colin
/
onnx-mlir
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
onnx-mlir/test/mlir/onnx/onnx_canonicalization.mlir

102 lines
7.7 KiB
MLIR
Raw Normal View History

[NFC] Change ONNF based names to ONNX-MLIR (#32) * Rename onnf to onnx-mlir. * Change workspace name.
2020-03-17 21:16:33 +08:00
// RUN: onnx-mlir-opt --canonicalize %s -split-input-file | FileCheck %s
[MLIR] compartmentalize build script (#369) * compartmentalize build script, temporarily remove dependency of onnf_opt on helper.cpp * fix test includes * fix op directory include * compiler -> op * compiler test depends on boost system * fix function name * specify libcompiler dependencies * let cmake take care of transitive dependencies * remove unnecessary includes * use ONNF_SRC_ROOT and ONNF_BIN_ROOT * allow whole-archive linked libraries to be appended * [MLIR] Support filecheck (#371) * support lit+FileCheck * add lit into build script * format MLIR.cmake * format cmake * [MLIR] Remove input/output ops (#372) * remove input/output ops * get output tensor type from symbol table
2019-11-19 08:37:58 +08:00
Add constraints for matmul-add fusion (#67) Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
2020-02-08 02:51:44 +08:00
// CHECK-LABEL: func @test_matmul_add_fused(%{{.*}}: tensor<10x10xf32>, %{{.*}}: tensor<10x10xf32>, %{{.*}}: tensor<10x10xf32>) -> tensor<10x10xf32> {
func @test_matmul_add_fused(%a0: tensor<10x10xf32>, %a1: tensor<10x10xf32>, %a2: tensor<10x10xf32>) -> tensor<10x10xf32> {
// CHECK-NEXT: %{{[0-9]+}} = "onnx.Gemm"(%{{.*}}, %{{.*}}, %{{.*}}) {alpha = 1.000000e+00 : f32, beta = 1.000000e+00 : f32, transA = 0 : i64, transB = 0 : i64} : (tensor<10x10xf32>, tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
[MLIR] Remove module from test (#379) * Remove module from test * Update onnx_canonicalization.mlir
2019-11-26 10:36:38 +08:00
%0 = "onnx.MatMul"(%a0, %a1) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
%1 = "onnx.Add"(%0, %a2) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
"std.return"(%1) : (tensor<10x10xf32>) -> ()
[MLIR] generate op from onnx document (#366) * generate op from onnx document * Restore FullGemm * update the op attribute for shape inference and canonicalizer * Update onnx_canonicalization.mlir
2019-11-19 10:08:21 +08:00
}
[MLIR] Canonicalization pattern for eliminating identity ops (#377) * Canonicalization pattern for eliminating identity ops * Add a test for the identity elimination rule * Remove frontend from test * Use CHECK-NEXT instead of CHECK
2019-11-21 10:57:13 +08:00
Add constraints for matmul-add fusion (#67) Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
2020-02-08 02:51:44 +08:00
// onnx.MatMul ops for non 2-D matrices should not get fused because Gemm only supports 2-D matrices.
// CHECK-LABEL: func @test_matmul_add_not_fused(%{{.*}}: tensor<10x10x10xf32>, %{{.*}}: tensor<10x10x10xf32>, %{{.*}}: tensor<10x10x10xf32>) -> tensor<10x10x10xf32> {
func @test_matmul_add_not_fused(%a0: tensor<10x10x10xf32>, %a1: tensor<10x10x10xf32>, %a2: tensor<10x10x10xf32>) -> tensor<10x10x10xf32> {
// CHECK-NEXT: %{{[0-9]+}} = "onnx.MatMul"(%{{.*}}, %{{.*}}) : (tensor<10x10x10xf32>, tensor<10x10x10xf32>) -> tensor<10x10x10xf32>
%0 = "onnx.MatMul"(%a0, %a1) : (tensor<10x10x10xf32>, tensor<10x10x10xf32>) -> tensor<10x10x10xf32>
%1 = "onnx.Add"(%0, %a2) : (tensor<10x10x10xf32>, tensor<10x10x10xf32>) -> tensor<10x10x10xf32>
"std.return"(%1) : (tensor<10x10x10xf32>) -> ()
}
// onnx.MatMul ops with more than one result uses should not get fused.
