[mlir][emitc] Support dense as init value for ShapedType #144826

jacquesguan · 2025-06-19T02:37:11Z

No description provided.

llvmbot · 2025-06-19T02:37:44Z

@llvm/pr-subscribers-mlir-emitc

Author: Jianjian Guan (jacquesguan)

Changes

Full diff: https://github.com/llvm/llvm-project/pull/144826.diff

2 Files Affected:

(modified) mlir/lib/Target/Cpp/TranslateToCpp.cpp (+14-4)
(modified) mlir/test/Target/Cpp/const.mlir (+8)

diff --git a/mlir/lib/Target/Cpp/TranslateToCpp.cpp b/mlir/lib/Target/Cpp/TranslateToCpp.cpp
index 067a0470b14e4..480962afb5319 100644
--- a/mlir/lib/Target/Cpp/TranslateToCpp.cpp
+++ b/mlir/lib/Target/Cpp/TranslateToCpp.cpp
@@ -1328,6 +1328,16 @@ LogicalResult CppEmitter::emitAttribute(Location loc, Attribute attr) {
     }
   };
 
+  auto getDenseElementType = [&](const Type &type) {
+    if (auto arrayType = dyn_cast<ArrayType>(type)) {
+      return arrayType.getElementType();
+    }
+    if (auto tensorType = dyn_cast<TensorType>(type)) {
+      return tensorType.getElementType();
+    }
+    return Type();
+  };
+
   // Print floating point attributes.
   if (auto fAttr = dyn_cast<FloatAttr>(attr)) {
     if (!isa<Float16Type, BFloat16Type, Float32Type, Float64Type>(
@@ -1362,8 +1372,8 @@ LogicalResult CppEmitter::emitAttribute(Location loc, Attribute attr) {
     }
   }
   if (auto dense = dyn_cast<DenseIntElementsAttr>(attr)) {
-    if (auto iType = dyn_cast<IntegerType>(
-            cast<TensorType>(dense.getType()).getElementType())) {
+    if (auto iType =
+            dyn_cast<IntegerType>(getDenseElementType(dense.getType()))) {
       os << '{';
       interleaveComma(dense, os, [&](const APInt &val) {
         printInt(val, shouldMapToUnsigned(iType.getSignedness()));
@@ -1371,8 +1381,8 @@ LogicalResult CppEmitter::emitAttribute(Location loc, Attribute attr) {
       os << '}';
       return success();
     }
-    if (auto iType = dyn_cast<IndexType>(
-            cast<TensorType>(dense.getType()).getElementType())) {
+    if (auto iType =
+            dyn_cast<IndexType>(getDenseElementType(dense.getType()))) {
       os << '{';
       interleaveComma(dense, os,
                       [&](const APInt &val) { printInt(val, false); });
diff --git a/mlir/test/Target/Cpp/const.mlir b/mlir/test/Target/Cpp/const.mlir
index d3656f830c48c..2a5ff1a0f0177 100644
--- a/mlir/test/Target/Cpp/const.mlir
+++ b/mlir/test/Target/Cpp/const.mlir
@@ -16,6 +16,8 @@ func.func @emitc_constant() {
   %c8 = "emitc.constant"(){value = dense<0> : tensor<i32>} : () -> tensor<i32>
   %c9 = "emitc.constant"(){value = dense<[0, 1]> : tensor<2xindex>} : () -> tensor<2xindex>
   %c10 = "emitc.constant"(){value = dense<[[0.0, 1.0], [2.0, 3.0]]> : tensor<2x2xf32>} : () -> tensor<2x2xf32>
+  %c11 = "emitc.constant"(){value = dense<[0, 1]> : !emitc.array<2xindex>} : () -> !emitc.array<2xindex>
+  %c12 = "emitc.constant"(){value = dense<[0.0, 1.0]> : !emitc.array<2xf32>} : () -> !emitc.array<2xf32>
   return
 }
 // CPP-DEFAULT: void emitc_constant() {
@@ -33,6 +35,8 @@ func.func @emitc_constant() {
 // CPP-DEFAULT-NEXT: Tensor<int32_t> [[V8:[^ ]*]] = {0};
 // CPP-DEFAULT-NEXT: Tensor<size_t, 2> [[V9:[^ ]*]] = {0, 1};
 // CPP-DEFAULT-NEXT: Tensor<float, 2, 2> [[V10:[^ ]*]] = {0.0e+00f, 1.000000000e+00f, 2.000000000e+00f, 3.000000000e+00f};
+// CPP-DEFAULT-NEXT: size_t [[V11:[^ ]*]][2] = {0, 1};
+// CPP-DEFAULT-NEXT: float [[V12:[^ ]*]][2] = {0.0e+00f, 1.000000000e+00f};
 
