Additional debug print information

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp

@@ -138,6 +138,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
             if(stream_config.log_level_ > 0)
             {
                 arg.Print();
+                GridwiseGemm::BlockwiseGemmPipe::HotLoopInstList::Print();
             }
 
             if(!GridwiseGemm::CheckValidity(arg))
@@ -733,7 +734,9 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
             << "BlkGemmPipelineVersion: "
             << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
             << "BlkGemmPipelinePrefetchStages: "
-            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
+            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages << ", "
+            << "Kpack: "
+            << GridwiseGemm::BlockwiseGemmPipe::AMmaKStride;
         // clang-format on
 
         return str.str();

include/ck/utility/blkgemmpipe_scheduler.hpp

@@ -90,14 +90,22 @@ struct BlockwiseGemmXdlops_pipeline_hotloop_inst
                KPerXDL);
 
         printf(" A/B buffer load inst: %d, %d\n A/B LDS write inst: %d, %d\n A/B LDS read inst: "
-               "%d, %d\n C MFMA inst: %d\n",
+               "%d, %d\n C MFMA inst: %d\n"
+               "A/B LDS read width: %d, %d, A/B LDS write width: %d, %d, A/B buffer load width: "
+               "%d/ %d\n",
                A_Buffer_Load_Inst_Num,
                B_Buffer_Load_Inst_Num,
                A_LDS_Write_Inst_Num,
                B_LDS_Write_Inst_Num,
                A_LDS_Read_Inst_Num,
                B_LDS_Read_Inst_Num,
-               C_MFMA_Inst_Num);
+               C_MFMA_Inst_Num,
+               A_LDS_Read_Width,
+               B_LDS_Read_Width,
+               ALDSWriteWidth,
+               BLDSWriteWidth,
+               ABufferLoadWidth,
+               BBufferLoadWidth);
     }
 };
 

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp

@@ -3,6 +3,9 @@
 
 #pragma once
 
+#include <string>
+#include <sstream>
+
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
@@ -83,6 +86,56 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
         return Policy::template GetSmemSize<Problem>();
     }
 
+    CK_TILE_HOST static std::string Print()
+    {
+        constexpr index_t MPerXDL = BlockGemm::WarpGemm::kM;
+        constexpr index_t NPerXDL = BlockGemm::WarpGemm::kN;
+        constexpr index_t KPerXDL = BlockGemm::WarpGemm::WarpGemmAttribute::Impl::kK;
+
+        constexpr index_t WaveSize = 64;
+        constexpr index_t WaveNumM = BlockGemmShape::BlockWarps::at(I0{});
+        constexpr index_t WaveNumN = BlockGemmShape::BlockWarps::at(I1{});
+
+        // Below should be equal to AK1|BK1
+        constexpr index_t A_LDS_Read_Width = Policy::template GetSmemPackA<Problem>();
+        constexpr index_t B_LDS_Read_Width = Policy::template GetSmemPackB<Problem>();
+
+        constexpr index_t A_LDS_Write_Width = Policy::template GetSmemPackA<Problem>();
+        constexpr index_t B_LDS_Write_Width = Policy::template GetSmemPackB<Problem>();
+
+        constexpr index_t A_Buffer_Load_Inst_Num =
+            MPerBlock * KPerBlock / (BlockSize * GetVectorSizeA());
+        constexpr index_t B_Buffer_Load_Inst_Num =
+            NPerBlock * KPerBlock / (BlockSize * GetVectorSizeB());
+
+        constexpr index_t A_LDS_Write_Inst_Num =
+            MPerBlock * KPerBlock / (BlockSize * A_LDS_Write_Width);
+        constexpr index_t B_LDS_Write_Inst_Num =
+            NPerBlock * KPerBlock / (BlockSize * B_LDS_Write_Width);
+
+        constexpr index_t A_LDS_Read_Inst_Num =
+            WaveNumN * MPerBlock * KPerBlock / (BlockSize * A_LDS_Read_Width);
+        constexpr index_t B_LDS_Read_Inst_Num =
+            WaveNumM * MPerBlock * KPerBlock / (BlockSize * B_LDS_Read_Width);
+
+        constexpr index_t C_MFMA_Inst_Num = MPerBlock * NPerBlock * KPerBlock /
+                                            (BlockSize / WaveSize) / (MPerXDL * NPerXDL * KPerXDL);
+
+        auto str = std::stringstream{};
+
+        str << "A/B vector size: " << GetVectorSizeA() << ", " << GetVectorSizeB() << "\n"
+            << "A/B LDS read/write width: " << A_LDS_Read_Width << ", " << B_LDS_Read_Width << "\n"
+            << "A/B buffer load inst: " << A_Buffer_Load_Inst_Num << ", " << B_Buffer_Load_Inst_Num
+            << "\n"
+            << "A/B LDS write inst: " << A_LDS_Write_Inst_Num << ", " << B_LDS_Write_Inst_Num
+            << "\n"
+            << "A/B LDS read inst: " << A_LDS_Read_Inst_Num << ", " << B_LDS_Read_Inst_Num << "\n"
+            << "C MFMA inst: " << C_MFMA_Inst_Num << "\n"
+            << "KPack: " << BlockGemm::Traits::KPack << "\n"
+            << "PrefetchStages: " << PrefetchStages << "\n";
+        return str.str();
+    }
+
     template <GemmPipelineScheduler Scheduler>
     struct PipelineImpl : public PipelineImplBase
     {

include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp

@@ -185,7 +185,6 @@ struct UniversalGemmPipelineAgBgCrPolicy
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
     {
-
         using ADataType = remove_cvref_t<typename Problem::ADataType>;
 
         constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;