cinn(dynamic): support run exp sub subgraph with dynamic shape graph (#59640)

6clc · web-flow · commit ab38416384ad · 2023-12-04T14:24:59.000+08:00
修改broadcast的compute，使得output shape和input shape 一致的计算支持动态形状
 联调bucket机制，在不进行op schedule、group schedule的情况下可以跑通流程
 增加exp sub动态形状的子图单测。
diff --git a/paddle/cinn/hlir/framework/pir/compilation_task.cc b/paddle/cinn/hlir/framework/pir/compilation_task.cc
@@ -56,7 +56,8 @@ void CompilationTask::operator()() {
 
 void CompilationTask::Lowering() {
   auto op_lowerer = CreateOpLowerer<pir::GroupPtr>(context_->target_);
-  context_->SetLoweredFuncs(op_lowerer.BucketLower(context_->group_));
+  context_->SetLoweredFuncs(
+      op_lowerer.BucketLower(context_->group_, false, false, false));
 }
 
 void CompilationTask::CodegenAndJit() {
diff --git a/paddle/cinn/hlir/framework/pir/op_lowering_impl.cc b/paddle/cinn/hlir/framework/pir/op_lowering_impl.cc
@@ -118,6 +118,16 @@ std::vector<ir::Tensor> CollectInputTensor(
       if (func_args != nullptr) {
         func_args->push_back(tensor);
       }
+    } else {
+      // TODO(6clc): After supporting symbolic calculation,
+      // 1. Check that the shape of the tensor with the same name is the same
+      // size
+      // 2. Or make the symbol expression in compute output tensor consistent
+      //    with the one inferred in shape_analysis
+      (*tensor_map)[in_value]->sym_shape = tensor->sym_shape;
+      (*tensor_map)[in_value]->shape = tensor->shape;
+      (*tensor_map)[in_value]->sym_domain = tensor->sym_domain;
+      (*tensor_map)[in_value]->domain = tensor->domain;
     }
     tensors.push_back(tensor);
   }
@@ -528,15 +538,22 @@ std::vector<ir::LoweredFunc> OpLowererImpl::PostProcess(
   // update args for dynamic dim
   int num_tensor_args = static_cast<int>(group_func_args.size());
   int non_tensor_arg_idx = group_func_args.size();
+  std::unordered_set<std::string> int_args_set;
   for (int tensor_arg_idx = 0; tensor_arg_idx < num_tensor_args;
        tensor_arg_idx++) {
     auto tensor_dim = (*group_func_arg_tensors)[tensor_arg_idx]->sym_shape;
     int tensor_dim_size = tensor_dim.size();
     for (int tensor_arg_dim_idx = 0; tensor_arg_dim_idx < tensor_dim_size;
          tensor_arg_dim_idx++) {
       if (tensor_dim[tensor_arg_dim_idx]->IsDynamic()) {
-        group_func_args.emplace_back(ir::_Var_::Make(
-            tensor_dim[tensor_arg_dim_idx]->GetSymbolName(), common::Int(32)));
+        const std::string symbol_name =
+            tensor_dim[tensor_arg_dim_idx]->GetSymbolName();
+        if (int_args_set.count(symbol_name) != 0) {
+          continue;
+        }
+        int_args_set.insert(symbol_name);
+        group_func_args.emplace_back(
+            ir::_Var_::Make(symbol_name, common::Int(32)));
         group->int_args_map[non_tensor_arg_idx++] = {tensor_arg_idx,
                                                      tensor_arg_dim_idx};
       }
diff --git a/paddle/cinn/hlir/pe/broadcast.cc b/paddle/cinn/hlir/pe/broadcast.cc
@@ -23,6 +23,7 @@
 #include "paddle/cinn/ir/utils/ir_copy.h"
 #include "paddle/cinn/lang/builtin.h"
 #include "paddle/cinn/lang/compute.h"
+PD_DECLARE_bool(cinn_bucket_compile);
 
 namespace cinn {
 namespace hlir {
@@ -241,13 +242,22 @@ Tensor Broadcast(const FuncOp& op,
   // the counts of left-shift of tensor b so as to right alignment
   int axis_offset = 0;
 