 // CPP-DECLTOP: void emitc_constant() {
 // CPP-DECLTOP-NEXT: int32_t [[V0:[^ ]*]];
@@ -49,6 +53,8 @@ func.func @emitc_constant() {
 // CPP-DECLTOP-NEXT: Tensor<int32_t> [[V8:[^ ]*]];
 // CPP-DECLTOP-NEXT: Tensor<size_t, 2> [[V9:[^ ]*]];
 // CPP-DECLTOP-NEXT: Tensor<float, 2, 2> [[V10:[^ ]*]];
+// CPP-DECLTOP-NEXT: size_t [[V11:[^ ]*]][2];
+// CPP-DECLTOP-NEXT: float [[V12:[^ ]*]][2];
 // CPP-DECLTOP-NEXT: [[V0]] = INT_MAX;
 // CPP-DECLTOP-NEXT: [[V1]] = 42;
 // CPP-DECLTOP-NEXT: [[V2]] = -1;
@@ -63,3 +69,5 @@ func.func @emitc_constant() {
 // CPP-DECLTOP-NEXT: [[V8]] = {0};
 // CPP-DECLTOP-NEXT: [[V9]] = {0, 1};
 // CPP-DECLTOP-NEXT: [[V10]] = {0.0e+00f, 1.000000000e+00f, 2.000000000e+00f, 3.000000000e+00f};
+// CPP-DECLTOP-NEXT: [[V11]] = {0, 1};
+// CPP-DECLTOP-NEXT: [[V12]] = {0.0e+00f, 1.000000000e+00f};

llvmbot · 2025-06-19T02:37:44Z

@llvm/pr-subscribers-mlir

Author: Jianjian Guan (jacquesguan)

Changes

Full diff: https://github.com/llvm/llvm-project/pull/144826.diff

2 Files Affected:

(modified) mlir/lib/Target/Cpp/TranslateToCpp.cpp (+14-4)
(modified) mlir/test/Target/Cpp/const.mlir (+8)

diff --git a/mlir/lib/Target/Cpp/TranslateToCpp.cpp b/mlir/lib/Target/Cpp/TranslateToCpp.cpp
index 067a0470b14e4..480962afb5319 100644
--- a/mlir/lib/Target/Cpp/TranslateToCpp.cpp
+++ b/mlir/lib/Target/Cpp/TranslateToCpp.cpp
@@ -1328,6 +1328,16 @@ LogicalResult CppEmitter::emitAttribute(Location loc, Attribute attr) {
     }
   };
 
