temp save

example/01_gemm/CMakeLists.txt

@@ -25,6 +25,7 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v2)
 add_example_executable(example_gemm_xdl_fp16_streamk_v3 gemm_xdl_fp16_streamk_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_streamk_v3)
 add_example_executable(example_gemm_xdl_fp16_v3 gemm_xdl_fp16_v3.cpp)
+target_compile_options(example_gemm_xdl_fp16_v3 PRIVATE -mllvm -greedy-reverse-local-assignment=1 -save-temps=$PWD -Wno-gnu-line-marker)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v3)
 add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_v3)

example/01_gemm/gemm_xdl_fp16_v3.cpp

@@ -19,7 +19,7 @@ using AElementOp = PassThrough;
 using BElementOp = PassThrough;
 using CElementOp = PassThrough;
 
-static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 
 // clang-format off
 using DeviceGemmV2Instance = 
@@ -35,7 +35,7 @@ using DeviceGemmV2Instance =
         S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
         2, 8, 8, 0,
         S<8, 32, 1>,  S<0, 2, 1>,  S<0, 2, 1>, 
-        1, 8, 2, 0,
+        1, 8, 8, 0,
         1, 2, S<1, 32, 1, 8>, 8,
         ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
 // clang-format on

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp

@@ -67,8 +67,8 @@ struct BlockwiseGemmXdlops_pipeline_base
                                                       KPerBlock,
                                                       ABlockTransferSrcScalarPerVector,
                                                       BBlockTransferSrcScalarPerVector,
-                                                      ABlockTransferSrcScalarPerVector,
-                                                      BBlockTransferSrcScalarPerVector,
+                                                      A_K1,
+                                                      B_K1,
                                                       A_K1,
                                                       B_K1,
                                                       MRepeat,

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.hpp

@@ -428,6 +428,194 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
         }
     }
 
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        const BBlockDesc& b_block_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        // Global prefetch 1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Local prefill 1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // Global prefetch 2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // Local prefetch 1
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k0) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, k0, I0),
+                                   a_thread_buf);
+            });
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                   make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(n0, I0, k0, I0),
+                                   b_thread_buf);
+            });
+        });
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                block_sync_lds();
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<ComputeDataType, KPack> a_thread_vec;
+                            vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, I0, k0, ik))>{}];
+                                b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, I0, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<ComputeDataType,
+                                                     xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+
+                block_sync_lds();
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k0, I0),
+                                           a_thread_buf);
+                    });
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
+                                           b_block_buf,
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k0, I0),
+                                           b_thread_buf);
+                    });
+                });
+
+                HotLoopScheduler();
+                __builtin_amdgcn_sched_barrier(0);
+
+                i += 1;
+            } while(i < (num_loop - 1));
+        }
+        // tail
+        if constexpr(TailNum == TailNumber::Full)
+        {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
     using Base::a_thread_copy_;
     using Base::a_thread_desc_;
     using Base::b_thread_copy_;

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp

@@ -562,6 +562,354 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
         }
     }
 
