Reland non-splat `tensor.from_elements` to `flow` (#18049)
https://github.com/iree-org/iree/issues/17086
We can support multipel `tensor.from_elements` through using
`flow.tensor.store` for each elements into an empty array.
---------
Signed-off-by: Rob Suderman <rob.suderman@gmail.com>
diff --git a/build_tools/pkgci/external_test_suite/onnx_cpu_llvm_sync.json b/build_tools/pkgci/external_test_suite/onnx_cpu_llvm_sync.json
index 8d461e7..3646f61 100644
--- a/build_tools/pkgci/external_test_suite/onnx_cpu_llvm_sync.json
+++ b/build_tools/pkgci/external_test_suite/onnx_cpu_llvm_sync.json
@@ -86,16 +86,13 @@
"onnx/node/generated/test_castlike_STRING_to_FLOAT_expanded",
"onnx/node/generated/test_center_crop_pad_crop",
"onnx/node/generated/test_center_crop_pad_crop_and_pad",
- "onnx/node/generated/test_center_crop_pad_crop_and_pad_expanded",
"onnx/node/generated/test_center_crop_pad_crop_axes_chw",
"onnx/node/generated/test_center_crop_pad_crop_axes_chw_expanded",
"onnx/node/generated/test_center_crop_pad_crop_axes_hwc",
"onnx/node/generated/test_center_crop_pad_crop_axes_hwc_expanded",
- "onnx/node/generated/test_center_crop_pad_crop_expanded",
"onnx/node/generated/test_center_crop_pad_crop_negative_axes_hwc",
"onnx/node/generated/test_center_crop_pad_crop_negative_axes_hwc_expanded",
"onnx/node/generated/test_center_crop_pad_pad",
- "onnx/node/generated/test_center_crop_pad_pad_expanded",
"onnx/node/generated/test_col2im",
"onnx/node/generated/test_col2im_5d",
"onnx/node/generated/test_col2im_dilations",
diff --git a/build_tools/pkgci/external_test_suite/onnx_gpu_cuda.json b/build_tools/pkgci/external_test_suite/onnx_gpu_cuda.json
index a41a6e0..92ccf6d 100644
--- a/build_tools/pkgci/external_test_suite/onnx_gpu_cuda.json
+++ b/build_tools/pkgci/external_test_suite/onnx_gpu_cuda.json
@@ -8,14 +8,16 @@
"--device=cuda"
],
"skip_compile_tests": [
- "onnx/node/generated/test_dequantizelinear"
+ "onnx/node/generated/test_dequantizelinear",
+ "onnx/node/generated/test_group_normalization_example_expanded",
+ "onnx/node/generated/test_group_normalization_epsilon_expanded",
],
"skip_run_tests": [
"onnx/node/generated/test_gather_elements_negative_indices",
"onnx/node/generated/test_gridsample_zeros_padding",
- "onnx/node/generated/test_scatter_elements_with_negative_indices",
"onnx/node/generated/test_resize_downsample_scales_linear_align_corners",
- "onnx/node/generated/test_resize_downsample_scales_linear_half_pixel_symmetric"
+ "onnx/node/generated/test_resize_downsample_scales_linear_half_pixel_symmetric",
+ "onnx/node/generated/test_scatter_elements_with_negative_indices"
],
"expected_compile_failures": [
"onnx/node/generated/test_adagrad",
@@ -92,16 +94,13 @@
"onnx/node/generated/test_castlike_STRING_to_FLOAT_expanded",
"onnx/node/generated/test_center_crop_pad_crop",
"onnx/node/generated/test_center_crop_pad_crop_and_pad",
- "onnx/node/generated/test_center_crop_pad_crop_and_pad_expanded",
"onnx/node/generated/test_center_crop_pad_crop_axes_chw",
"onnx/node/generated/test_center_crop_pad_crop_axes_chw_expanded",
"onnx/node/generated/test_center_crop_pad_crop_axes_hwc",
"onnx/node/generated/test_center_crop_pad_crop_axes_hwc_expanded",
- "onnx/node/generated/test_center_crop_pad_crop_expanded",
"onnx/node/generated/test_center_crop_pad_crop_negative_axes_hwc",
"onnx/node/generated/test_center_crop_pad_crop_negative_axes_hwc_expanded",
"onnx/node/generated/test_center_crop_pad_pad",
- "onnx/node/generated/test_center_crop_pad_pad_expanded",
"onnx/node/generated/test_col2im",
"onnx/node/generated/test_col2im_5d",
"onnx/node/generated/test_col2im_dilations",
@@ -158,9 +157,7 @@
"onnx/node/generated/test_gridsample_volumetric_nearest_align_corners_0",
"onnx/node/generated/test_gridsample_volumetric_nearest_align_corners_1",
"onnx/node/generated/test_group_normalization_epsilon",
- "onnx/node/generated/test_group_normalization_epsilon_expanded",
"onnx/node/generated/test_group_normalization_example",
- "onnx/node/generated/test_group_normalization_example_expanded",
"onnx/node/generated/test_gru_batchwise",
"onnx/node/generated/test_gru_defaults",