-  GetBroadcastShape(a->shape,
-                    b->shape,
-                    &common_shape,
-                    &broadcast_flags1,
-                    &broadcast_flags2,
-                    &axis_offset,
-                    axis);
+  if (FLAGS_cinn_bucket_compile) {
+    // TODO(6clc): After supporting symbolic calculation,
+    // perfect the logic of shape equal judgment
+    common_shape = a->shape;
+    broadcast_flags1.resize(common_shape.size(), true);
+    broadcast_flags2.resize(common_shape.size(), true);
+  } else {
+    GetBroadcastShape(a->shape,
+                      b->shape,
+                      &common_shape,
+                      &broadcast_flags1,
+                      &broadcast_flags2,
+                      &axis_offset,
+                      axis);
+  }
+
   auto fn = [=](const std::vector<Expr>& indice) {
     std::vector<Expr> broadcast_indice1;
     std::vector<Expr> broadcast_indice2;
diff --git a/paddle/cinn/runtime/cuda/cuda_util.cc b/paddle/cinn/runtime/cuda/cuda_util.cc
@@ -80,7 +80,7 @@ class CublasHandle {
 
 int32_t cinn_get_value_in_cuda_kernel_args(void *v_args, int idx) {
   cinn_pod_value_t *args = static_cast<cinn_pod_value_t *>(v_args);
-  return args[idx].operator int32_t();
+  return args[idx].operator int64_t();
 }
 
 void cinn_call_cuda_kernel(void *kernel_fn,
diff --git a/paddle/fluid/framework/new_executor/instruction/cinn_jit_instruction.cc b/paddle/fluid/framework/new_executor/instruction/cinn_jit_instruction.cc
@@ -24,6 +24,7 @@
 #if defined(PADDLE_WITH_CUDA)
 #include "paddle/cinn/runtime/cinn_runtime.h"
 #endif
+PD_DECLARE_bool(cinn_bucket_compile);
 
 namespace paddle {
 namespace framework {
@@ -50,6 +51,9 @@ class CinnJitInstruction::FnPtrImpl {
     for (const auto& int_arg_mp : cinn_kernel_info_.int_args_map) {
       func_args_.emplace_back(kernel_args[int_arg_mp.second.arg_idx]->dims().at(
           int_arg_mp.second.dim_idx));
+      func_args_.emplace_back(static_cast<int64_t>(
+          kernel_args[int_arg_mp.second.arg_idx]->dims().at(
+              int_arg_mp.second.dim_idx)));
     }
 
     // 3. Launch host kernel
@@ -100,13 +104,13 @@ CinnJitInstruction::CinnJitInstruction(
 
     tensor_args_.push_back(tensor);
 
-    out_tensor_ = tensor;
-
-    auto alloc_tensor_type =
-        result.type().dyn_cast<paddle::dialect::AllocatedDenseTensorType>();
-    tensor->set_type(
-        paddle::dialect::TransToPhiDataType(alloc_tensor_type.dtype()));
-    tensor->Resize(alloc_tensor_type.dims());
+    if (!FLAGS_cinn_bucket_compile) {
+      auto alloc_tensor_type =
+          result.type().dyn_cast<paddle::dialect::AllocatedDenseTensorType>();
+      tensor->set_type(
+          paddle::dialect::TransToPhiDataType(alloc_tensor_type.dtype()));
+      tensor->Resize(alloc_tensor_type.dims());
+    }
   }
 }
 
@@ -115,7 +119,14 @@ void CinnJitInstruction::Run() {
   auto gpu_ctx = static_cast<phi::GPUContext*>(dev_ctx_);
 
   auto stream = gpu_ctx->stream();
+
   for (size_t i = 0; i < tensor_args_.size(); ++i) {
+    // TODO(6clc): template infer shape from tensor_args_[0].
+    // After supporting symbolic calculation, perfect the code to query shape
+    // of output tensor
+    if (FLAGS_cinn_bucket_compile) {
+      tensor_args_[i]->Resize(tensor_args_[0]->dims());
+    }
     gpu_ctx->Alloc(tensor_args_[i], tensor_args_[i]->dtype());
   }
 
diff --git a/test/ir/pir/cinn/CMakeLists.txt b/test/ir/pir/cinn/CMakeLists.txt
@@ -8,14 +8,29 @@ if(WITH_GPU)
 