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        const BBlockDesc& b_block_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+
+        // Global prefetch 1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Global prefetch 2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Local prefill 1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0), I0);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0), I0);
+
+        // Local prefill 2
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1), I1);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1), I1);
+
+        // Local prefetch 1
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                   a_block_buf.At(I0),
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, k, I0),
+                                   a_thread_bufs(I0));
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf.At(I0),
+                                       b_thread_desc_,
+                                       make_tuple(n0, I0, k, I0),
+                                       b_thread_bufs(I0));
+                });
+            });
+        });
+
+        // Global prefetch 3
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Global prefetch 4
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            // This hot loop has two legacy loopover, to implement the double local buffer strategy
+            do
+            {
+                auto LoopFunc = [&](auto lds_read_buf,
+                                    auto lds_read_reg_buf,
+                                    auto lds_write_buf,
+                                    auto vmem_buf,
+                                    auto mfma_reg_buf,
+                                    auto schedule_group) {
+                    block_sync_lds();
+
+                    static_for<0, KRepeat, 1>{}([&](auto k) {
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                               make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                               a_block_buf.At(lds_read_buf),
+                                               a_thread_desc_,
+                                               make_tuple(m0, I0, k, I0),
+                                               a_thread_bufs(lds_read_reg_buf));
+                            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                                b_thread_copy_.Run(
+                                    b_block_desc_n0_n1_n2_k,
+                                    make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                    b_block_buf.At(lds_read_buf),
+                                    b_thread_desc_,
+                                    make_tuple(n0, I0, k, I0),
+                                    b_thread_bufs(lds_read_reg_buf));
+                            });
+                        });
+                    });
+
+                    a_blockwise_copy.RunWrite(
+                        a_block_desc, a_block_buf.At(lds_write_buf), vmem_buf);
+                    b_blockwise_copy.RunWrite(
+                        b_block_desc, b_block_buf.At(lds_write_buf), vmem_buf);
+
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, vmem_buf);
+                    b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, vmem_buf);
+
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                                vector_type<ComputeDataType, KPack> a_thread_vec;
+                                vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                                static_for<0, KPack, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        a_thread_bufs[mfma_reg_buf]
+                                                     [Number<a_thread_desc_.CalculateOffset(
+                                                         make_tuple(m0, I0, k0, ik))>{}];
+                                    b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                using mfma_input_type =
+                                    typename vector_type<ComputeDataType,
+                                                         xdlops_gemm.K1PerXdlops>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                xdlops_gemm.Run(
+                                    a_thread_vec.template AsType<mfma_input_type>(),
+                                    b_thread_vec.template AsType<mfma_input_type>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+                        });
+                    });
+
+                    HotLoopScheduler(schedule_group);
+                };
+
+                LoopFunc(I1, I1, I0, I0, I0, I0);
+                LoopFunc(I0, I0, I1, I1, I1, I0);
+
+                i += HotloopUnroll;
+            } while(i < (num_loop - PrefetchStages));
+        }
+
+        auto ReadWriteCompFunc = [&](auto lds_read_buf,
+                                     auto lds_read_reg_buf,
+                                     auto lds_write_buf,
+                                     auto vmem_buf,
+                                     auto mfma_reg_buf,
+                                     auto schedule_group) {
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf.At(lds_read_buf),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, I0),
+                                       a_thread_bufs(lds_read_reg_buf));
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf.At(lds_read_buf),
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k, I0),
+                                           b_thread_bufs(lds_read_reg_buf));
+                    });
+                });
+            });
+
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf), vmem_buf);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf), vmem_buf);
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            HotLoopScheduler(schedule_group);
+        };
+
+        auto ReadCompFunc = [&](auto lds_read_buf,
+                                auto lds_read_reg_buf,
+                                auto mfma_reg_buf,
+                                auto schedule_group) {
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf.At(lds_read_buf),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, I0),
+                                       a_thread_bufs(lds_read_reg_buf));
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf.At(lds_read_buf),
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k, I0),
+                                           b_thread_bufs(lds_read_reg_buf));
+                    });
+                });
+            });
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            HotLoopScheduler(schedule_group);
+        };
+
+        auto CompFunc = [&](auto mfma_reg_buf) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+        };
+        // tail
+        if constexpr(TailNum == TailNumber::Odd)
+        {
+            ReadWriteCompFunc(I1, I1, I0, I0, I0, I1);
+            ReadCompFunc(I0, I0, I1, I1);
+            CompFunc(I0);
+        }
+        else if constexpr(TailNum == TailNumber::Even)
+        {
+            ReadWriteCompFunc(I1, I1, I0, I0, I0, I1);
+            ReadWriteCompFunc(I0, I0, I1, I1, I1, I1);
+            ReadCompFunc(I1, I1, I0, I1);
+            CompFunc(I1);
+        }
+    }
+
     using Base::a_thread_copy_;
     using Base::a_thread_desc_;
     using Base::b_thread_copy_;