[MLIR] Add one more test case for MatMul-Add fusion (#380) * Add one more testcase for matmul-add fusion * Code format for identity elimination testcase
2019-11-26 10:49:48 +08:00
// CHECK-LABEL: func @test_sigmoid_add(%{{.*}}: tensor<10x10xf32>, %{{.*}}: tensor<10x10xf32>, %{{.*}}: tensor<10x10xf32>) -> tensor<10x10xf32>
func @test_sigmoid_add(%a0: tensor<10x10xf32>, %a1: tensor<10x10xf32>, %a2: tensor<10x10xf32>) -> tensor<10x10xf32> {
Add constraints for matmul-add fusion (#67) Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
2020-02-08 02:51:44 +08:00
// CHECK-NEXT: %{{[0-9]+}} = "onnx.MatMul"(%{{.*}}, %{{.*}}) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
[MLIR] Add one more test case for MatMul-Add fusion (#380) * Add one more testcase for matmul-add fusion * Code format for identity elimination testcase
2019-11-26 10:49:48 +08:00
%0 = "onnx.MatMul"(%a0, %a1) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
%1 = "onnx.Add"(%0, %a2) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
%2 = "onnx.Add"(%0, %a1) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
%3 = "onnx.Add"(%1, %2) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
"std.return"(%3) : (tensor<10x10xf32>) -> ()
}
// CHECK-LABEL: @test_identity_identity(%{{.*}}: tensor<10x10xf32>, %{{.*}}: tensor<10x10xf32>) -> tensor<10x10xf32>
[MLIR] Canonicalization pattern for eliminating identity ops (#377) * Canonicalization pattern for eliminating identity ops * Add a test for the identity elimination rule * Remove frontend from test * Use CHECK-NEXT instead of CHECK
2019-11-21 10:57:13 +08:00
func @test_identity_identity(%a0: tensor<10x10xf32>, %a1: tensor<10x10xf32>) -> tensor<10x10xf32> {
[MLIR] Add one more test case for MatMul-Add fusion (#380) * Add one more testcase for matmul-add fusion * Code format for identity elimination testcase
2019-11-26 10:49:48 +08:00
// CHECK-NEXT: %{{[0-9]+}} = "onnx.Add"(%{{.*}}, %{{.*}}) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
%0 = "onnx.Identity"(%a0) : (tensor<10x10xf32>) -> tensor<10x10xf32>
%1 = "onnx.Identity"(%a1) : (tensor<10x10xf32>) -> tensor<10x10xf32>
%2 = "onnx.Add"(%0, %1) : (tensor<10x10xf32>, tensor<10x10xf32>) -> tensor<10x10xf32>
"std.return"(%2) : (tensor<10x10xf32>) -> ()
[MLIR] Canonicalization pattern for eliminating identity ops (#377) * Canonicalization pattern for eliminating identity ops * Add a test for the identity elimination rule * Remove frontend from test * Use CHECK-NEXT instead of CHECK
2019-11-21 10:57:13 +08:00
}
Rewrite ReduceL1, ReduceL2, ReduceLogSum, ReduceLogSumExp, ReduceSumSquare in the ONNX dialect (#38) * Rewrite ReduceSumSquare * Edit gen_doc.py * Revise the code * Do shape inference after canonicalization so that there is no need to implement shape inference of rewritten ops * Rewrite ReduceL2 * Add onnx_rewrite.cpp for all rewriting for ONNX ops * Rewrite ReduceL1, ReduceLogSum, ReduceLogSumExp * Edit comments * Change the use of -> to . * Checkout gen_doc.py from the master branch * Use emplace_back instead of push_back * Revise the code * Edit comments Co-authored-by: Tian Jin <tjingrant@gmail.com>