+  auto getDenseElementType = [&](const Type &type) {
+    if (auto arrayType = dyn_cast<ArrayType>(type)) {
+      return arrayType.getElementType();
+    }
+    if (auto tensorType = dyn_cast<TensorType>(type)) {
+      return tensorType.getElementType();
+    }
+    return Type();
+  };
+
   // Print floating point attributes.
   if (auto fAttr = dyn_cast<FloatAttr>(attr)) {
     if (!isa<Float16Type, BFloat16Type, Float32Type, Float64Type>(
@@ -1362,8 +1372,8 @@ LogicalResult CppEmitter::emitAttribute(Location loc, Attribute attr) {
     }
   }
   if (auto dense = dyn_cast<DenseIntElementsAttr>(attr)) {
-    if (auto iType = dyn_cast<IntegerType>(
-            cast<TensorType>(dense.getType()).getElementType())) {
+    if (auto iType =
+            dyn_cast<IntegerType>(getDenseElementType(dense.getType()))) {
       os << '{';
       interleaveComma(dense, os, [&](const APInt &val) {
         printInt(val, shouldMapToUnsigned(iType.getSignedness()));
@@ -1371,8 +1381,8 @@ LogicalResult CppEmitter::emitAttribute(Location loc, Attribute attr) {
       os << '}';
       return success();
     }
-    if (auto iType = dyn_cast<IndexType>(
-            cast<TensorType>(dense.getType()).getElementType())) {
+    if (auto iType =
+            dyn_cast<IndexType>(getDenseElementType(dense.getType()))) {
       os << '{';
       interleaveComma(dense, os,
                       [&](const APInt &val) { printInt(val, false); });
diff --git a/mlir/test/Target/Cpp/const.mlir b/mlir/test/Target/Cpp/const.mlir
index d3656f830c48c..2a5ff1a0f0177 100644
--- a/mlir/test/Target/Cpp/const.mlir
+++ b/mlir/test/Target/Cpp/const.mlir
@@ -16,6 +16,8 @@ func.func @emitc_constant() {
   %c8 = "emitc.constant"(){value = dense<0> : tensor<i32>} : () -> tensor<i32>
   %c9 = "emitc.constant"(){value = dense<[0, 1]> : tensor<2xindex>} : () -> tensor<2xindex>
   %c10 = "emitc.constant"(){value = dense<[[0.0, 1.0], [2.0, 3.0]]> : tensor<2x2xf32>} : () -> tensor<2x2xf32>
+  %c11 = "emitc.constant"(){value = dense<[0, 1]> : !emitc.array<2xindex>} : () -> !emitc.array<2xindex>
+  %c12 = "emitc.constant"(){value = dense<[0.0, 1.0]> : !emitc.array<2xf32>} : () -> !emitc.array<2xf32>
   return
 }
 // CPP-DEFAULT: void emitc_constant() {
@@ -33,6 +35,8 @@ func.func @emitc_constant() {
 // CPP-DEFAULT-NEXT: Tensor<int32_t> [[V8:[^ ]*]] = {0};
 // CPP-DEFAULT-NEXT: Tensor<size_t, 2> [[V9:[^ ]*]] = {0, 1};
 // CPP-DEFAULT-NEXT: Tensor<float, 2, 2> [[V10:[^ ]*]] = {0.0e+00f, 1.000000000e+00f, 2.000000000e+00f, 3.000000000e+00f};
+// CPP-DEFAULT-NEXT: size_t [[V11:[^ ]*]][2] = {0, 1};
+// CPP-DEFAULT-NEXT: float [[V12:[^ ]*]][2] = {0.0e+00f, 1.000000000e+00f};
 
 // CPP-DECLTOP: void emitc_constant() {
 // CPP-DECLTOP-NEXT: int32_t [[V0:[^ ]*]];
@@ -49,6 +53,8 @@ func.func @emitc_constant() {
 // CPP-DECLTOP-NEXT: Tensor<int32_t> [[V8:[^ ]*]];
 // CPP-DECLTOP-NEXT: Tensor<size_t, 2> [[V9:[^ ]*]];
 // CPP-DECLTOP-NEXT: Tensor<float, 2, 2> [[V10:[^ ]*]];
+// CPP-DECLTOP-NEXT: size_t [[V11:[^ ]*]][2];
+// CPP-DECLTOP-NEXT: float [[V12:[^ ]*]][2];
 // CPP-DECLTOP-NEXT: [[V0]] = INT_MAX;
 // CPP-DECLTOP-NEXT: [[V1]] = 42;
 // CPP-DECLTOP-NEXT: [[V2]] = -1;
@@ -63,3 +69,5 @@ func.func @emitc_constant() {
 // CPP-DECLTOP-NEXT: [[V8]] = {0};
 // CPP-DECLTOP-NEXT: [[V9]] = {0, 1};
 // CPP-DECLTOP-NEXT: [[V10]] = {0.0e+00f, 1.000000000e+00f, 2.000000000e+00f, 3.000000000e+00f};
+// CPP-DECLTOP-NEXT: [[V11]] = {0, 1};
+// CPP-DECLTOP-NEXT: [[V12]] = {0.0e+00f, 1.000000000e+00f};

tpopp

Left one comment but removed myself as a reviewer as I'm not appropriate to guide or approve this.