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp

@@ -580,6 +580,317 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
         }
     }
 
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        const BBlockDesc& b_block_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        // Global prefetch 1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Local prefill 1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
+
+        // Global prefetch 2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Global prefetch 3
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // Local prefetch 1
+        block_sync_lds();
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                               make_tuple(m0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(m0, I0, I0, I0),
+                               a_thread_buf);
+        });
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                               make_tuple(n0, I0, I0, I0),
+                               b_block_buf,
+                               b_thread_desc_,
+                               make_tuple(n0, I0, I0, I0),
+                               b_thread_buf);
+        });
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto vmem_buf) {
+                    vector_type<ComputeDataType, KPack> a_thread_vec;
+                    vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        if constexpr(k0 == (KRepeat - 1))
+                        {
+                            block_sync_lds();
+
+                            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, vmem_buf);
+                            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, vmem_buf);
+
+                            a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, vmem_buf);
+                            b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, vmem_buf);
+
+                            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                            block_sync_lds();
+                        }
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                                static_for<0, KPack, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                            make_tuple(m0, I0, I0, ik))>{}];
+                                });
+                                static_for<0, KPack, 1>{}([&](auto ik) {
+                                    b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                            make_tuple(n0, I0, I0, ik))>{}];
+                                });
+
+                                using mfma_input_type =
+                                    typename vector_type<ComputeDataType,
+                                                         xdlops_gemm.K1PerXdlops>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                xdlops_gemm.Run(
+                                    a_thread_vec.template AsType<mfma_input_type>(),
+                                    b_thread_vec.template AsType<mfma_input_type>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+
+                            a_thread_copy_.Run(
+                                a_block_desc_m0_m1_m2_k,
+                                make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
+                                a_block_buf,
+                                a_thread_desc_,
+                                make_tuple(m0, I0, I0, I0),
+                                a_thread_buf);
+                        });
+
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            b_thread_copy_.Run(
+                                b_block_desc_n0_n1_n2_k,
+                                make_tuple(n0, I0, I0, Number<(k0 + 1) % KRepeat * BMmaKStride>{}),
+                                b_block_buf,
+                                b_thread_desc_,
+                                make_tuple(n0, I0, I0, I0),
+                                b_thread_buf);
+                        });
+                    });
+
+                    HotLoopScheduler();
+                };
+
+                LoopFunc(I0);
+                LoopFunc(I1);
+
+                i += HotloopUnroll;
+            } while(i < (num_loop - PrefetchStages));
+        }
+        // tail
+        auto ReadWriteCompFunc = [&](auto vmem_buf) {
+            vector_type<ComputeDataType, KPack> a_thread_vec;
+            vector_type<ComputeDataType, KPack> b_thread_vec;
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                if constexpr(k0 == (KRepeat - 1))
+                {
+                    block_sync_lds();
+
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, vmem_buf);
+                    b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, vmem_buf);
+
+                    block_sync_lds();
+                }
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, I0, ik))>{}];
+                        });
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, I0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                    a_thread_copy_.Run(
+                        a_block_desc_m0_m1_m2_k,
+                        make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
+                        a_block_buf,
+                        a_thread_desc_,
+                        make_tuple(m0, I0, I0, I0),
+                        a_thread_buf);
+                });
+
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_thread_copy_.Run(
+                        b_block_desc_n0_n1_n2_k,
+                        make_tuple(n0, I0, I0, Number<(k0 + 1) % KRepeat * BMmaKStride>{}),
+                        b_block_buf,
+                        b_thread_desc_,
+                        make_tuple(n0, I0, I0, I0),
+                        b_thread_buf);
+                });
+            });
+
+            HotLoopScheduler();
+        };
+        auto ReadCompFunc = [&]() {
+            vector_type<ComputeDataType, KPack> a_thread_vec;
+            vector_type<ComputeDataType, KPack> b_thread_vec;
+
+            static_for<0, KRepeat - 1, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, I0, ik))>{}];
+                        });
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, I0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+
+                    a_thread_copy_.Run(
+                        a_block_desc_m0_m1_m2_k,
+                        make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
+                        a_block_buf,
+                        a_thread_desc_,
+                        make_tuple(m0, I0, I0, I0),
+                        a_thread_buf);
+                });
+
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_thread_copy_.Run(
+                        b_block_desc_n0_n1_n2_k,
+                        make_tuple(n0, I0, I0, Number<(k0 + 1) % KRepeat * BMmaKStride>{}),
+                        b_block_buf,
+                        b_thread_desc_,
+                        make_tuple(n0, I0, I0, I0),
+                        b_thread_buf);
+                });
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KPack, 1>{}([&](auto ik) {
+                        a_thread_vec.template AsType<ComputeDataType>()(ik) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(m0, I0, I0, ik))>{}];
+                    });
+                    static_for<0, KPack, 1>{}([&](auto ik) {
+                        b_thread_vec.template AsType<ComputeDataType>()(ik) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(n0, I0, I0, ik))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                    b_thread_vec.template AsType<mfma_input_type>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+
+            HotLoopScheduler();
+        };
+
+        if constexpr(TailNum == TailNumber::Odd)
+        {
+            ReadWriteCompFunc(I0);
+            ReadWriteCompFunc(I1);
+            ReadCompFunc();
+        }
+        else if constexpr(TailNum == TailNumber::Even)
+        {
+            ReadWriteCompFunc(I0);
+            ReadCompFunc();
+        }
+    }
+
     // A[MRepeat, I1, I1, KPack]
     static constexpr auto a_thread_desc_loop =
         make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, I1, I1, Number<KPack>{}));

