[cherry-pick]fix cusparse compile bug in CUDA11.2, test=release/2.2 (#36913)

* fix cusparse compile bug in CUDA11.2, test=develop

* fix bug

paddle/fluid/operators/CMakeLists.txt

@@ -97,7 +97,7 @@ if (WITH_GPU OR WITH_ROCM)
     endif()
     op_library(sync_batch_norm_op)
     file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(sync_batch_norm);\n")
-    if ((NOT WIN32) AND (NOT WITH_ROCM) AND (NOT PADDLE_WITH_ARM) AND (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_LESS 11.2) )
+    if ((NOT WIN32) AND (NOT WITH_ROCM) AND (NOT PADDLE_WITH_ARM) AND (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_LESS 11.3) )
         op_library(sparse_attention_op)
         file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(sparse_attention);\n")
     endif()

paddle/fluid/platform/dynload/cusparse.h

@@ -56,8 +56,8 @@ extern void *cusparse_dso_handle;
 
 CUSPARSE_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUSPARSE_WRAP);
 
-// APIs available after CUDA 11.2
-#if CUDA_VERSION >= 11020
+// APIs available after CUDA 11.3
+#if CUDA_VERSION >= 11030
 #define CUSPARSE_ROUTINE_EACH_R2(__macro) \
   __macro(cusparseSDDMM_bufferSize);      \
   __macro(cusparseSDDMM_preprocess);      \

python/paddle/fluid/tests/unittests/test_sparse_attention_op.py

@@ -128,8 +128,8 @@ def init_csr_format(batch_size, num_heads, rows, blocksize):
 
 
 @unittest.skipIf(
-    not core.is_compiled_with_cuda() or get_cuda_version() < 11020,
-    "core is not compiled with CUDA and cuda version need larger than or equal to 11.2"
+    not core.is_compiled_with_cuda() or get_cuda_version() < 11030,
+    "core is not compiled with CUDA and cuda version need larger than or equal to 11.3"
 )
 class TestSparseAttentionOp(OpTest):
     def config(self):
@@ -190,8 +190,8 @@ class TestSparseAttentionOpShapeTest(TestSparseAttentionOp):
 
 
 @unittest.skipIf(
-    not core.is_compiled_with_cuda() or get_cuda_version() < 11020,
-    "core is not compiled with CUDA and cuda version need larger than or equal to 11.2"
+    not core.is_compiled_with_cuda() or get_cuda_version() < 11030,
+    "core is not compiled with CUDA and cuda version need larger than or equal to 11.3"
 )
 class TestSparseAttentionAPI(unittest.TestCase):
     def setUp(self):