tpopp · 2025-06-20T09:43:38Z

mlir/lib/Target/Cpp/TranslateToCpp.cpp

Should this be ShapedType instead? Also, it might be worth considering adding an assert

Thanks for comment, changed to use ShapedType and added assert.

jacquesguan · 2025-08-12T03:24:36Z

ping

aniragil

Sorry for the delay, slipped under the radar.
Title mismatch? IIUC this patch is extending existing support of dense as init value from tensor type to shaped type?
LGTM with that assert nit. Curious about motivation: lowering? a step towards emitc-only module before translation?

aniragil · 2025-08-12T08:56:24Z

mlir/lib/Target/Cpp/TranslateToCpp.cpp

cast asserts by itself, no need to assert separately.

jacquesguan · 2025-08-18T09:11:58Z

Sorry for the delay, slipped under the radar. Title mismatch? IIUC this patch is extending existing support of dense as init value from tensor type to shaped type? LGTM with that assert nit. Curious about motivation: lowering? a step towards emitc-only module before translation?

Thanks for comment, title has been changed now. The motivation of this pr is lowering, for some custom ops, we need to generate some array type in emitc.

aniragil

LGTM, thanks!

jacquesguan requested review from TinaAMD, mgehre-amd and tpopp June 19, 2025 02:37

llvmbot added mlir mlir:emitc labels Jun 19, 2025

jacquesguan requested a review from marbre June 20, 2025 06:27

tpopp removed their request for review June 20, 2025 09:42

tpopp reviewed Jun 20, 2025

View reviewed changes

jacquesguan force-pushed the mlir-emitc-constant-dense branch from 2c07aa7 to 802cd0d Compare June 25, 2025 03:33

rengolin requested a review from aniragil August 12, 2025 08:32

aniragil requested changes Aug 12, 2025

View reviewed changes

[mlir][emitc] Support dense as init value for ShapedType

ac19730

jacquesguan force-pushed the mlir-emitc-constant-dense branch from 802cd0d to ac19730 Compare August 18, 2025 09:03

jacquesguan changed the title ~~[mlir][emitc] Support dense as init value of constantOp~~ [mlir][emitc] Support dense as init value for ShapedType Aug 18, 2025

aniragil approved these changes Aug 18, 2025

View reviewed changes

jacquesguan merged commit 1eb5b18 into llvm:main Aug 19, 2025
9 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[mlir][emitc] Support dense as init value for ShapedType #144826

[mlir][emitc] Support dense as init value for ShapedType #144826

Uh oh!

jacquesguan commented Jun 19, 2025

Uh oh!

llvmbot commented Jun 19, 2025

Uh oh!

llvmbot commented Jun 19, 2025

Uh oh!

tpopp left a comment

Uh oh!

tpopp Jun 20, 2025

Uh oh!

jacquesguan Jun 25, 2025

Uh oh!

jacquesguan commented Aug 12, 2025

Uh oh!

aniragil left a comment

Uh oh!

aniragil Aug 12, 2025

Uh oh!

jacquesguan Aug 18, 2025

Uh oh!

jacquesguan commented Aug 18, 2025

Uh oh!

aniragil left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[mlir][emitc] Support dense as init value for ShapedType #144826

[mlir][emitc] Support dense as init value for ShapedType #144826

Uh oh!

Conversation

jacquesguan commented Jun 19, 2025

Uh oh!

llvmbot commented Jun 19, 2025

Uh oh!

llvmbot commented Jun 19, 2025

Uh oh!

tpopp left a comment

Choose a reason for hiding this comment

Uh oh!

tpopp Jun 20, 2025

Choose a reason for hiding this comment

Uh oh!

jacquesguan Jun 25, 2025

Choose a reason for hiding this comment

Uh oh!

jacquesguan commented Aug 12, 2025

Uh oh!

aniragil left a comment

Choose a reason for hiding this comment

Uh oh!

aniragil Aug 12, 2025

Choose a reason for hiding this comment

Uh oh!

jacquesguan Aug 18, 2025

Choose a reason for hiding this comment

Uh oh!

jacquesguan commented Aug 18, 2025

Uh oh!

aniragil left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants