arith::FPToSIOp>(b.getI32Type(), index);
+ Value val = b.create<arith::MulIOp>(in, scaleD);
+ val = b.create<arith::AddIOp>(val, offset);
+ index = b.create<arith::FloorDivSIOp>(val, scaleN);
+
+ // rx = x % scale_n
+ // dx = rx / scale_n
+ Value r = b.create<arith::RemSIOp>(val, scaleN);
+ Value rFp = b.create<arith::SIToFPOp>(floatTy, r);
+ Value scaleNfp = b.create<arith::UIToFPOp>(floatTy, scaleN);
+ delta = b.create<arith::DivFOp>(rFp, scaleNfp);
};
// Compute the ix and dx values for the X and Y dimensions - int case.
diff --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-resize.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-resize.mlir
index 468e92e2a2661f..d42d0a46692d47 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-resize.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-resize.mlir
@@ -304,42 +304,36 @@ func.func @resize_nearest_fp32(%input: tensor<1x50x48x1xf32>) -> () {
// CHECK-DAG: %[[XMAX:.*]] = arith.constant 47
// CHECK: %[[Y:.+]] = arith.index_cast %[[IDX1]]
// CHECK: %[[X:.+]] = arith.index_cast %[[IDX2]]
- // CHECK-DAG: %[[ISCALE_Y_N:.*]] = arith.constant 64
- // CHECK-DAG: %[[ISCALE_Y_D:.*]] = arith.constant 2
- // CHECK-DAG: %[[ISCALE_X_N:.*]] = arith.constant 64
- // CHECK-DAG: %[[ISCALE_X_D:.*]] = arith.constant 2
- // CHECK-DAG: %[[IOFFSET_Y:.*]] = arith.constant -31
- // CHECK-DAG: %[[IOFFSET_X:.*]] = arith.constant -31
- // CHECK-DAG: %[[IBORDER_Y:.*]] = arith.constant 31
- // CHECK-DAG: %[[IBORDER_X:.*]] = arith.constant 31
-
- // CHECK: %[[Y0:.+]] = arith.uitofp %[[Y]]
- // CHECK: %[[SCALE_Y_N:.*]] = arith.uitofp %[[ISCALE_Y_N]]
- // CHECK: %[[SCALE_Y_D:.*]] = arith.uitofp %[[ISCALE_Y_D]]
- // CHECK: %[[OFFSET_Y:.*]] = arith.sitofp %[[IOFFSET_Y]]
- // CHECK: %[[VAL_29:.*]] = arith.mulf %[[Y0]], %[[SCALE_Y_D]]
- // CHECK: %[[VAL_31:.*]] = arith.addf %[[VAL_29]], %[[OFFSET_Y]]
- // CHECK: %[[VAL_33:.*]] = arith.divf %[[VAL_31]], %[[SCALE_Y_N]]
- // CHECK: %[[VAL_35:.*]] = math.floor %[[VAL_33]]
- // CHECK: %[[D_Y:.*]] = arith.subf %[[VAL_33]], %[[VAL_35]]
- // CHECK: %[[VAL_39:.*]] = arith.fptosi %[[VAL_35]]
-
- // CHECK: %[[X0:.+]] = arith.uitofp %[[X]]
- // CHECK: %[[SCALE_X_N:.*]] = arith.uitofp %[[ISCALE_X_N]]
- // CHECK: %[[SCALE_X_D:.*]] = arith.uitofp %[[ISCALE_X_D]]
- // CHECK: %[[OFFSET_X:.*]] = arith.sitofp %[[IOFFSET_X]]
- // CHECK: %[[VAL_30:.*]] = arith.mulf %[[X0]], %[[SCALE_X_D]]
- // CHECK: %[[VAL_32:.*]] = arith.addf %[[VAL_30]], %[[OFFSET_X]]
- // CHECK: %[[VAL_34:.*]] = arith.divf %[[VAL_32]], %[[SCALE_X_N]]
- // CHECK: %[[VAL_36:.*]] = math.floor %[[VAL_34]]
- // CHECK: %[[D_X:.*]] = arith.subf %[[VAL_34]], %[[VAL_36]]
- // CHECK: %[[VAL_40:.*]] = arith.fptosi %[[VAL_36]]
+ // CHECK-DAG: %[[SCALE_Y_N:.*]] = arith.constant 64
+ // CHECK-DAG: %[[SCALE_Y_D:.*]] = arith.constant 2
