Merge branch 'improve_4k_gemm' of...

Merge branch 'improve_4k_gemm' of https://github.com/ROCmSoftwarePlatform/composable_kernel into gemm_opt_exp

include/ck/host_utility/kernel_launch.hpp

@@ -33,14 +33,13 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
         printf("Warm up 1 time\n");
 #endif
         // warm up
-        const int warmup_repeat = 50;
-        const int profile_repeat = 200;
-        for(int i = 0; i < warmup_repeat; ++i)
+        for(int i = 0; i < stream_config.cold_niters_; ++i)
         {
             kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
             hip_check_error(hipGetLastError());
         }
 
+        const int nrepeat = stream_config.nrepeat_;
 #if DEBUG_LOG
         printf("Start running %d times...\n", nrepeat);
 #endif

include/ck/stream_config.hpp

@@ -11,4 +11,6 @@ struct StreamConfig
     hipStream_t stream_id_ = nullptr;
     bool time_kernel_      = false;
     int log_level_         = 0;
+    int cold_niters_       = 50;
+    int nrepeat_           = 200;
 };

library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp

@@ -377,7 +377,7 @@ struct DeviceOperationInstanceFactory<
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
             {
-                add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+                /// add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
 #endif
@@ -386,7 +386,7 @@ struct DeviceOperationInstanceFactory<
             else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
-                add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+                /// add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
 #endif
@@ -395,7 +395,7 @@ struct DeviceOperationInstanceFactory<
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
                               is_same_v<CLayout, Row>)
             {
-                add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+                /// add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
 #endif
@@ -404,7 +404,7 @@ struct DeviceOperationInstanceFactory<
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
-                add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+                /// add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
 #endif
@@ -418,7 +418,7 @@ struct DeviceOperationInstanceFactory<
             if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
                          is_same_v<CLayout, Row>)
             {
-                add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                /// add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs);
@@ -430,7 +430,7 @@ struct DeviceOperationInstanceFactory<
             else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
-                add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                /// add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs);
@@ -443,7 +443,7 @@ struct DeviceOperationInstanceFactory<
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
                               is_same_v<CLayout, Row>)
             {
-                add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+                /// add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs);
@@ -455,7 +455,7 @@ struct DeviceOperationInstanceFactory<
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
-                add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+                /// add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs);

profiler/include/profiler/profile_gemm_impl.hpp

@@ -6,6 +6,7 @@
 #include <iomanip>
 #include <iostream>
 #include <typeinfo>
+#include <unistd.h>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
@@ -20,6 +21,7 @@
 #include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/fill.hpp"
 
 namespace ck {
 namespace profiler {
@@ -69,14 +71,17 @@ int profile_gemm_impl(int do_verification,
 
     switch(init_method)
     {
-    case 0: break;
+    case 0:
+        ck::utils::FillConstant<ADataType>{static_cast<ADataType>(1.f)}(a_m_k);
+        ck::utils::FillConstant<BDataType>{static_cast<BDataType>(1.f)}(b_k_n);
+        break;
     case 1:
-        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
         break;
     default:
-        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 0.1});
-        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.01, 0.01});
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
     }
 
     using AElementOp = ck::tensor_operation::element_wise::PassThrough;
@@ -130,11 +135,10 @@ int profile_gemm_impl(int do_verification,
         ref_invoker.Run(ref_argument);
     }
 
-    std::string best_op_name;
-    float best_avg_time   = 0;
     float best_tflops    = 0;
-    float best_gb_per_sec = 0;
+    int best_instance_id = 0;
 
+    int instance_id = 0;
     // profile device op instances
     for(auto& op_ptr : op_ptrs)
     {
@@ -178,10 +182,8 @@ int profile_gemm_impl(int do_verification,
 
             if(tflops > best_tflops)
             {
-                best_op_name    = op_name;
+                best_instance_id = instance_id;
                 best_tflops      = tflops;
-                best_avg_time   = avg_time;
-                best_gb_per_sec = gb_per_sec;
             }
 
             if(do_verification)
@@ -205,8 +207,47 @@ int profile_gemm_impl(int do_verification,
         {
             std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
         }
+
+        instance_id++;
     }
 
+    sleep(2);
+
+    // Run the best instance again
+    {
+        auto& op_ptr = op_ptrs[best_instance_id];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        StrideA,
+                                        StrideB,
+                                        StrideC,
+                                        a_element_op,
+                                        b_element_op,
+                                        c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float avg_time = invoker_ptr->Run(argument_ptr.get(),
+                                              StreamConfig{nullptr, time_kernel, 0, 50, 200});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
             if constexpr(is_same<CDataType, float>::value)
             {
                 std::cout << "Best Perf for datatype = f32";
@@ -249,9 +290,11 @@ int profile_gemm_impl(int do_verification,
             }
 
             std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
-              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_avg_time
-              << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
-              << best_op_name << std::endl;
+                      << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << avg_time
+                      << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " << op_name
+                      << std::endl;
+        }
+    }
 
     return pass ? 0 : 1;
 }