INFRT/add LLVM lit to infrt (#39149)

.gitignore

@@ -39,3 +39,10 @@ model_test
 
 Testing
 tools/__pycache__
+
+# This file is automatically generated.
+# TODO(zhiqiang) Move this file to build directory.
+paddle/infrt/dialect/pd_ops.td
+.lit_test_times.txt
+paddle/infrt/tests/dialect/Output
+paddle/infrt/tests/lit.cfg.py

paddle/infrt/CMakeLists.txt

@@ -76,6 +76,7 @@ add_subdirectory(tensor)
 add_subdirectory(support)
 add_subdirectory(external_kernels)
 add_subdirectory(paddle)
+add_subdirectory(tests)
 
 
 # MLIR td file generations
@@ -86,13 +87,14 @@ set(infrt_mlir_incs
         tensor_shape_inc
         dense_tensor_inc
         pd_ops_inc
+        pd_extra_ops_inc
         rewrite_inc
         trt_ops_inc
         )
 message(STATUS "infrt srcs:\n${infrt_src}")
 
-cc_library(infrt SHARED SRCS ${infrt_src} DEPS glog boost ${mlir_libs} paddle_framework_proto pten dense_tensor)
-cc_library(infrt_static SRCS ${infrt_src} DEPS glog boost ${mlir_libs} paddle_framework_proto pten dense_tensor)
+cc_library(infrt SHARED SRCS ${infrt_src} DEPS glog boost ${mlir_libs} paddle_framework_proto)
+cc_library(infrt_static SRCS ${infrt_src} DEPS glog boost ${mlir_libs} paddle_framework_proto)
 add_dependencies(infrt ${infrt_mlir_incs})
 
 add_custom_target(test_infrt_exec DEPENDS ${INFRT_TEST_TARGETS})

paddle/infrt/dialect/CMakeLists.txt

@@ -20,6 +20,7 @@ mlir_tablegen_on(tensor_shape DIALECT ts)
 mlir_tablegen_on(dense_tensor DIALECT dt)
 mlir_tablegen_on(pd_op_base DIALECT pd)
 mlir_tablegen_on(pd_ops)
+mlir_tablegen_on(pd_extra_ops)
 mlir_add_rewriter(rewrite)
 
 # TODO(Superjomn) add a cmake function cc_executable to ecapsulate the following code
@@ -30,22 +31,6 @@ add_executable(print-ir print_ir.cc)
 target_link_libraries(print-ir infrt ${mlir_libs})
 add_dependencies(print-ir pd_ops_inc)
 
-
-# MLIR opt tests
-# %{
-set(infrt_opt_path ${CMAKE_CURRENT_BINARY_DIR}/infrtopt)
-
-add_test(test_infrt_mlir_opt_on_basic ${infrt_opt_path} ${CMAKE_CURRENT_SOURCE_DIR}/mlir_tests/basic.mlir)
-add_test(test_infrt_mlir_opt_on_tensor_shape ${infrt_opt_path} ${CMAKE_CURRENT_SOURCE_DIR}/mlir_tests/tensor_shape.mlir)
-add_test(test_infrt_mlir_opt_on_paddle_op ${infrt_opt_path} ${CMAKE_CURRENT_SOURCE_DIR}/mlir_tests/paddle_ops.mlir)
-# %}
-
 cc_test_tiny(test_infrt_mlir_loader SRCS mlir_loader_test.cc DEPS infrt ${MLIR_IR_LIBS})
 
-# execute mlir and run FileCheck
-infrt_exec_check(test_infrt_tensor_type mlir_tests/tensor_type.mlir)
-infrt_exec_check(test_infrt__basic mlir_tests/basic.mlir)
-infrt_exec_check(test_infrt_benchmark mlir_tests/benchmark.mlir)
-infrt_exec_check(test_infrt_mlir_dense_tensor mlir_tests/dense_tensor.mlir)
-
 add_subdirectory(tensorrt)

paddle/infrt/dialect/pd_extra_ops.td

+#ifndef PD_EXTRA_OPS
+#define PD_EXTRA_OPS
+
+include "mlir/Interfaces/InferTypeOpInterface.td"
+include "mlir/Interfaces/LoopLikeInterface.td"
+include "mlir/IR/OpBase.td"
+include "paddle/infrt/dialect/pd_op_base.td"
+
+def PD_FusedFC : PD_Op<"FC", [NoSideEffect]> {
+    let summary = "Computes the Fully Connected result of two tensors";
+    let description = [{
+    }];
+
+    let arguments = (ins PD_Tensor:$input, PD_Tensor:$w, PD_Tensor:$bias, DefaultValuedAttr<I32Attr, "1">:$in_num_col_dims);
+    let results = (outs PD_Tensor:$out);
+}
+
+#endif

paddle/infrt/dialect/pd_ops.cc

@@ -20,6 +20,8 @@
 
 #define GET_OP_CLASSES
 #include "paddle/infrt/dialect/pd_ops.cpp.inc"  // NOLINT
+#define GET_OP_CLASSES
+#include "paddle/infrt/dialect/pd_extra_ops.cpp.inc"  // NOLINT
 
