Infrt registers pten kernels (#39588)

* the mlir representation of pten, test=develop

* fixes an error, test=develop

* infrt registers pten kernels
Co-authored-by: NShixiaowei02 <39303645+Shixiaowei02@users.noreply.github.com>

.gitignore

@@ -52,5 +52,6 @@ paddle/infrt/dialect/pd_ops_info.h
 .lit_test_times.txt
 paddle/infrt/tests/dialect/Output
 paddle/infrt/tests/lit.cfg.py
+paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launchers.cc
 paddle/fluid/pybind/eager_final_state_op_function_impl.h
 paddle/fluid/pybind/tmp_eager_final_state_op_function_impl.h

paddle/infrt/host_context/value.h

@@ -28,13 +28,18 @@
 #include "paddle/infrt/tensor/dense_tensor_view.h"
 #include "paddle/infrt/tensor/tensor_map.h"
 #include "paddle/infrt/tensor/tensor_shape.h"
-#include "paddle/pten/core/meta_tensor.h"
 
 #ifdef INFRT_WITH_PTEN
 #include "paddle/infrt/backends/host/pten_allocator.h"
 #include "paddle/infrt/backends/host/pten_context.h"
-#include "paddle/pten/backends/cpu/cpu_context.h"
+#include "paddle/pten/backends/all_context.h"
+#include "paddle/pten/common/backend.h"
+#include "paddle/pten/common/data_type.h"
+#include "paddle/pten/common/layout.h"
+#include "paddle/pten/common/scalar.h"
+#include "paddle/pten/common/scalar_array.h"
 #include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/meta_tensor.h"
 #endif
 
 namespace infrt {
@@ -42,7 +47,8 @@ namespace host_context {
 
 struct MlirFunctionExecutable;
 
-using ValueVariantType = Variant<int16_t,
+using ValueVariantType =
+    Variant<int16_t,
             int32_t,
             int64_t,
             float,
@@ -61,6 +67,14 @@ using ValueVariantType = Variant<int16_t,
             backends::CpuPtenAllocator,
             backends::CpuPtenContext,
             ::pten::CPUContext,
+            std::vector<pten::DenseTensor>,
+            paddle::experimental::ScalarBase<pten::DenseTensor>,
+            paddle::experimental::ScalarArrayBase<pten::DenseTensor>,
+            std::vector<pten::MetaTensor>,
+            pten::MetaConfig,
+            paddle::experimental::Backend,
+            paddle::experimental::DataLayout,
+            paddle::experimental::DataType,
 #endif
             std::vector<int16_t>,
             std::vector<int32_t>,

paddle/infrt/kernel/pten/CMakeLists.txt

@@ -11,9 +11,21 @@ gather_srcs(infrt_src SRCS
     allocator_kernels.cc
 )
 
+set(infrt_register_pten_kernels_gen_source_file ${CMAKE_SOURCE_DIR}/paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launchers.cc)
+set(infrt_register_pten_kernels_gen_file ${CMAKE_SOURCE_DIR}/tools/infrt/get_pten_kernel_function.sh)
+set(wrapped_infermeta_header_file ${CMAKE_SOURCE_DIR}/paddle/pten/infermeta/generated.h)
+set(wrapped_infermeta_source_file ${CMAKE_SOURCE_DIR}/paddle/pten/infermeta/generated.cc)
+
+add_custom_command(
+        OUTPUT ${infrt_register_pten_kernels_gen_source_file}
+        COMMAND sh ${infrt_register_pten_kernels_gen_file}
+        DEPENDS ${wrapped_infermeta_header_file} ${wrapped_infermeta_source_file}
+        COMMENT "infrt generate ${infrt_register_pten_kernels_gen_source_file}"
+        VERBATIM)
+
 cc_library(infrt_naive SRCS infershaped/infershaped_kernel_launcher.cc
         infershaped/infershaped_kernel_launchers.cc
-        )
+        DEPS pten wrapped_infermeta)
 
 cc_test_tiny(test_infrt_infershape_launchers SRCS
 infershaped/infershape_launchers_test.cc DEPS infrt)

paddle/infrt/kernel/pten/infershaped/infershape_launchers_test.cc

@@ -17,6 +17,10 @@
 #include "paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launcher.h"
 #include "paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launchers.h"
 #include "paddle/infrt/kernel/pten/infershaped/infershaped_utils.h"
+#include "paddle/pten/backends/cpu/cpu_context.h"
+#include "paddle/pten/common/place.h"
+#include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/meta_tensor.h"
 
 namespace infrt {
 namespace kernel {
@@ -33,23 +37,62 @@ TEST(utils, registry) {
   CHECK_EQ(count, 2U);
 }
 