+ // CHECK-DAG: %[[SCALE_X_N:.*]] = arith.constant 64
+ // CHECK-DAG: %[[SCALE_X_D:.*]] = arith.constant 2
+ // CHECK-DAG: %[[OFFSET_Y:.*]] = arith.constant -31
+ // CHECK-DAG: %[[OFFSET_X:.*]] = arith.constant -31
+ // CHECK-DAG: %[[BORDER_Y:.*]] = arith.constant 31
+ // CHECK-DAG: %[[BORDER_X:.*]] = arith.constant 31
+
+ // CHECK: %[[VAL_29:.*]] = arith.muli %[[Y]], %[[SCALE_Y_D]]
+ // CHECK: %[[Y_TEMP:.*]] = arith.addi %[[VAL_29]], %[[OFFSET_Y]]
+ // CHECK: %[[IY_TEMP:.*]] = arith.floordivsi %[[Y_TEMP]], %[[SCALE_Y_N]]
+ // CHECK: %[[RY:.*]] = arith.remsi %[[Y_TEMP]], %[[SCALE_Y_N]]
+ // CHECK: %[[RY_FP:.*]] = arith.sitofp %[[RY]]
+ // CHECK: %[[SCALE_Y_N_FP:.*]] = arith.uitofp %[[SCALE_Y_N]]
+ // CHECK: %[[D_Y:.*]] = arith.divf %[[RY_FP]], %[[SCALE_Y_N_FP]]
+
+ // CHECK: %[[VAL_30:.*]] = arith.muli %[[X]], %[[SCALE_X_D]]
+ // CHECK: %[[X_TEMP:.*]] = arith.addi %[[VAL_30]], %[[OFFSET_X]]
+ // CHECK: %[[IX_TEMP:.*]] = arith.floordivsi %[[X_TEMP]], %[[SCALE_X_N]]
+ // CHECK: %[[RX:.*]] = arith.remsi %[[X_TEMP]], %[[SCALE_X_N]]
+ // CHECK: %[[RX_FP:.*]] = arith.sitofp %[[RX]]
+ // CHECK: %[[SCALE_X_N_FP:.*]] = arith.uitofp %[[SCALE_X_N]]
+ // CHECK: %[[D_X:.*]] = arith.divf %[[RX_FP]], %[[SCALE_X_N_FP]]
// CHECK-DAG: %[[ONE:.*]] = arith.constant 1
// CHECK-DAG: %[[HALF:.*]] = arith.constant 5.000000e-01
// CHECK: %[[PRED_Y:.*]] = arith.cmpf oge, %[[D_Y]], %[[HALF]]
// CHECK: %[[ROUND_Y:.*]] = arith.select %[[PRED_Y]], %[[ONE]], %[[ZERO]]
- // CHECK: %[[VAL_48:.*]] = arith.addi %[[VAL_39]], %[[ROUND_Y]]
+ // CHECK: %[[VAL_48:.*]] = arith.addi %[[IY_TEMP]], %[[ROUND_Y]]
// CHECK: %[[LOWER:.*]] = arith.maxsi %[[ZERO]], %[[VAL_48]]
// CHECK: %[[CLAMPED:.*]] = arith.minsi %[[YMAX]], %[[LOWER]]
// CHECK: %[[IDY:.*]] = arith.index_cast %[[CLAMPED]]
@@ -347,7 +341,7 @@ func.func @resize_nearest_fp32(%input: tensor<1x50x48x1xf32>) -> () {
// CHECK-DAG: %[[HALF:.*]] = arith.constant 5.000000e-01
// CHECK: %[[PRED_X:.*]] = arith.cmpf oge, %[[D_X]], %[[HALF]]
// CHECK: %[[ROUND_X:.*]] = arith.select %[[PRED_X]], %[[ONE]], %[[ZERO]]
- // CHECK: %[[VAL_49:.*]] = arith.addi %[[VAL_40]], %[[ROUND_X]]
+ // CHECK: %[[VAL_49:.*]] = arith.addi %[[IX_TEMP]], %[[ROUND_X]]
// CHECK: %[[LOWER:.*]] = arith.maxsi %[[ZERO]], %[[VAL_49]]
// CHECK: %[[CLAMPED:.*]] = arith.minsi %[[XMAX]], %[[LOWER]]
// CHECK: %[[IDX:.*]] = arith.index_cast %[[CLAMPED]]
@@ -374,36 +368,30 @@ func.func @resize_bilinear_fp(%input: tensor<1x23x24x1xf32>) -> () {
// CHECK-DAG: %[[X_MAX:.*]] = arith.constant 23
// CHECK: %[[Y:.+]] = arith.index_cast %[[IDX_1]]
// CHECK: %[[X:.+]] = arith.index_cast %[[IDX_2]]
- // CHECK-DAG: %[[ISCALE_Y_N:.*]] = arith.constant 4
- // CHECK-DAG: %[[ISCALE_Y_D:.*]] = arith.constant 1
- // CHECK-DAG: %[[ISCALE_X_N:.*]] = arith.constant 4
- // CHECK-DAG: %[[ISCALE_X_D:.*]] = arith.constant 1