 #include "paddle/infrt/dialect/rewrite.hpp.inc"  // NOLINT
 
@@ -31,8 +33,10 @@ PaddleDialect::PaddleDialect(MLIRContext *context)
   addOperations<
 #define GET_OP_LIST
 #include "paddle/infrt/dialect/pd_ops.cpp.inc"  // NOLINT
+      ,
+#define GET_OP_LIST
+#include "paddle/infrt/dialect/pd_extra_ops.cpp.inc"  // NOLINT
       >();
-#undef GET_OP_LIST
 }
 
 mlir::Operation *PaddleDialect::materializeConstant(mlir::OpBuilder &builder,
@@ -81,11 +85,14 @@ LogicalResult ElementwiseAdd::inferReturnTypes(
   inferredReturnTypes.push_back(operands[0].getType());
   return success();
 }
-void ElementwiseAdd::getCanonicalizationPatterns(
+*/
+
+void Elementwise_addOp::getCanonicalizationPatterns(
     mlir::OwningRewritePatternList &results, mlir::MLIRContext *context) {
   results.insert<FuseMulAdd>(context);
 }
 
+/*
 mlir::OpFoldResult ElementwiseAdd::fold(
     llvm::ArrayRef<mlir::Attribute> operands) {
   if (getElementTypeOrSelf(getType()).isa<FloatType>()) {

paddle/infrt/dialect/pd_ops.h

@@ -57,3 +57,5 @@ class PaddleDialect : public Dialect {
 
 #define GET_OP_CLASSES
 #include "paddle/infrt/dialect/pd_ops.hpp.inc"
+#define GET_OP_CLASSES
+#include "paddle/infrt/dialect/pd_extra_ops.hpp.inc"

paddle/infrt/dialect/rewrite.td

@@ -4,7 +4,8 @@
 include "paddle/infrt/dialect/infrt_base.td"
 include "mlir/Interfaces/SideEffectInterfaces.td"
 include "paddle/infrt/dialect/pd_ops.td"
-/*
+include "paddle/infrt/dialect/pd_extra_ops.td"
+
 //===----------------------------------------------------------------------===//
 // This is to fuse the composition: 'Matmul o ElementwiseAdd' into 'PD_FusedFC'.
 //
@@ -23,10 +24,11 @@ include "paddle/infrt/dialect/pd_ops.td"
 //  1. Make the constrait more completely.
 //  2. Consider the case of : out = bias + z
 //===----------------------------------------------------------------------===//
-def FuseMulAdd : Pat<(PD_ElementwiseAdd (PD_MatmulOp $x, $y, ConstBoolAttrFalse:$_, ConstBoolAttrFalse:$_, $alpha), $bias, $axis),
+def FuseMulAdd : Pat<(PD_Elementwise_addOp (PD_Matmul_v2Op $x, $y, ConstBoolAttrFalse:$_, ConstBoolAttrFalse:$_), $bias, $axis),
                      (PD_FusedFC $x, $y, $bias, (INFRT_createI32Attr<"1">))>;
 
 
+/*
 //===----------------------------------------------------------------------===//
 // This is to fuse the composition: 'FusedFC o Relu' into 'FusedRepeatedFCRelu'.
 //

paddle/infrt/external_kernels/CMakeLists.txt

@@ -9,5 +9,5 @@ message(STATUS "basic_mlir: ${basic_mlir}")
 message(STATUS "external_kernels_lib: ${external_kernels_lib}")
 add_test(
     NAME run_and_check_external_kernels
-    COMMAND sh -c "${CMAKE_BINARY_DIR}/infrt/host_context/infrt-exec -i ${basic_mlir} --shared_libs=${external_kernels_lib} | ${LLVM_PATH}/bin/FileCheck ${basic_mlir}"
+    COMMAND sh -c "${CMAKE_BINARY_DIR}/infrt/host_context/infrtexec -i ${basic_mlir} --shared_libs=${external_kernels_lib} | ${LLVM_PATH}/bin/FileCheck ${basic_mlir}"
 )

paddle/infrt/host_context/CMakeLists.txt

@@ -21,9 +21,5 @@ cc_test_tiny(test_infrt_op_executable SRCS op_executable_test.cc DEPS infrt ${ML
 cc_test_tiny(test_infrt_core_runtime SRCS core_runtime_test.cc DEPS infrt ${MLIR_IR_LIBS})
 cc_test_tiny(test_infrt_mlir_to_runtime_translate SRCS mlir_to_runtime_translate_test.cc DEPS infrt ${MLIR_IR_LIBS})
 