   foreach(cinn_pir_test_name ${CINN_PIR_TEST})
     string(REGEX REPLACE ".py" "" cinn_pir_test_name ${cinn_pir_test_name})
-    add_test(
-      NAME ${cinn_pir_test_name}
-      COMMAND
-        ${CMAKE_COMMAND} -E env
-        PYTHONPATH=${CMAKE_BINARY_DIR}:${CMAKE_BINARY_DIR}/python/:$ENV{PYTHONPATH}
-        FLAGS_enable_pir_api=1 FLAGS_prim_all=True ${PYTHON_EXECUTABLE}
-        ${CMAKE_CURRENT_SOURCE_DIR}/${cinn_pir_test_name}.py
-      WORKING_DIRECTORY ${CMAKE_BINARY_DIR})
+    string(FIND "${cinn_pir_test_name}" "symbolic" index_of_substr)
+
+    message(${index_of_substr} "liuchao liuchao" ${cinn_pir_test_name})
+    if(index_of_substr EQUAL -1)
+      add_test(
+        NAME ${cinn_pir_test_name}
+        COMMAND
+          ${CMAKE_COMMAND} -E env
+          PYTHONPATH=${CMAKE_BINARY_DIR}:${CMAKE_BINARY_DIR}/python/:$ENV{PYTHONPATH}
+          FLAGS_enable_pir_api=1 FLAGS_prim_all=True ${PYTHON_EXECUTABLE}
+          ${CMAKE_CURRENT_SOURCE_DIR}/${cinn_pir_test_name}.py
+        WORKING_DIRECTORY ${CMAKE_BINARY_DIR})
+    else()
+      add_test(
+        NAME ${cinn_pir_test_name}
+        COMMAND
+          ${CMAKE_COMMAND} -E env
+          PYTHONPATH=${CMAKE_BINARY_DIR}:${CMAKE_BINARY_DIR}/python/:$ENV{PYTHONPATH}
+          FLAGS_cinn_bucket_compile=True FLAGS_enable_pir_api=1
+          FLAGS_prim_all=True ${PYTHON_EXECUTABLE}
+          ${CMAKE_CURRENT_SOURCE_DIR}/${cinn_pir_test_name}.py
+        WORKING_DIRECTORY ${CMAKE_BINARY_DIR})
+    endif()
     set_tests_properties(${cinn_pir_test_name} PROPERTIES LABELS
                                                           "RUN_TYPE=CINN")
   endforeach()
diff --git a/test/ir/pir/cinn/test_cinn_sub_graph_symbolic.py b/test/ir/pir/cinn/test_cinn_sub_graph_symbolic.py
@@ -0,0 +1,81 @@
+# 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
+#
+#     https://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 unittest
+
+import numpy as np
+
+import paddle
+from paddle.static import InputSpec
+
+
+def apply_to_static(net, use_cinn, input_spec=None):
+    build_strategy = paddle.static.BuildStrategy()
+    build_strategy.build_cinn_pass = use_cinn
+    return paddle.jit.to_static(
+        net,
+        input_spec=input_spec,
+        build_strategy=build_strategy,
+        full_graph=True,
+    )
+
+
+def exp_sub(x):
+    y = paddle.exp(x)
+    z = y - x
+    return z
+
+
+class CINNSubGraphNet(paddle.nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.fn = exp_sub
+
+    def forward(self, x):
+        out = self.fn(x)
+        return out
+
+
+class TestCinnSubGraphBase(unittest.TestCase):
+    """
+    Test Pir API + @to_static + CINN.
+    """
+
+    def setUp(self):
+        paddle.seed(2022)
+        self.prepare_data()
+
+    def prepare_data(self):
+        self.shape = [64, 128]
+        self.axis = -1
+        self.x = paddle.randn(self.shape, dtype="float32")
+        self.x.stop_gradient = False
+
+    def eval_symbolic(self, use_cinn):
+        paddle.seed(2022)
+        net = CINNSubGraphNet()
+        input_spec = [InputSpec(shape=[None, 128], dtype='float32')]
+        net = apply_to_static(net, use_cinn, input_spec)
+        net.eval()
+        out = net(self.x)
+        return out
+
+    def test_eval_symolic(self):
+        cinn_out = self.eval_symbolic(use_cinn=True)
+        dy_out = self.eval_symbolic(use_cinn=False)
+        np.testing.assert_allclose(cinn_out.numpy(), dy_out.numpy(), atol=1e-8)
+
+
+if __name__ == '__main__':
+    unittest.main()

Original file line number	Diff line number	Diff line change
`@@ -56,7 +56,8 @@ void CompilationTask::operator()() {`
`56`	`56`
`57`	`57`	`void CompilationTask::Lowering() {`
`58`	`58`	`auto op_lowerer = CreateOpLowerer<pir::GroupPtr>(context_->target_);`
`59`		`- context_->SetLoweredFuncs(op_lowerer.BucketLower(context_->group_));`
	`59`	`+ context_->SetLoweredFuncs(`
	`60`	`+ op_lowerer.BucketLower(context_->group_, false, false, false));`
`60`	`61`	`}`
`61`	`62`
`62`	`63`	`void CompilationTask::CodegenAndJit() {`
Original file line number	Diff line number	Diff line change
`@@ -80,7 +80,7 @@ class CublasHandle {`
`80`	`80`
`81`	`81`	`int32_t cinn_get_value_in_cuda_kernel_args(void *v_args, int idx) {`
`82`	`82`	`cinn_pod_value_t args = static_cast<cinn_pod_value_t >(v_args);`
`83`		`- return args[idx].operator int32_t();`
	`83`	`+ return args[idx].operator int64_t();`
`84`	`84`	`}`
`85`	`85`
`86`	`86`	`void cinn_call_cuda_kernel(void *kernel_fn,`