CUTLASS 2.8 (#363)

CUTLASS 2.8

tools/profiler/src/cublas_helpers.cu

@@ -220,7 +220,7 @@ cublasGemmExDispatcher::cublasGemmExDispatcher(
 
   // cuBLAS introduces a separate cublasComputeType enumerant to more precisely describe
   // internal numerical data types used in the computation.
-#if (__CUDA_VER_MAJOR__ >= 11)
+#if (__CUDACC_VER_MAJOR__ >= 11)
   library::OpcodeClassID const & opcode_class =
     op_desc.tile_description.math_instruction.opcode_class;
 
@@ -253,7 +253,7 @@ cublasGemmExDispatcher::cublasGemmExDispatcher(
         break;
     }
   }
-#endif // __CUDA_VER_MAJOR__ >= 11
+#endif // __CUDACC_VER_MAJOR__ >= 11
 
   if (!good) {
     status = Status::kErrorNotSupported;
@@ -286,7 +286,7 @@ cublasStatus_t cublasGemmExDispatcher::operator()(cublasHandle_t handle) {
       int(configuration.ldc),
       arguments.batch_stride_C,
       configuration.batch_count,
-  #if (__CUDA_VER_MAJOR__ >= 11)
+  #if (__CUDACC_VER_MAJOR__ >= 11)
       compute_type,
   #else
       compute_data_type,
@@ -313,7 +313,7 @@ cublasStatus_t cublasGemmExDispatcher::operator()(cublasHandle_t handle) {
       arguments.D,
       data_type_C,
       int(configuration.ldc),
-  #if (__CUDA_VER_MAJOR__ >= 11)
+  #if (__CUDACC_VER_MAJOR__ >= 11)
       compute_type,
   #else
       compute_data_type,

tools/profiler/src/cublas_helpers.h

@@ -187,7 +187,7 @@ struct cublasGemmExDispatcher {
   cudaDataType_t data_type_C;
   cudaDataType_t compute_data_type;
 
-#if (__CUDA_VER_MAJOR__ >= 11)
+#if (__CUDACC_VER_MAJOR__ >= 11)
   cublasComputeType_t compute_type;
 #endif
 

tools/profiler/src/gemm_operation_profiler.cu

@@ -254,6 +254,10 @@ int64_t GemmOperationProfiler::GemmProblem::flops(library::GemmDescription const
     flops_ *= 4;
     break;
 
+  case library::MathOperationID::kMultiplyAddComplexFastF32:
+    flops_ *= 4;
+    break;
+
   case library::MathOperationID::kMultiplyAddGaussianComplex:
     flops_ *= 3;
     break;

tools/profiler/src/operation_profiler.cu

@@ -260,6 +260,9 @@ int OperationProfiler::profile_all(
       auto min_cc = operation->description().tile_description.minimum_compute_capability;
       auto max_cc = operation->description().tile_description.maximum_compute_capability;
 
+      // Clear named allocations
+      device_context.free();
+
       // Execute compatible cutlass operations if they satisfy the current device's compute capability
       if (operation->description().kind == kind_ &&
         operation->description().provider == library::Provider::kCUTLASS &&
@@ -301,9 +304,12 @@ int OperationProfiler::profile_all(
           problem);
 
         if (status == Status::kErrorInternal) {
-          // Stop profiling if there was an internal error
-          internal_error = true;
-          break;
+          
+          // If there was an internal error, consume the CUDA error and move to the next operation.
+          (void)cudaGetLastError();
+          
+          report.append_results(results_);
+          continue;
         }
         else if (status != Status::kSuccess) {
           // If the workspace could not be initialized for any other reason, continue to
@@ -322,9 +328,12 @@ int OperationProfiler::profile_all(
             problem);
 
           if (status == Status::kErrorInternal) {
-            // Stop profiling if there was an internal error
-            internal_error = true;
-            break;
+
+            // If there was an internal error, consume the CUDA error and move to the next operation.
+            (void)cudaGetLastError();
+
+            report.append_results(results_);
+            continue;
           }
           else if (status != Status::kSuccess) {
             // If the workspace could not be initialized for any other reason, continue to
@@ -383,9 +392,6 @@ int OperationProfiler::profile_all(
             problem);
         }
 