-infrt_exec_check(test_infrt_mlir_exec_on_basic mlir_tests/basic.mlir)
-infrt_exec_check(test_infrt_mlir_exec_on_shape mlir_tests/shape.mlir)
-infrt_exec_check(test_infrt_mlir_exec_on_dense_tensor mlir_tests/dense_tensor.mlir)
-
-add_executable(infrt-exec mlir_exec.cc)
-target_link_libraries(infrt-exec infrt ${MLIR_IR_LIBS})
+add_executable(infrtexec mlir_exec.cc)
+target_link_libraries(infrtexec infrt ${MLIR_IR_LIBS})

paddle/infrt/host_context/value.h

@@ -29,9 +29,6 @@
 #include "paddle/infrt/tensor/tensor_map.h"
 #include "paddle/infrt/tensor/tensor_shape.h"
 
-#include "paddle/pten/backends/cpu/cpu_context.h"
-#include "paddle/pten/core/dense_tensor.h"
-
 namespace infrt {
 namespace host_context {
 
@@ -48,8 +45,8 @@ using ValueVariantType = Variant<int16_t,
                                  tensor::DenseHostTensor,
                                  MlirFunctionExecutable*,
                                  tensor::TensorMap,
-                                 pten::CPUContext,
-                                 pten::DenseTensor,
+                                 // pten::CPUContext,
+                                 // pten::DenseTensor,
                                  std::vector<int16_t>,
                                  std::vector<int32_t>,
                                  std::vector<int64_t>,

paddle/infrt/kernel/CMakeLists.txt

@@ -2,7 +2,7 @@ core_gather_headers()
 
 gather_srcs(infrt_src SRCS
     basic_kernels.cc
-    pten_kernels.cc
+    # pten_kernels.cc
     test_kernels.cc
     tensor_shape_kernels.cc
     tensor_kernels.cc

paddle/infrt/kernel/pten_kernels.cc

@@ -12,14 +12,17 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include "paddle/infrt/kernel/pten_kernels.h"
+
 #include <iostream>
 #include <string>
 
 #include "paddle/infrt/host_context/kernel_registry.h"
 #include "paddle/infrt/host_context/kernel_utils.h"
-#include "paddle/infrt/kernel/pten_kernels.h"
-#include "paddle/pten/backends/cpu/cpu_context.h"
-#include "paddle/pten/kernels/math_kernel.h"
+
+// Disable temporarily.
+// #include "paddle/pten/backends/cpu/cpu_context.h"
+// #include "paddle/pten/kernels/math_kernel.h"
 
 using infrt::host_context::Attribute;
 

paddle/infrt/tests/CMakeLists.txt

+configure_file(lit.cfg.py.in "${CMAKE_SOURCE_DIR}/paddle/infrt/tests/lit.cfg.py")
+
+add_test(NAME test_infrt_by_lit COMMAND sh -c "lit -v ${CMAKE_SOURCE_DIR}/paddle/infrt/tests --filter-out \"disabled_*\""
+    DEPENDS infrtopt infrtexec)

paddle/infrt/dialect/mlir_tests/basic.mlir → paddle/infrt/tests/dialect/basic.mlir

+// RUN: infrtexec -i %s | FileCheck %s
 // CHECK-LABEL: @basic_f32
 func @basic_f32() -> f32 {
   %v0 = infrt.constant.f32 1.0

paddle/infrt/dialect/mlir_tests/benchmark.mlir → paddle/infrt/tests/dialect/benchmark.mlir

+// RUN: infrtexec -i %s | FileCheck %s
 // CHECK-LABEL: @benchmark
 func @benchmark() {
   // CHECK-LABEL: BM:add.f32:Count: 3

paddle/infrt/dialect/mlir_tests/dense_tensor.mlir → paddle/infrt/tests/dialect/dense_tensor.mlir

+// RUN: infrtexec -i %s | FileCheck %s
 // CHECK-LABEL: dense_shape0
 func @dense_shape0() {
   %shape = ts.build_shape [1:i64, 57:i64]

paddle/infrt/dialect/mlir_tests/paddle_ops.mlir → paddle/infrt/tests/dialect/paddle_ops.mlir

