[NPU] fix compile and update to release 2.6 (#853)

Author: qili93

.gitmodules

@@ -1,4 +1,4 @@
 [submodule "Paddle"]
 	path = Paddle
 	url = https://github.com/PaddlePaddle/Paddle.git
-	branch = develop
+	branch = release/2.6

Paddle

@@ -61,7 +61,7 @@ void BatchNormKernel(const Context& dev_ctx,
   auto* Bias = bias.get_ptr();
 
   phi::DenseTensor new_scale, new_bias;
-  const auto data_layout = phi::StringToDataLayout(data_layout_str);
+  const auto data_layout = common::StringToDataLayout(data_layout_str);
 
   int C;
   if (x_dims.size() == 2) {
@@ -308,7 +308,7 @@ void BatchNormGradKernel(
   auto* Bias = bias.get_ptr();
 
   phi::DenseTensor new_scale, new_bias;
-  const auto data_layout = phi::StringToDataLayout(data_layout_str);
+  const auto data_layout = common::StringToDataLayout(data_layout_str);
 
   int C;
   if (x_dims.size() == 2) {

backends/npu/kernels/batch_norm_kernel.cc

@@ -119,7 +119,7 @@ void Conv2dTransposeGradKernel(const Context& dev_ctx,
   auto dilations = dilation;
   if ((!dx) && (!dfilter)) return;
 
-  const phi::DataLayout data_layout = phi::StringToDataLayout(data_format);
+  const phi::DataLayout data_layout = common::StringToDataLayout(data_format);
 
   auto in_dims = x.dims();
   auto filter_dims = filter.dims();

backends/npu/kernels/conv_transpose_kernel.cc

@@ -40,7 +40,8 @@ void FillDiagonalTensorKernel(const Context &dev_ctx,
   int64_t new_dims[2], strides[2];
   std::vector<int64_t> matdim;
   matdim.resize(fill_dims[0]);
-  CalMatDims(out_dims, dim1, dim2, &offset, new_dims, strides, matdim.data());
+  phi::CalMatDims(
+      out_dims, dim1, dim2, &offset, new_dims, strides, matdim.data());
   PADDLE_ENFORCE_EQ(
       new_dims[0],
       fill_dims[0],
@@ -125,7 +126,8 @@ void FillDiagonalTensorGradKernel(const Context &dev_ctx,
     int64_t new_dims[2], strides[2];
     std::vector<int64_t> matdim;
     matdim.resize(matrows);
-    CalMatDims(dx_dims, dim1, dim2, &offset, new_dims, strides, matdim.data());
+    phi::CalMatDims(
+        dx_dims, dim1, dim2, &offset, new_dims, strides, matdim.data());
 
     auto size = x_grad->numel();
 

backends/npu/kernels/fill_diagonal_tensor_kernel.cc

@@ -173,8 +173,8 @@ std::string GetPDTensorString(const Context& dev_ctx,
 
   log_stream << "  - place: " << print_tensor.place() << std::endl;
   log_stream << "  - shape: " << print_tensor.dims().to_str() << std::endl;
-  log_stream << "  - layout: " << phi::DataLayoutToString(print_tensor.layout())
-             << std::endl;
+  log_stream << "  - layout: "
+             << common::DataLayoutToString(print_tensor.layout()) << std::endl;
 
   auto dtype = print_tensor.dtype();
   log_stream << "  - dtype: " << dtype << std::endl;

backends/npu/kernels/funcs/string_helper.cc

@@ -152,7 +152,8 @@ void GroupNormKernel(const Context& dev_ctx,
                      phi::DenseTensor* mean,
                      phi::DenseTensor* variance) {
   auto x_dims = phi::vectorize(x.dims());
-  const phi::DataLayout data_layout_data = phi::StringToDataLayout(data_layout);
+  const phi::DataLayout data_layout_data =
+      common::StringToDataLayout(data_layout);
 
   if (x_dims.size() > 3) {
     phi::DenseTensor x_tmp(x);

backends/npu/kernels/group_norm_kernel.cc

@@ -164,11 +164,11 @@ void Pool2dKernel(const Context& dev_ctx,
     // AdaptiveAvgPool2d only support NCHW
     phi::DenseTensor transformed_input, transformed_output;
     if (pooling_type == "avg" && channel_last) {
-      transformed_input.Resize(phi::make_dim(
+      transformed_input.Resize(common::make_dim(
           in_x_dims[0], in_x_dims[3], in_x_dims[1], in_x_dims[2]));
       dev_ctx.template Alloc<T>(&transformed_input);
       transformed_output.Resize(
-          phi::make_dim(out_dims[0], out_dims[3], out_dims[1], out_dims[2]));
+          common::make_dim(out_dims[0], out_dims[3], out_dims[1], out_dims[2]));
       dev_ctx.template Alloc<T>(&transformed_output);
 
       const auto& trans_runner =

backends/npu/kernels/pool2d_kernel.cc

@@ -34,7 +34,7 @@ FLAGS_DEFINE_uint64(npu_profiling_dtypes,
                         ACL_PROF_HCCL_TRACE | ACL_PROF_RUNTIME_API,
                     "ACL datatypes to profile");
 FLAGS_DEFINE_uint64(npu_profiling_metrics,
-                    static_cast<uint64_t>(ACL_AICORE_ARITHMETIC_UTILIZATION),
+                    static_cast<uint64_t>(ACL_AICORE_PIPE_UTILIZATION),
                     "AI Core metric to profile");
 
 FLAGS_DEFINE_bool(set_to_1d, true, "set_to_1d");
@@ -199,7 +199,10 @@ aclrtStream SecondaryStream::Get(aclrtStream aicore_stream) {
 void SecondaryStream::Create(aclrtStream aicore_stream) {
   RUN_CHECK(aicpu_streams.find(aicore_stream) == aicpu_streams.cend());
   aclrtStream aicpu_stream;
-  ACL_CHECK(aclrtCreateStream(&aicpu_stream));
+  ACL_CHECK(aclrtCreateStreamWithConfig(
+      reinterpret_cast<aclrtStream *>(&aicpu_stream),
+      0,
+      (ACL_STREAM_FAST_LAUNCH | ACL_STREAM_FAST_SYNC)));
   aicpu_streams[aicore_stream] = aicpu_stream;
 }
 
@@ -597,7 +600,10 @@ C_Status HostDeallocate(const C_Device device, void *ptr, size_t size) {
 }
 
 C_Status CreateStream(const C_Device device, C_Stream *stream) {
-  ACL_CHECK(aclrtCreateStream(reinterpret_cast<aclrtStream *>(stream)));
+  ACL_CHECK(aclrtCreateStreamWithConfig(
+      reinterpret_cast<aclrtStream *>(stream),
+      0,
+      (ACL_STREAM_FAST_LAUNCH | ACL_STREAM_FAST_SYNC)));
   LOG_IF(INFO, FLAGS_npu_runtime_debug)
       << "[RUNTIME] CreateStream: device=" << device->id
       << ", stream=" << *stream;

backends/npu/runtime/runtime.cc

@@ -9,3 +9,6 @@ test_zero_dim_tensor_npu
 test_momentum_op_npu
 test_matmul_op_npu
 test_linear_op_npu
+test_compare_op_npu
+test_elementwise_sub_op_npu
+test_index_sample_op_npu