- // CHECK-DAG: %[[IOFFSET_Y:.*]] = arith.constant 0
- // CHECK-DAG: %[[IOFFSET_X:.*]] = arith.constant 0
- // CHECK-DAG: %[[IBORDER_Y:.*]] = arith.constant 0
- // CHECK-DAG: %[[IBORDER_X:.*]] = arith.constant 0
-
- // CHECK: %[[Y0:.+]] = arith.uitofp %[[Y]]
- // CHECK: %[[SCALE_Y_N:.*]] = arith.uitofp %[[ISCALE_Y_N]]
- // CHECK: %[[SCALE_Y_D:.*]] = arith.uitofp %[[ISCALE_Y_D]]
- // CHECK: %[[OFFSET_Y:.*]] = arith.sitofp %[[IOFFSET_Y]]
- // CHECK: %[[VAL_29:.*]] = arith.mulf %[[Y0]], %[[SCALE_Y_D]]
- // CHECK: %[[VAL_31:.*]] = arith.addf %[[VAL_29]], %[[OFFSET_Y]]
- // CHECK: %[[VAL_33:.*]] = arith.divf %[[VAL_31]], %[[SCALE_Y_N]]
- // CHECK: %[[VAL_35:.*]] = math.floor %[[VAL_33]]
- // CHECK: %[[D_Y:.*]] = arith.subf %[[VAL_33]], %[[VAL_35]]
- // CHECK: %[[I_Y:.*]] = arith.fptosi %[[VAL_35]]
-
- // CHECK: %[[X0:.+]] = arith.uitofp %[[X]]
- // CHECK: %[[SCALE_X_N:.*]] = arith.uitofp %[[ISCALE_X_N]]
- // CHECK: %[[SCALE_X_D:.*]] = arith.uitofp %[[ISCALE_X_D]]
- // CHECK: %[[OFFSET_X:.*]] = arith.sitofp %[[IOFFSET_X]]
- // CHECK: %[[VAL_30:.*]] = arith.mulf %[[X0]], %[[SCALE_X_D]]
- // CHECK: %[[VAL_32:.*]] = arith.addf %[[VAL_30]], %[[OFFSET_X]]
- // CHECK: %[[VAL_34:.*]] = arith.divf %[[VAL_32]], %[[SCALE_X_N]]
- // CHECK: %[[VAL_36:.*]] = math.floor %[[VAL_34]]
- // CHECK: %[[D_X:.*]] = arith.subf %[[VAL_34]], %[[VAL_36]]
- // CHECK: %[[I_X:.*]] = arith.fptosi %[[VAL_36]]
+ // CHECK-DAG: %[[SCALE_Y_N:.*]] = arith.constant 4
+ // CHECK-DAG: %[[SCALE_Y_D:.*]] = arith.constant 1
+ // CHECK-DAG: %[[SCALE_X_N:.*]] = arith.constant 4
+ // CHECK-DAG: %[[SCALE_X_D:.*]] = arith.constant 1
+ // CHECK-DAG: %[[OFFSET_Y:.*]] = arith.constant 0
+ // CHECK-DAG: %[[OFFSET_X:.*]] = arith.constant 0
+ // CHECK-DAG: %[[BORDER_Y:.*]] = arith.constant 0
+ // CHECK-DAG: %[[BORDER_X:.*]] = arith.constant 0
+
+ // CHECK: %[[VAL_29:.*]] = arith.muli %[[Y]], %[[SCALE_Y_D]]
+ // CHECK: %[[Y_TEMP:.*]] = arith.addi %[[VAL_29]], %[[OFFSET_Y]]
+ // CHECK: %[[I_Y:.*]] = arith.floordivsi %[[Y_TEMP]], %[[SCALE_Y_N]]
+ // CHECK: %[[RY:.*]] = arith.remsi %[[Y_TEMP]], %[[SCALE_Y_N]]
+ // CHECK: %[[RY_FP:.*]] = arith.sitofp %[[RY]]
+ // CHECK: %[[SCALE_Y_N_FP:.*]] = arith.uitofp %[[SCALE_Y_N]]
+ // CHECK: %[[D_Y:.*]] = arith.divf %[[RY_FP]], %[[SCALE_Y_N_FP]]
+
+ // CHECK: %[[VAL_30:.*]] = arith.muli %[[X]], %[[SCALE_X_D]]
+ // CHECK: %[[X_TEMP:.*]] = arith.addi %[[VAL_30]], %[[OFFSET_X]]
+ // CHECK: %[[I_X:.*]] = arith.floordivsi %[[X_TEMP]], %[[SCALE_X_N]]
+ // CHECK: %[[RX:.*]] = arith.remsi %[[X_TEMP]], %[[SCALE_X_N]]
+ // CHECK: %[[RX_FP:.*]] = arith.sitofp %[[RX]]
+ // CHECK: %[[SCALE_X_N_FP:.*]] = arith.uitofp %[[SCALE_X_N]]
+ // CHECK: %[[D_X:.*]] = arith.divf %[[RX_FP]], %[[SCALE_X_N_FP]]
// Compute the left, right, and top indices for the bilinear interpolation.
More information about the Mlir-commits
mailing list