+class FancyAllocator : public pten::Allocator {
+ public:
+  static void Delete(pten::Allocation* allocation) {
+    ::operator delete(allocation->ptr());
+  }
+
+  AllocationPtr Allocate(size_t bytes_size) override {
+    void* data = ::operator new(bytes_size);
+    auto* allocation = new pten::Allocation(data, bytes_size, pten::CPUPlace());
+    return AllocationPtr(allocation, Delete);
+  }
+};
+
 TEST(ElementwiseAdd, launcher_registry) {
   host_context::KernelRegistry registry;
   RegisterInferShapeLaunchers(&registry);
-  ASSERT_EQ(registry.size(), 1UL);
-  auto creator = registry.GetKernel("elementwise_add");
+  ASSERT_GE(registry.size(), 1UL);
+  auto creator = registry.GetKernel("add.cpu.any.fp32");
+
+  const pten::DDim dims({1, 2});
+  const pten::DataType dtype{pten::DataType::FLOAT32};
+  const pten::DataLayout layout{pten::DataLayout::NHWC};
+  const pten::LoD lod{};
+  pten::DenseTensorMeta meta(dtype, dims, layout, lod);
+
+  auto fancy_allocator = std::unique_ptr<pten::Allocator>(new FancyAllocator);
+  auto* alloc = fancy_allocator.get();
 
-  ::pten::CPUContext ctx{};
-  ::pten::DenseTensor a{};
-  ::pten::DenseTensor b{};
-  ::pten::DenseTensor c{};
+  pten::DenseTensor a(alloc, meta);
+  pten::DenseTensor b(alloc, meta);
+  pten::DenseTensor c(alloc, meta);
+
+  auto place = pten::CPUPlace();
+  float* a_data = a.mutable_data<float>(place);
+  float* b_data = b.mutable_data<float>(place);
+  float* c_data = c.mutable_data<float>(place);
+  for (size_t i = 0; i < 2; ++i) {
+    a_data[i] = 1.f;
+    b_data[i] = 2.f;
+  }
+
+  pten::CPUContext context;
+  context.SetAllocator(alloc);
+  context.Init();
 
   host_context::KernelFrameBuilder kernel_frame_builder;
-  kernel_frame_builder.AddArgument(new host_context::Value(std::move(ctx)));
+  kernel_frame_builder.AddArgument(new host_context::Value(std::move(context)));
   kernel_frame_builder.AddArgument(new host_context::Value(std::move(a)));
   kernel_frame_builder.AddArgument(new host_context::Value(std::move(b)));
   kernel_frame_builder.SetResults({new host_context::Value(std::move(c))});
+
   creator(&kernel_frame_builder);
+
+  for (size_t i = 0; i < 2; ++i) {
+    CHECK_EQ(c_data[i], 3.f);
+  }
 }
 
 }  // namespace kernel

paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launcher.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launcher.h"
+#include "paddle/pten/core/dense_tensor.h"
 
 namespace infrt {
 namespace kernel {
@@ -26,6 +27,9 @@ void InferShapedKernelLauncher::CreateKernelFrameForInferShape(
       values.emplace_back(
           ::pten::MetaTensor{&value->get<::pten::DenseTensor>()});
       infershape_kernel_frame_builder.AddArgument(values.back().get());
+    } else if (value->is_type<pten::DenseTensor>()) {
+      values.emplace_back(pten::MetaTensor{&value->get<pten::DenseTensor>()});
+      infershape_kernel_frame_builder.AddArgument(values.back().get());
     } else {
       infershape_kernel_frame_builder.AddArgument(value);
     }

paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launchers.cc

-// 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.
-
-#include "paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launchers.h"
-#include "paddle/infrt/kernel/pten/infershaped/elementwise_add.h"
-
-namespace infrt {
-namespace kernel {
-
-void RegisterInferShapeLaunchers(host_context::KernelRegistry* registry) {
-  registry->AddKernel(
-      "elementwise_add",
-      std::bind(&KernelLauncherFunc<decltype(&ElementwiseAdd),
-                                    &ElementwiseAdd,
-                                    decltype(&ElementwiseAddInferShape),
-                                    &ElementwiseAddInferShape>,
-                KernelLauncher<decltype(&ElementwiseAdd),
-                               &ElementwiseAdd,
-                               decltype(&ElementwiseAddInferShape),
-                               &ElementwiseAddInferShape>(),
-                std::placeholders::_1));
-}
-
-}  // namespace kernel
-}  // namespace infrt

paddle/infrt/kernel/pten/infershaped/infershaped_utils.h

@@ -16,6 +16,7 @@
 
 #include <type_traits>
 #include "paddle/infrt/tensor/dense_host_tensor.h"
+#include "paddle/pten/core/dense_tensor.h"
 
 namespace infrt {
 namespace kernel {

paddle/infrt/kernel/pten/infershaped/elementwise_add.h → paddle/infrt/kernel/pten/infershaped/pten_kernel_launcher.h

@@ -19,21 +19,9 @@
 #include "paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launcher.h"
 #include "paddle/infrt/kernel/pten/infershaped/infershaped_utils.h"
 
-// This file contains a example of the infershape ElementwiseAdd kernel.
-// Some of the following code should be generated from PTEN by script.
-
 namespace infrt {
 namespace kernel {
 
-static void ElementwiseAddInferShape(const ::pten::MetaTensor& a,
-                                     const ::pten::MetaTensor& b,
-                                     ::pten::MetaTensor* c) {}
-
-static void ElementwiseAdd(const ::pten::CPUContext& /*Context*/,
-                           const ::pten::DenseTensor& a,
-                           const ::pten::DenseTensor& b,
-                           ::pten::DenseTensor* c) {}
-
 template <typename KernelFunc,
           KernelFunc kernel,
           typename InferShapedFunc,
@@ -55,7 +43,6 @@ class KernelLauncher : public InferShapedKernelLauncher {
         BuildInferShapeCache(num_input_tensors);
       }
     }
-
     ::infrt::host_context::KernelImpl<KernelFunc, kernel>::Invoke(frame);
   }
 };

paddle/infrt/kernel/pten/registry.cc

@@ -22,7 +22,7 @@
 #include "paddle/infrt/kernel/pten/allocator_kernels.h"
 #include "paddle/infrt/kernel/pten/context_kernels.h"
 #include "paddle/infrt/kernel/pten/dense_tensor_kernels.h"
-#include "paddle/infrt/kernel/pten/infershaped/elementwise_add.h"
+#include "paddle/infrt/kernel/pten/infershaped/pten_kernel_launcher.h"
 #include "paddle/pten/include/infermeta.h"
 #include "paddle/pten/include/kernels.h"
 #include "paddle/pten/kernels/matmul_kernel.h"

tools/infrt/get_pten_kernel_function.sh

@@ -22,13 +22,13 @@ set -e
 
 #step 1:get kernel registered info
 kernel_register_info_file=`mktemp`
-PADDLE_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}")/../../" && pwd )"
+PADDLE_ROOT="$( cd "$( dirname "$0" )/../../" && pwd )"
 unset GREP_OPTIONS && find ${PADDLE_ROOT}/paddle/pten/kernels -name "*.c*" \
   | xargs sed -e '/PT_REGISTER_\(GENERAL_\)\?KERNEL(/,/)/!d' \
   | awk 'BEGIN { RS="{" }{ gsub(/\n /,""); print $0 }' \
   | grep PT_REGISTER \
-   | awk -F ",|\(" '{gsub(/ /,"");print $2, $3, $4, $5}' \
-   |  sort -u | awk '{gsub(/pten::/,"");print $0}' \
+  | awk -F ",|\(|\)" '{gsub(/ /,"");$1="";print}' \
+  | sort -u  | awk '{gsub(/pten::/,"");gsub(/paddle::platform::/,"");gsub(/dtype::/,"");gsub(/paddle::/,"");print $0}' \
   | grep -v "_grad" > $kernel_register_info_file
 
