v3.9 update (#2213)

Co-authored-by: yuzhai <yuzhai@nvidia.com>

tools/profiler/src/options.cu

@@ -31,7 +31,8 @@
 /* \file
    \brief Command line options for performance test program
 */
-
+#include <cuda.h>
+#include <cuda_runtime_api.h>
 #include <algorithm>
 #include <fstream>
 #include <set>
@@ -165,9 +166,11 @@ void Options::Device::print_usage(std::ostream &out) const {
         break;
       }
       else {
+        int32_t clock_KHz;
+        cudaDeviceGetAttribute(&clock_KHz, cudaDevAttrClockRate, 0);
         out << "    [" << idx << "] - "
           << prop.name << " - SM " << prop.major << "." << prop.minor << ", "
-          << prop.multiProcessorCount << " SMs @ " << (prop.clockRate / 1000.0) << " MHz, "
+          << prop.multiProcessorCount << " SMs @ " << (clock_KHz / 1000.0) << " MHz, "
           << "L2 cache: " << (prop.l2CacheSize >> 20) << " MB, Global Memory: " << (prop.totalGlobalMem >> 30) << " GB"
           << std::endl;
       }
@@ -216,9 +219,11 @@ void Options::Device::print_options(std::ostream &out, int indent) const {
   for (int device : devices) {
     out << device << ',';
   }
+  int32_t clock_KHz;
+  cudaDeviceGetAttribute(&clock_KHz, cudaDevAttrClockRate, 0);
   out
     << "\n"
-    << indent_str(indent) << "clock: " << int(double(properties[0].clockRate) / 1000.0) << "\n"
+    << indent_str(indent) << "clock: " << int(double(clock_KHz) / 1000.0) << "\n"
     << indent_str(indent) << "compute-capability: " << compute_capability(0) << "\n";
 }
 

tools/util/include/cutlass/util/reference/device/tensor_compare.h

@@ -109,7 +109,8 @@ bool BlockCompareEqual(
   Element const *ptr_B,
   size_t capacity,
   int grid_size = 0, 
-  int block_size = 0) {
+  int block_size = 0,
+  cudaStream_t stream = nullptr) {
 
   int equal_flag = 1;
   int *device_equal_flag = nullptr;
@@ -146,7 +147,9 @@ bool BlockCompareEqual(
   dim3 grid(grid_size, 1, 1);
   dim3 block(block_size, 1, 1);
 
-  kernel::BlockCompareEqual<Element><<< grid, block >>>(device_equal_flag, ptr_A, ptr_B, capacity);
+  kernel::BlockCompareEqual<Element><<< grid, block, 0, stream >>>(device_equal_flag, ptr_A, ptr_B, capacity);
+
+  cudaStreamSynchronize(stream);
 
   if (cudaMemcpy(
     &equal_flag, 
@@ -175,7 +178,8 @@ bool BlockCompareRelativelyEqual(
   Element epsilon,
   Element nonzero_floor,
   int grid_size = 0, 
-  int block_size = 0) {
+  int block_size = 0,
+  cudaStream_t stream = nullptr) {
 
   int equal_flag = 1;
   int *device_equal_flag = nullptr;
@@ -212,7 +216,7 @@ bool BlockCompareRelativelyEqual(
   dim3 grid(grid_size, 1, 1);
   dim3 block(block_size, 1, 1);
 
-  kernel::BlockCompareRelativelyEqual<Element><<< grid, block >>>(
+  kernel::BlockCompareRelativelyEqual<Element><<< grid, block, 0, stream >>>(
     device_equal_flag, 
     ptr_A, 
     ptr_B, 
@@ -221,6 +225,8 @@ bool BlockCompareRelativelyEqual(
     nonzero_floor
   );
 
+  cudaStreamSynchronize(stream);
+
   if (cudaMemcpy(
     &equal_flag, 
     device_equal_flag,

tools/util/include/cutlass/util/reference/device/tensor_reduce.h

@@ -232,6 +232,8 @@ ComputeType TensorTransformReduce(
     workspace, identity, workspace_size, reduce
   );
 
+  cudaStreamSynchronize(stream);
+
   if (copy_out) {
     cudaError_t result = cudaMemcpy(&identity, workspace, sizeof(identity), cudaMemcpyDeviceToHost);
     if (result != cudaSuccess) {
@@ -285,6 +287,8 @@ ComputeType TensorTransformReduce(
     workspace, identity, workspace_size, reduce
   );
 
+  cudaStreamSynchronize(stream);
+
   if (copy_out) {
     cudaError_t result = cudaMemcpy(&identity, workspace, sizeof(identity), cudaMemcpyDeviceToHost);
     if (result != cudaSuccess) {