2020-01-31 19:00:39 +08:00
Create constant pad (#75) * handle pad op which does not have the optional third argment * rewrite PadConstantValue with constant pad into PadConstantValuePad * add test for PadConstantValuePad * update onnx.md
2020-02-12 04:32:01 +08:00
// CHECK-LABEL: @test_constant_pad(%{{.*}}: tensor<?x?xf32>) -> tensor<*xf32> {
func @test_constant_pad(%arg0 : tensor<?x?xf32>) -> tensor<*xf32> {
shape inference for pad with constant pads
2020-02-14 01:08:29 +08:00
// CHECK-NEXT: [[SQUARE:%.+]] = "onnx.PadConstantValuePad"(%arg0) {constant_value = 0.000000e+00 : f32, mode = "constant", pads = [0, 2, 0, 0]} : (tensor<?x?xf32>) -> tensor<*xf32>
Create constant pad (#75) * handle pad op which does not have the optional third argment * rewrite PadConstantValue with constant pad into PadConstantValuePad * add test for PadConstantValuePad * update onnx.md
2020-02-12 04:32:01 +08:00
%0 ="onnx.Constant"() {value=[0, 2, 0, 0]} : ()-> tensor<?xi64>
%2 = "onnx.PadConstantValue"(%arg0, %0) {constant_value=0. : f32, mode = "constant"} : (tensor<?x?xf32>, tensor<?xi64>)-> tensor<*xf32>
"std.return"(%2) : (tensor<*xf32>) -> ()
}
Split convolution into explicit padding and unpaded convolution. (#82) * Split convolution into explicit padding and unpaded convolution. * Refactor code. Add test.
2020-02-15 05:06:38 +08:00
// CHECK-LABEL: @test_conv_split(%{{.*}}: tensor<1x9x32x64xf32>, %{{.*}}: tensor<5x9x6x7xf32>) -> tensor<*xf32> {
func @test_conv_split(%arg0 : tensor<1x9x32x64xf32>, %arg1 : tensor<5x9x6x7xf32>) -> tensor<*xf32> {
Unify Conv implementation (#54) * fixed readme for new git repo * conv with bias as an optional input
2020-03-26 23:03:19 +08:00
%cst = constant unit
%0 = "onnx.Conv"(%arg0, %arg1, %cst) {auto_pad = "NOTSET", group = 1 : i64, pads = [2, 3, 4, 5]} : (tensor<1x9x32x64xf32>, tensor<5x9x6x7xf32>, none) -> tensor<*xf32>
Split convolution into explicit padding and unpaded convolution. (#82) * Split convolution into explicit padding and unpaded convolution. * Refactor code. Add test.
2020-02-15 05:06:38 +08:00
"std.return"(%0) : (tensor<*xf32>) -> ()
Unify Conv implementation (#54) * fixed readme for new git repo * conv with bias as an optional input
2020-03-26 23:03:19 +08:00
// CHECK-NEXT: %cst = constant unit
fix test case for changes in Op definition
2020-02-17 22:07:58 +08:00
// CHECK-NEXT: %0 = "onnx.PadConstantValuePad"(%arg0) {constant_value = 0.000000e+00 : f32, mode = "constant", pads = [0, 0, 2, 3, 0, 0, 4, 5]} : (tensor<1x9x32x64xf32>) -> tensor<1x9x38x72xf32>
Unify Conv implementation (#54) * fixed readme for new git repo * conv with bias as an optional input
2020-03-26 23:03:19 +08:00
// CHECK-NEXT: %1 = "onnx.Conv"(%0, %arg1, %cst) {auto_pad = "NOTSET", group = 1 : i64, pads = [0, 0, 0, 0]} : (tensor<1x9x38x72xf32>, tensor<5x9x6x7xf32>, none) -> tensor<*xf32>
Split convolution into explicit padding and unpaded convolution. (#82) * Split convolution into explicit padding and unpaded convolution. * Refactor code. Add test.