-        // Clear named allocations
-        device_context.free();
-
         report.append_results(results_);
         results_.clear();
       }

tools/profiler/src/options.cu

@@ -577,6 +577,8 @@ Options::Report::Report(cutlass::CommandLine const &cmdline) {
   cmdline.get_cmd_line_argument("report-not-run", report_not_run, false);
 
   cmdline.get_cmd_line_argument("verbose", verbose, true);
+
+  cmdline.get_cmd_line_argument("sort-results", sort_results, false);
 }
 
 void Options::Report::print_usage(std::ostream &out) const {
@@ -602,7 +604,10 @@ void Options::Report::print_usage(std::ostream &out) const {
     << "      column. Useful for generating pivot tables.\n\n"
 
     << "  --verbose=<bool>                             "
-    << "    Prints human-readable text to stdout. If false, nothing is written to stdout.\n\n";
+    << "    Prints human-readable text to stdout. If false, nothing is written to stdout.\n\n"
+
+    << "  --sort-results=<bool>                        "
+    << "    Sorts results (by flops-per-byte).\n\n";
 }
 
 void Options::Report::print_options(std::ostream &out, int indent) const {

tools/profiler/src/options.h

@@ -231,6 +231,9 @@ public:
     /// Prints human-readable text to stdout. If false, nothing is written to stdout
     bool verbose;
 
+    /// Sort results by (currently by flops-per-byte)
+    bool sort_results;
+
     //
     // Methods
     //

tools/profiler/src/performance_report.cpp

@@ -154,6 +154,22 @@ void PerformanceReport::append_result(PerformanceResult result) {
   }
 }
 
+void PerformanceReport::sort_results(PerformanceResultVector &results) {
+
+  struct FlopsPerByteCompare
+  {
+    bool operator()(const PerformanceResult &a, const PerformanceResult &b)
+    {
+      double a_flops_per_byte = double(a.flops) / double(a.bytes);
+      double b_flops_per_byte = double(b.flops) / double(b.bytes);
+
+      return (a_flops_per_byte < b_flops_per_byte);
+    }
+  };
+
+  std::stable_sort(results.begin(), results.end(), FlopsPerByteCompare());
+}
+
 void PerformanceReport::append_results(PerformanceResultVector const &results) {
 
   if (options_.report.verbose) {
@@ -173,12 +189,16 @@ PerformanceReport::~PerformanceReport() {
   //
   if (options_.report.verbose && !concatenated_results_.empty()) {
 
+    if (options_.report.sort_results) {
+      sort_results(concatenated_results_);
+    }
+
     std::cout << "\n\n";
     std::cout << "=============================\n\n";
     std::cout << "CSV Results:\n\n";
 
     print_csv_header_(std::cout) << std::endl;
-    
+
     for (auto const &result : concatenated_results_) {
       print_result_csv_(std::cout, result) << "\n";
     }
@@ -273,7 +293,8 @@ std::ostream & PerformanceReport::print_result_pretty_(
 
   out 
     << "           Bytes: " << result.bytes << "  bytes\n"
-    << "           FLOPs: " << result.flops << "  flops\n\n";
+    << "           FLOPs: " << result.flops << "  flops\n"
+    << "           FLOPs/Byte: " << (result.flops / result.bytes) << "\n\n";
 
   if (result.good()) {
 
@@ -309,6 +330,7 @@ std::ostream & PerformanceReport::print_csv_header_(
   out 
     << ",Bytes"
     << ",Flops"
+    << ",Flops/Byte"
     << ",Runtime"
     << ",GB/s"
     << ",GFLOPs"
@@ -345,6 +367,7 @@ std::ostream & PerformanceReport::print_result_csv_(
   out 
     << "," << result.bytes
     << "," << result.flops
+    << "," << result.flops / result.bytes
     << "," << result.runtime;
 
   if (result.good()) {

tools/profiler/src/performance_report.h

@@ -86,6 +86,7 @@ public:
 
   void next_problem();
   void append_result(PerformanceResult result);
+  void sort_results(PerformanceResultVector &results);
   void append_results(PerformanceResultVector const &results);
 
 public: