CUDA12.0 integration (#49539)

* Update warpctc for cuda-12

* Deprecate cudaProfilerInitialize for CUDA > 11

* Deprecate CUSPARSE_MV_ALG_DEFAULT for CUDA_VERSION >= 11040

* Add the missing thrust header

cmake/external/warpctc.cmake

@@ -23,7 +23,7 @@ set(WARPCTC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/warpctc)
 # in case of low internet speed
 #set(WARPCTC_REPOSITORY  https://gitee.com/tianjianhe/warp-ctc.git)
 set(WARPCTC_REPOSITORY ${GIT_URL}/baidu-research/warp-ctc.git)
-set(WARPCTC_TAG 37ece0e1bbe8a0019a63ac7e6462c36591c66a5b)
+set(WARPCTC_TAG bdc2b4550453e0ef2d3b5190f9c6103a84eff184)
 
 set(WARPCTC_INCLUDE_DIR
     "${WARPCTC_INSTALL_DIR}/include"

paddle/fluid/platform/device/gpu/cuda/cuda_profiler.cc

@@ -20,6 +20,7 @@ namespace platform {
 void CudaProfilerInit(const std::string& output_file,
                       const std::string& output_mode,
                       const std::string& config_file) {
+#if CUDA_VERSION < 11000
   PADDLE_ENFORCE(output_mode == "kvp" || output_mode == "csv",
                  platform::errors::InvalidArgument(
                      "Unsupported cuda profiler output mode, expect `kvp` or "
@@ -28,6 +29,7 @@ void CudaProfilerInit(const std::string& output_file,
   cudaOutputMode_t mode = output_mode == "csv" ? cudaCSV : cudaKeyValuePair;
   PADDLE_ENFORCE_GPU_SUCCESS(
       cudaProfilerInitialize(config_file.c_str(), output_file.c_str(), mode));
+#endif
 }
 
 void CudaProfilerStart() { PADDLE_ENFORCE_GPU_SUCCESS(cudaProfilerStart()); }

paddle/phi/kernels/funcs/sparse/sparse_blas_impl.cu.h

@@ -381,7 +381,11 @@ void SparseBlas<phi::GPUContext>::SPMV(bool transa,
                                           &beta,
                                           out_descriptor.descriptor(),
                                           gpu_type,
+#if CUDA_VERSION >= 11040
+                                          CUSPARSE_SPMV_ALG_DEFAULT,
+#else
                                           CUSPARSE_MV_ALG_DEFAULT,
+#endif
                                           &buffer_size);
   });
 
@@ -399,7 +403,11 @@ void SparseBlas<phi::GPUContext>::SPMV(bool transa,
                                &beta,
                                out_descriptor.descriptor(),
                                gpu_type,
+#if CUDA_VERSION >= 11040
+                               CUSPARSE_SPMV_ALG_DEFAULT,
+#else
                                CUSPARSE_MV_ALG_DEFAULT,
+#endif
                                tmp_buffer_ptr);
   });
 }

paddle/phi/kernels/gpu/send_u_recv_kernel.cu

@@ -15,6 +15,7 @@
 #include "paddle/phi/kernels/send_u_recv_kernel.h"
 
 #include <thrust/device_vector.h>
+#include <thrust/execution_policy.h>
 #include <thrust/fill.h>
 
 #include <algorithm>

paddle/phi/kernels/gpu/send_ue_recv_kernel.cu

@@ -15,6 +15,7 @@
 #include "paddle/phi/kernels/send_ue_recv_kernel.h"
 
 #include <thrust/device_vector.h>
+#include <thrust/execution_policy.h>
 #include <thrust/fill.h>
 #include <algorithm>
 #include <vector>

paddle/phi/kernels/gpu/unique_kernel.cu

@@ -20,6 +20,7 @@
 #include <thrust/functional.h>
 #include <thrust/scatter.h>
 #include <thrust/sequence.h>
+#include <thrust/sort.h>
 #include <thrust/unique.h>
 
 #include <iostream>

paddle/phi/kernels/sparse/gpu/coalesce_kernel.cu

@@ -14,6 +14,9 @@ limitations under the License. */
 
 #include "paddle/phi/kernels/sparse/coalesce_kernel.h"
 
+#include <thrust/sort.h>
+#include <thrust/unique.h>
+
 #include "paddle/phi/backends/gpu/gpu_info.h"
 #include "paddle/phi/backends/gpu/gpu_launch_config.h"
 #include "paddle/phi/core/kernel_registry.h"