 #step 2:get simple general inferMeta function wrap info
@@ -49,4 +49,5 @@ grep PT_REGISTER_INFER_META_FN ${temp_path}/generate.cc  \
 python3 ${PADDLE_ROOT}/tools/infrt/get_pten_kernel_info.py \
   --paddle_root_path ${PADDLE_ROOT} \
   --kernel_info_file $kernel_register_info_file \
-  --infermeta_wrap_file ${temp_path}/wrap_info.txt
+  --infermeta_wrap_file ${temp_path}/wrap_info.txt \
+  --generate_file ${PADDLE_ROOT}/paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launchers.cc

tools/infrt/get_pten_kernel_info.py

@@ -17,6 +17,7 @@
 import argparse
 import json
 import yaml
+from typing import List, Dict, Any
 
 
 def parse_args():
@@ -25,17 +26,23 @@ def parse_args():
         "--paddle_root_path",
         type=str,
         required=True,
-        help="root path of paddle src[WORK_PATH/Paddle] .")
+        help="root path of paddle src[WORK_PATH/Paddle].")
     parser.add_argument(
         "--kernel_info_file",
         type=str,
         required=True,
-        help="kernel info file generated by get_pten_kernel_function.sh .")
+        help="kernel info file generated by get_pten_kernel_function.sh.")
     parser.add_argument(
         "--infermeta_wrap_file",
         type=str,
         required=True,
-        help="inferMeta wrap info file .")
+        help="inferMeta wrap info file.")
+    parser.add_argument(
+        "--generate_file",
+        type=str,
+        required=True,
+        default="../paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launchers.cc",
+        help="generated file.")
     args = parser.parse_args()
     return args
 
@@ -76,10 +83,179 @@ def merge(infer_meta_data, kernel_data, wrap_data):
     return full_kernel_data
 