"onnx/node/generated/test_gru_seq_length",
diff --git a/build_tools/pkgci/external_test_suite/onnx_gpu_rocm_rdna3.json b/build_tools/pkgci/external_test_suite/onnx_gpu_rocm_rdna3.json
index c6a9a35..6f31aad 100644
--- a/build_tools/pkgci/external_test_suite/onnx_gpu_rocm_rdna3.json
+++ b/build_tools/pkgci/external_test_suite/onnx_gpu_rocm_rdna3.json
@@ -9,7 +9,9 @@
"--device=hip"
],
"skip_compile_tests": [
- "onnx/node/generated/test_dequantizelinear"
+ "onnx/node/generated/test_dequantizelinear",
+ "onnx/node/generated/test_group_normalization_epsilon_expanded",
+ "onnx/node/generated/test_group_normalization_example_expanded"
],
"skip_run_tests": [],
"expected_compile_failures": [
@@ -87,16 +89,13 @@
"onnx/node/generated/test_castlike_STRING_to_FLOAT_expanded",
"onnx/node/generated/test_center_crop_pad_crop",
"onnx/node/generated/test_center_crop_pad_crop_and_pad",
- "onnx/node/generated/test_center_crop_pad_crop_and_pad_expanded",
"onnx/node/generated/test_center_crop_pad_crop_axes_chw",
"onnx/node/generated/test_center_crop_pad_crop_axes_chw_expanded",
"onnx/node/generated/test_center_crop_pad_crop_axes_hwc",
"onnx/node/generated/test_center_crop_pad_crop_axes_hwc_expanded",
- "onnx/node/generated/test_center_crop_pad_crop_expanded",
"onnx/node/generated/test_center_crop_pad_crop_negative_axes_hwc",
"onnx/node/generated/test_center_crop_pad_crop_negative_axes_hwc_expanded",
"onnx/node/generated/test_center_crop_pad_pad",
- "onnx/node/generated/test_center_crop_pad_pad_expanded",
"onnx/node/generated/test_col2im",
"onnx/node/generated/test_col2im_5d",
"onnx/node/generated/test_col2im_dilations",
@@ -153,9 +152,7 @@
"onnx/node/generated/test_gridsample_volumetric_nearest_align_corners_0",
"onnx/node/generated/test_gridsample_volumetric_nearest_align_corners_1",
"onnx/node/generated/test_group_normalization_epsilon",
- "onnx/node/generated/test_group_normalization_epsilon_expanded",
"onnx/node/generated/test_group_normalization_example",
- "onnx/node/generated/test_group_normalization_example_expanded",
"onnx/node/generated/test_gru_batchwise",
"onnx/node/generated/test_gru_defaults",
"onnx/node/generated/test_gru_seq_length",
diff --git a/build_tools/pkgci/external_test_suite/onnx_gpu_vulkan.json b/build_tools/pkgci/external_test_suite/onnx_gpu_vulkan.json
index bf88b19..c62935e 100644
--- a/build_tools/pkgci/external_test_suite/onnx_gpu_vulkan.json
+++ b/build_tools/pkgci/external_test_suite/onnx_gpu_vulkan.json
@@ -8,11 +8,12 @@
"--device=vulkan"
],
"skip_compile_tests": [
- "onnx/node/generated/test_dequantizelinear"
+ "onnx/node/generated/test_dequantizelinear",
+ "onnx/node/generated/test_det_nd",
+ "onnx/node/generated/test_group_normalization_epsilon_expanded",
+ "onnx/node/generated/test_group_normalization_example_expanded"
],
- "skip_run_tests": [
- "onnx/node/generated/test_det_nd"
- ],
+ "skip_run_tests": [],
"expected_compile_failures": [
"onnx/node/generated/test_adagrad",
"onnx/node/generated/test_adagrad_multiple",
@@ -187,9 +188,7 @@
"onnx/node/generated/test_gridsample_volumetric_nearest_align_corners_1",
"onnx/node/generated/test_gridsample_zeros_padding",
"onnx/node/generated/test_group_normalization_epsilon",
- "onnx/node/generated/test_group_normalization_epsilon_expanded",
"onnx/node/generated/test_group_normalization_example",
- "onnx/node/generated/test_group_normalization_example_expanded",
"onnx/node/generated/test_gru_batchwise",
"onnx/node/generated/test_gru_defaults",
"onnx/node/generated/test_gru_seq_length",
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Conversion/TensorToFlow/Patterns.cpp b/compiler/src/iree/compiler/Dialect/Flow/Conversion/TensorToFlow/Patterns.cpp
index 0a03362..d5794af 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Conversion/TensorToFlow/Patterns.cpp
+++ b/compiler/src/iree/compiler/Dialect/Flow/Conversion/TensorToFlow/Patterns.cpp
@@ -187,17 +187,33 @@
}
auto tensorType = op.getType();
if (!tensorType.hasRank()) {
- return failure();
+ return rewriter.notifyMatchFailure(op,
+ "unranked result type not supported");
}
- // Check that all the dimensions are 1.
- if (!llvm::all_of(tensorType.getShape(),
- [](int64_t dim) { return dim == 1; })) {
- return failure();
+ if (op.getNumOperands() == 1) {
+ rewriter.replaceOpWithNewOp<IREE::Flow::TensorSplatOp>(
+ op, tensorType, op.getOperand(0), ValueRange());
+ return success();
}
- rewriter.replaceOpWithNewOp<IREE::Flow::TensorSplatOp>(
- op, tensorType, op.getOperand(0), ValueRange());
+ const int64_t rank = tensorType.getRank();
+ Value result = rewriter.create<tensor::EmptyOp>(
+ op.getLoc(), tensorType.getShape(), tensorType.getElementType());
+ SmallVector<Value> ivs(rank);
+ for (int i = 0, s = op.getNumOperands(); i < s; ++i) {
+ int64_t index = i;
+ for (int j = rank - 1; j >= 0; --j) {
+ int64_t iv = index % tensorType.getDimSize(j);
+ index = index / tensorType.getDimSize(j);
+ ivs[j] = rewriter.create<arith::ConstantIndexOp>(op.getLoc(), iv);
+ }
+
+ result = rewriter.create<Flow::TensorStoreOp>(
+ op.getLoc(), op.getOperand(i), result, ivs);
+ }
+
+ rewriter.replaceOp(op, result);
return success();
}
};
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Conversion/TensorToFlow/test/from_elements.mlir b/compiler/src/iree/compiler/Dialect/Flow/Conversion/TensorToFlow/test/from_elements.mlir
index 13489f8..087d0e3 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Conversion/TensorToFlow/test/from_elements.mlir
+++ b/compiler/src/iree/compiler/Dialect/Flow/Conversion/TensorToFlow/test/from_elements.mlir
@@ -10,18 +10,6 @@
}
// -----
-// CHECK: util.func public @tensor.from_elements__not_convertible(%[[arg0:.*]]: i8)
-util.func public @tensor.from_elements__not_convertible(%arg0: i8) -> (i8) {
- // CHECK: %[[c0:.*]] = arith.constant 0
- %c0 = arith.constant 0 : index
- // CHECK: %[[res:.*]] = tensor.from_elements %[[arg0]], %[[arg0]] : tensor<2xi8>
- %0 = tensor.from_elements %arg0, %arg0 : tensor<2xi8>
- // CHECK: flow.tensor.load %[[res]][%[[c0]]]
- %1 = flow.tensor.load %0[%c0] : tensor<2xi8>
- util.return %1 : i8
-}
-
-// -----
util.func public @tensor.from_elements__within_dispatch_workgroups_not_converted() -> tensor<f32> {
%x = arith.constant 100 : index
%0 = flow.dispatch.workgroups[%x]() : () -> (tensor<f32>) = () {
@@ -44,3 +32,21 @@
// CHECK-SAME: %[[ARG0:.+]]: f32
// CHECK: %[[SPLAT:.+]] = flow.tensor.splat %[[ARG0]] : tensor<f32>
// CHECK: util.return %[[SPLAT]]
+
+// -----
+
+// CHECK-LABEL: util.func public @tensor.from_elements_2D
+util.func @tensor.from_elements_2D(%arg0 : f32, %arg1 : f32, %arg2 : f32, %arg3 : f32, %arg4 : f32, %arg5 : f32) -> tensor<2x3xf32> {
+ %0 = tensor.from_elements %arg0, %arg1, %arg2, %arg3, %arg4, %arg5 : tensor<2x3xf32>
+ // CHECK-DAG: %[[C2:.+]] = arith.constant 2 : index
+ // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
+ // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
+ // CHECK: %[[EMPTY:.+]] = tensor.empty() : tensor<2x3xf32>
+ // CHECK: %[[STORE0:.+]] = flow.tensor.store %arg0, %[[EMPTY]][%[[C0]], %[[C0]]] : tensor<2x3xf32>
+ // CHECK: %[[STORE1:.+]] = flow.tensor.store %arg1, %[[STORE0]][%[[C0]], %[[C1]]] : tensor<2x3xf32>
+ // CHECK: %[[STORE2:.+]] = flow.tensor.store %arg2, %[[STORE1]][%[[C0]], %[[C2]]] : tensor<2x3xf32>
+ // CHECK: %[[STORE3:.+]] = flow.tensor.store %arg3, %[[STORE2]][%[[C1]], %[[C0]]] : tensor<2x3xf32>
+ // CHECK: %[[STORE4:.+]] = flow.tensor.store %arg4, %[[STORE3]][%[[C1]], %[[C1]]] : tensor<2x3xf32>
+ // CHECK: %[[STORE5:.+]] = flow.tensor.store %arg5, %[[STORE4]][%[[C1]], %[[C2]]] : tensor<2x3xf32>
+ util.return %0 : tensor<2x3xf32>
+}