2020-02-15 05:06:38 +08:00
// CHECK-NEXT: return %1 : tensor<*xf32>
}
Support Optional Inputs (#94) * 1. Combine variadicIn/Out with expectedNumOperands/Results to simplify import function arguments. 2. Generic improvements to code readability in gen_doc.py. * Update ONNX Dialect doc. * Remove redundant code in ImportNode. * Prettify op_build_table.inc. * 1. Remove irrelevant code in gen_doc.py * Refactor code to be more readable. * Further refactoring for readability improvements. * Allow gemm to have an optional operand (bias term), and include an example of declarative optimization pattern targeting gemm with bias term ommitted. * Make shape inference/lowering of gemm op compatible with optional operand declaration. * Apply canonicalization again after lowering from onnx -> std dialects. * Make hasBias compatible with the situation of GemmNoBias op. * Update doc. * Add a canonicalization test. * Remove special handler for importing Gemm op, as it's redundant now.
2020-02-24 23:46:48 +08:00
//CHECK-LABEL: @test_gemm_add_fusion(%{{.*}}: tensor<128x128xf32>, %{{.*}}: tensor<128x128xf32>, %{{.*}}: tensor<128xf32>) -> tensor<*xf32> {
func @test_gemm_add_fusion(%arg0: tensor<128x128xf32>, %arg1: tensor<128x128xf32>, %arg2: tensor<128xf32>) -> tensor<*xf32> {
%cst = constant unit
%0 = "onnx.Gemm"(%arg0, %arg1, %cst) : (tensor<128x128xf32>, tensor<128x128xf32>, none) -> tensor<*xf32>
%1 = "onnx.Add"(%0, %arg2) : (tensor<*xf32>, tensor<128xf32>) -> tensor<*xf32>
return %1 : tensor<*xf32>
// CHECK-NEXT: [[GEMM:%.+]] = "onnx.Gemm"(%{{.*}}, %{{.*}}, %{{.*}}) {alpha = 1.000000e+00 : f32, beta = 1.000000e+00 : f32, transA = 0 : i64, transB = 0 : i64} : (tensor<128x128xf32>, tensor<128x128xf32>, tensor<128xf32>) -> tensor<*xf32>
// return [[GEMM]] : tensor<*xf32>
Remove rank constraints in gemm fusion (#101) * Remove rank constraints in gemm fusion * Add an MLIR test Co-authored-by: Tian Jin <tjingrant@gmail.com> Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
2020-02-27 00:40:52 +08:00
}
//CHECK-LABEL: @test_gemm_add_fusion_rank3(%{{.*}}: tensor<128x128x256xf32>, %{{.*}}: tensor<128x128x256xf32>, %{{.*}}: tensor<256xf32>) -> tensor<*xf32> {
func @test_gemm_add_fusion_rank3(%arg0: tensor<128x128x256xf32>, %arg1: tensor<128x128x256xf32>, %arg2: tensor<256xf32>) -> tensor<*xf32> {
%cst = constant unit
%0 = "onnx.Gemm"(%arg0, %arg1, %cst) : (tensor<128x128x256xf32>, tensor<128x128x256xf32>, none) -> tensor<*xf32>
%1 = "onnx.Add"(%0, %arg2) : (tensor<*xf32>, tensor<256xf32>) -> tensor<*xf32>
return %1 : tensor<*xf32>
// CHECK-NEXT: [[GEMM:%.+]] = "onnx.Gemm"(%{{.*}}, %{{.*}}, %{{.*}}) {alpha = 1.000000e+00 : f32, beta = 1.000000e+00 : f32, transA = 0 : i64, transB = 0 : i64} : (tensor<128x128x256xf32>, tensor<128x128x256xf32>, tensor<256xf32>) -> tensor<*xf32>
// return [[GEMM]] : tensor<*xf32>
}
Fix issues #15 and #16 (#29) * fix issue #15 and #16 * fix format Co-authored-by: Gheorghe-Teodor Bercea <gt.bercea@gmail.com>
2020-03-13 22:19:27 +08:00
//CHECK-LABEL: @test_maxpoolsingleout_split(%{{.*}}: tensor<5x5x32x32xf32>) -> tensor<5x5x36x38xf32> {
func @test_maxpoolsingleout_split(%arg0: tensor<5x5x32x32xf32>) -> tensor<5x5x36x38xf32> {