+// RUN: infrtopt %s | FileCheck %s
+// CHECK-LABEL: @ops
 func @ops() {
   %a = pd.feed() {name="input0"} : tensor<?xf32>
   %b = pd.feed() {name="input1"}: tensor<?xf32>

paddle/infrt/dialect/mlir_tests/rewrite.mlir → paddle/infrt/tests/dialect/rewrite.mlir

+// RUN: infrtopt --canonicalize %s | FileCheck %s
 // CHECK-LABEL: @main
 func @main() -> tensor<?xf32> {
   %a = "pd.feed"() {name="input0"} : () -> tensor<?xf32>
@@ -9,16 +10,19 @@ func @main() -> tensor<?xf32> {
   %bias1 = "pd.feed"() {name="input5"} : () -> tensor<?xf32>
   %bias2 = "pd.feed"() {name="input6"} : () -> tensor<?xf32>
 
-  %c = "pd.matmul"(%a, %b) {transpose_y=false} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
-  %d = "pd.elementwise_add"(%c, %bias) {axis=1:i32} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  // CHECK: %{{[0-9]+}} = "pd.FC"(%{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}) {in_num_col_dims = 1 : i32} : (tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  %c = "pd.matmul_v2"(%a, %b) {transpose_y=false} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  %d = "pd.elementwise_add"(%c, %bias) {axis=1:si32} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
   %e = "pd.relu6"(%d) {} : (tensor<?xf32>) -> tensor<?xf32>
 
-  %c1 = "pd.matmul"(%e, %b1) {transpose_x=false, transpose_y=false} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
-  %d1 = "pd.elementwise_add"(%c1, %bias1) {axis=1:i32} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  // CHECK: %{{[0-9]+}} = "pd.FC"(%{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}) {in_num_col_dims = 1 : i32} : (tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  %c1 = "pd.matmul_v2"(%e, %b1) {transpose_x=false, transpose_y=false} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  %d1 = "pd.elementwise_add"(%c1, %bias1) {axis=1:si32} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
   %e1 = "pd.relu"(%d1) {} : (tensor<?xf32>) -> tensor<?xf32>
 
-  %c2 = "pd.matmul"(%e1, %b2) {transpose_x=true, transpose_y=false} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
-  %d2 = "pd.elementwise_add"(%c2, %bias2) {axis=1:i32} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  // CHECK: %{{[0-9]+}} = "pd.FC"(%{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}) {in_num_col_dims = 1 : i32} : (tensor<?xf32>, tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  %c2 = "pd.matmul_v2"(%e1, %b2) {transpose_x=true, transpose_y=false} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
+  %d2 = "pd.elementwise_add"(%c2, %bias2) {axis=1:si32} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
   %e2 = "pd.relu"(%d2) {} : (tensor<?xf32>) -> tensor<?xf32>
   "pd.fetch"(%e2) {name="output"} :(tensor<?xf32>)->()
 }

paddle/infrt/dialect/mlir_tests/tensor_shape.mlir → paddle/infrt/tests/dialect/tensor_shape.mlir

+// RUN: infrtexec -i %s | FileCheck %s
+// CHECK-LABEL: @build_tensor1
 func @build_tensor1() {
   %a = ts.build_shape [1:i64, 57:i64, 92:i64]
+  // CHECK: shape[1,57,92]
   ts.print_shape %a
   infrt.return
 }

paddle/infrt/dialect/mlir_tests/tensor_type.mlir → paddle/infrt/tests/dialect/tensor_type.mlir

+// RUN: infrtexec -i %s | FileCheck %s
 // CHECK-LABEL: test_tensor_type
 func @test_tensor_type() {
   %a = dt.create_uninit_tensor.f32 [3, 4] -> !infrt.tensor<X86, NCHW, F32>

paddle/infrt/tests/lit.cfg.py.in

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import lit.formats
+import os
+
+config.name = "MLIR tests"
+config.test_format = lit.formats.ShTest(True)
+build_dir = "@CMAKE_BINARY_DIR@"
+config.llvm_tools_dir = os.path.join(build_dir, "third_party/install/llvm/bin")
+config.llvm_tools_dir = os.path.join(build_dir, "/third_party/install/llvm/lib")
+infrtopt_bin = os.path.join(build_dir, "paddle/infrt/dialect/")
+infrtexec_bin = os.path.join(build_dir, "paddle/infrt/host_context/")
+
+llvm_bin = os.path.join(build_dir, "third_party/install/llvm/bin/")
+config.environment['PATH'] = os.path.pathsep.join(
+    (infrtopt_bin, infrtexec_bin, llvm_bin, config.environment['PATH']))
+
+config.suffixes = ['.mlir']