 
+def gen_warn_info():
+    return """// Generated by tools/infrt/gen_pten_kernel_register.py for infrt.
+// DO NOT edit or include it within paddle.
+"""
+
+
+def gen_include_headers():
+    return """
+#include "paddle/infrt/kernel/pten/infershaped/infershaped_kernel_launchers.h"
+#include "paddle/infrt/kernel/pten/infershaped/pten_kernel_launcher.h"
+#include "paddle/pten/backends/all_context.h"
+#include "paddle/pten/include/kernels.h"
+#include "paddle/pten/include/infermeta.h"
+#include "paddle/pten/infermeta/generated.h"
+"""
+
+
+def gen_namespace():
+    return ("""
+namespace infrt {
+namespace kernel {
+
+""", """
+
+}  // namespace kernel
+}  // namespace infrt
+""")
+
+
+def gen_context(val):
+    if val == "CPU":
+        return "pten::CPUContext"
+    # elif val == "GPU":
+    #     return "pten::GPUContext"
+    # elif val == "XPU":
+    #     return "pten::XPUContext"
+    else:
+        # raise Exception(f"Unknown context type {val}")
+        return ""
+
+
+def gen_layout(val):
+    if val == "ALL_LAYOUT":
+        return 'any'
+    else:
+        # TODO(wilber): now only process ALL_LAYOUT
+        raise Exception(f"Unknown layout type {val}")
+
+
+def gen_kernel_func(val, ctx_name, dtype_name):
+    if '<' in val and '>' in val:
+        st = val.index('<')
+        ed = val.index('>')
+        func_name = val[:st]
+        template_name = val[st + 1:ed]
+        if 'pten::' in template_name:
+            return "&pten::" + val
+        else:
+            return "&pten::" + func_name + "<pten::" + template_name + ">"
+    else:
+        return "&pten::" + val + "<" + dtype_name + ", " + ctx_name + ">"
+
+
+def gen_dtype(vals: List[str]):
+    ir_dtypes, origin_dtypes = [], []
+    for val in vals:
+        if val == "float":
+            ir_dtypes.append("fp32")
+            origin_dtypes.append("float")
+        elif val == "double":
+            ir_dtypes.append("fp64")
+            origin_dtypes.append("double")
+        elif val == "float16":
+            ir_dtypes.append("fp16")
+            origin_dtypes.append("paddle::experimental::float16")
+        elif val == "bfloat16":
+            ir_dtypes.append("bf16")
+            origin_dtypes.append("paddle::experimental::bfloat16")
+        elif val == "bool":
+            ir_dtypes.append("int1")
+            origin_dtypes.append("bool")
+        elif val == "int8_t":
+            ir_dtypes.append("int8")
+            origin_dtypes.append("int8_t")
+        elif val == "uint8_t":
+            ir_dtypes.append("uint8")
+            origin_dtypes.append("uint8_t")
+        elif val == "int16_t":
+            ir_dtypes.append("int16")
+            origin_dtypes.append("int16_t")
+        elif val == "int" or val == "int32_t":
+            ir_dtypes.append("int32")
+            origin_dtypes.append("int32_t")
+        elif val == "int64_t":
+            ir_dtypes.append("int64")
+            origin_dtypes.append("int64_t")
+        elif val == "complex<float>" or val == "complex64":
+            ir_dtypes.append("complex64")
+            origin_dtypes.append("paddle::experimental::complex64")
+        elif val == "complex<double>" or val == "complex128":
+            ir_dtypes.append("complex128")
+            origin_dtypes.append("paddle::experimental::complex128")
+        elif val == "ALL_DTYPE":
+            ir_dtypes.append("all")
+            origin_dtypes.append("all")
+        else:
+            if "VA_ARGS" in val:
+                continue
+            raise Exception(f"Unknown data type {val}")
+    return ir_dtypes, origin_dtypes
+
+
+# TODO(wilber): Now only process CPUContext.
+def gen_register_info(resources: List[List[str]]):
+    """
+    resources: [['add', 'CPU', 'ALL_LAYOUT', 'AddKernel', 'float', 'double', '...'(varaidic types), 'ElementwiseInferMeta'], ...]
+    """
+    res = "void RegisterInferShapeLaunchers(host_context::KernelRegistry* registry) {"
+    for item in resources:
+        # The output string is polluted by C++ macros, here the \ is removed
+        update_item = [v.strip('\\') for v in item]
+
+        ctx_name = gen_context(update_item[1])
+        if (ctx_name == ""):
+            continue
+        update_item[2] = gen_layout(update_item[2])
+        ir_dtypes, origin_dtypes = gen_dtype(update_item[4:-1])
+        infer_shape_func = "&pten::" + update_item[-1]
+
+        if update_item[-1] == "unknown":
+            # TODO(wilber): handle the unknown inferShape func.
+            continue
+
+        for ir_dtype, origin_dtype in zip(ir_dtypes, origin_dtypes):
+            kernel_func = gen_kernel_func(update_item[3], ctx_name,
+                                          origin_dtype)
+            ir_name = '.'.join(
+                [it.lower() for it in update_item[:3]]) + "." + ir_dtype
+            res += f"""
+  registry->AddKernel("{ir_name}","""
+
+            res += f"""
+    std::bind(&KernelLauncherFunc<decltype({kernel_func}),
+                                  {kernel_func},
+                                  decltype({infer_shape_func}),
+                                  {infer_shape_func}>,
+              KernelLauncher<decltype({kernel_func}),
+                                  {kernel_func},
+                                  decltype({infer_shape_func}),
+                                  {infer_shape_func}>(),
+              std::placeholders::_1));
+"""
+
+    res += "\n}"
+    return res
+
+
+def gen_pten_kernel_register_code(resources: List[List[str]],
+                                  src_file_path: str):
+    source_file = open(src_file_path, 'w')
+    source_file.write(gen_warn_info())
+    source_file.write(gen_include_headers())
+    namespace = gen_namespace()
+    source_file.write(namespace[0])
+    source_file.write(gen_register_info(resources))
+    source_file.write(namespace[1])
+    source_file.close()
+
+
 if __name__ == "__main__":
     args = parse_args()
     infer_meta_data = get_api_yaml_info(args.paddle_root_path)
     kernel_data = get_kernel_info(args.kernel_info_file)
     info_meta_wrap_data = get_kernel_info(args.infermeta_wrap_file)
     out = merge(infer_meta_data, kernel_data, info_meta_wrap_data)
-    print(json.dumps(out))
+    gen_pten_kernel_register_code(out, args.generate_file)