%0 = "onnx.MaxPoolSingleOut"(%arg0) {auto_pad = "NOTSET", ceil_mode = 0, kernel_shape = [5,3], pads = [1, 2, 3, 4] } : (tensor<5x5x32x32xf32>) -> tensor<5x5x36x38xf32>
"std.return"(%0) : (tensor<5x5x36x38xf32>) -> ()
// CHECK-NEXT: %0 = "onnx.PadConstantValuePad"(%arg0) {constant_value = 0xFF800000 : f32, mode = "constant", pads = [0, 0, 1, 2, 0, 0, 3, 4]} : (tensor<5x5x32x32xf32>) -> tensor<5x5x36x38xf32>
// CHECK-NEXT: %1 = "onnx.MaxPoolSingleOut"(%0) {auto_pad = "NOTSET", ceil_mode = 0 : i64, kernel_shape = [5, 3], pads = [0, 0, 0, 0], storage_order = 0 : i64} : (tensor<5x5x36x38xf32>) -> tensor<5x5x36x38xf32>
// CHECK-NEXT: return %1 : tensor<5x5x36x38xf32>
Support Pads for MaxPoolSingleOut (#14) * Support Pads for MaxPoolSingleOut * Regenerate onnx.md to include the new op * Edit comments * Undo redundant parts that were unintentionally changed * Move declarative rewriting rules into canonicalize to avoid creating a new op * Reformat the rewriting rule pattern of MaxPoolSingleOut * Put ONNXPadConstantValuePadOp's build method into a .cpp file instead of a tablegen file * Use the same helper function as the one in inferShape for the ONNXPadConstantValuePadOp's build method * Change function names and fix padding for the spatial dimensions * Call shape-inference again after canonicalization to infer shape for newly added ops during canonicalization. * Fix typos
2020-03-10 08:15:58 +08:00
}
//CHECK-LABEL: @test_maxpoolsingleout_split_unknown_dims(%{{.*}}: tensor<*xf32>) -> tensor<*xf32> {
func @test_maxpoolsingleout_split_unknown_dims(%arg0: tensor<*xf32>) -> tensor<*xf32> {
%0 = "onnx.MaxPoolSingleOut"(%arg0) {auto_pad = "NOTSET", ceil_mode = 0, kernel_shape = [5,3], pads = [1, 2, 3, 4] } : (tensor<*xf32>) -> tensor<*xf32>
"std.return"(%0) : (tensor<*xf32>) -> ()
Pad value for MaxPool must be negative infinity instead of zero (#20) Co-authored-by: Alexandre Eichenberger <alexe@us.ibm.com>
2020-03-12 21:30:02 +08:00
// CHECK-NEXT: %0 = "onnx.PadConstantValuePad"(%arg0) {constant_value = 0xFF800000 : f32, mode = "constant", pads = [0, 0, 1, 2, 0, 0, 3, 4]} : (tensor<*xf32>) -> tensor<*xf32>
Support Pads for MaxPoolSingleOut (#14) * Support Pads for MaxPoolSingleOut * Regenerate onnx.md to include the new op * Edit comments * Undo redundant parts that were unintentionally changed * Move declarative rewriting rules into canonicalize to avoid creating a new op * Reformat the rewriting rule pattern of MaxPoolSingleOut * Put ONNXPadConstantValuePadOp's build method into a .cpp file instead of a tablegen file * Use the same helper function as the one in inferShape for the ONNXPadConstantValuePadOp's build method * Change function names and fix padding for the spatial dimensions * Call shape-inference again after canonicalization to infer shape for newly added ops during canonicalization. * Fix typos
2020-03-10 08:15:58 +08:00
// CHECK-NEXT: %1 = "onnx.MaxPoolSingleOut"(%0) {auto_pad = "NOTSET", ceil_mode = 0 : i64, kernel_shape = [5, 3], pads = [0, 0, 0, 0], storage_order = 0 : i64} : (tensor<*xf32>) -> tensor<*xf32>
// CHECK-NEXT: return %1 : tensor<*xf32>
}