paddle/scripts/infrt_build.sh

@@ -32,8 +32,8 @@ function update_pd_ops() {
    # compile and install paddle
    rm -rf ${PADDLE_ROOT}/build && mkdir -p ${PADDLE_ROOT}/build
    cd ${PADDLE_ROOT}/build
-   cmake .. -DWITH_PYTHON=ON -DWITH_GPU=OFF -DPYTHON_EXECUTABLE=`which python3`
-   make -j8
+   cmake .. -DWITH_PYTHON=ON -DWITH_GPU=OFF -DPYTHON_EXECUTABLE=`which python3` -DWITH_XBYAK=OFF -DWITH_NCCL=OFF -DWITH_RCCL=OFF -DWITH_CRYPTO=OFF
+   make -j8 paddle_python
    cd ${PADDLE_ROOT}/build
    cd python/dist/
    python3 -m pip uninstall -y paddlepaddle
@@ -90,7 +90,7 @@ function infrt_gen_and_build() {
         exit 7;
     fi
 
-    make -j ${parallel_number} infrt infrtopt infrt-exec test_infrt_exec trt-exec infrt_lib_dist;build_error=$?
+    make -j ${parallel_number} infrt infrtopt infrtexec test_infrt_exec trt-exec infrt_lib_dist;build_error=$?
     if [ "$build_error" != 0 ];then
         exit 7;
     fi
@@ -101,6 +101,9 @@ function infrt_gen_and_build() {
 }
 
 function test_infrt() {
+    # install llvm-lit toolkit
+    python3 -m pip install lit
+
     mkdir -p ${PADDLE_ROOT}/build
     cd ${PADDLE_ROOT}/build
     if [ ${WITH_TESTING:-ON} == "ON" ] ; then

tools/infrt/generate_pd_op_dialect_from_paddle_op_maker.py

@@ -16,6 +16,8 @@ import paddle.fluid.framework as framework
 from paddle.fluid import core
 from paddle import compat as cpt
 
+ops_having_canonicalization = {"elementwise_add", }
+
 
 # collect original ops: op which has both inference and grid defination
 def get_original_ops():
@@ -120,7 +122,18 @@ def convert_op_proto_into_mlir(op_descs):
 |*                                                                            *|\n\
 \*===----------------------------------------------------------------------===*/\n"
 
-    start_ = comment_ + "#ifndef PD_OPS\n#define PD_OPS\ninclude \"mlir/Interfaces/InferTypeOpInterface.td\"\ninclude \"mlir/Interfaces/LoopLikeInterface.td\"\ninclude \"mlir/IR/OpBase.td\"\ninclude \"paddle/infrt/dialect/pd_op_base.td\"\n\n"
+    lines = [
+        "#ifndef PD_OPS",
+        "#define PD_OPS",
+        "include \"mlir/Interfaces/InferTypeOpInterface.td\"",
+        "include \"mlir/Interfaces/LoopLikeInterface.td\"",
+        "include \"mlir/IR/OpBase.td\"",
+        "include \"paddle/infrt/dialect/pd_op_base.td\"",
+        "",
+    ]
+
+    start_ = comment_ + "\n".join(lines)
+
     with open(dst_dialect_file, 'w') as ops_mlir_file:
         ops_mlir_file.write(start_)
 
@@ -134,6 +147,7 @@ def convert_op_proto_into_mlir(op_descs):
         "trainable_statistics", "use_global_stats", "is_test", "use_mkldnn",
         "use_cudnn"
     ]
+
     original_ops_ = get_original_ops()
     automatically_generated_op_dialect = []
     for op_type, op_proto in op_descs.items():
@@ -144,6 +158,7 @@ def convert_op_proto_into_mlir(op_descs):
         HEAD = "def PD_" + op_type.capitalize(
         ) + "Op : PD_Op<\"" + op_type + "\", [NoSideEffect]> {\n"
         SUMMARY = "  let summary = \"" + op_type + " op\";\n"
+        CANONICALIZATION = "let hasCanonicalizer = 1;" if op_type in ops_having_canonicalization else ""
 
         # 2.2 Description
         DESCRIPTION = "  let description = [{\n"
@@ -245,6 +260,7 @@ def convert_op_proto_into_mlir(op_descs):
             ops_mlir_file.write(DESCRIPTION)
             ops_mlir_file.write(ARGUMENTS)
             ops_mlir_file.write(RESULTS)
+            ops_mlir_file.write(CANONICALIZATION)
             ops_mlir_file.write("}\n")
 
     print("Skipped ops num: " + str(len(skipped_op_list)))