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp

@@ -752,18 +752,11 @@ struct GridwiseGemm_xdl_cshuffle_v3
     __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
     {
         // B matrix in LDS memory, dst of blockwise copy
-        // if constexpr(BBlockLdsExtraN)
-        // {
-        //     return make_naive_tensor_descriptor(
-        //         make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
-        //         make_tuple(BK1Number, Number<KPerBlock + BBlockLdsExtraN>{}, I1));
-        // }
-        // else
-        if constexpr(BBlockLdsExtraN && is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        if constexpr(BBlockLdsExtraN)
         {
             return make_naive_tensor_descriptor(
                 make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
-                make_tuple(BK1Number * Number<NPerBlock>{}, I1, Number<NPerBlock>{}));
+                make_tuple(BK1Number, Number<KPerBlock + BBlockLdsExtraN>{}, I1));
         }
         else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
         {
@@ -1325,9 +1318,9 @@ struct GridwiseGemm_xdl_cshuffle_v3
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
-                                                BBlockTransferSrcAccessOrder,
-                                                BBlockTransferSrcVectorDim,
+                                                Sequence<0, 1, 2>,
                                                 BBlockTransferSrcVectorDim,
+                                                2,
                                                 BBlockTransferSrcScalarPerVector,
                                                 BBlockTransferDstScalarPerVector_BK1,
                                                 1,

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp

@@ -71,7 +71,7 @@ using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_instances = std:
 
         // Latency friendly 
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   256,    64,   16,    256, 16,  16,  16,   16,    1,    1,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 64, 1, 4>,      S<4, 4, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
-        DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   128,    32,   16,    256, 16,  16,  16,   16,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,      S<4, 4, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   128,    32,   16,    256, 16,  16,  16,   16,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 4>,      S<4, 4, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,    64,    16,   16,    256, 16,  16,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,      S<4, 4, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   128,    16,   32,    256, 16,  16,  16,   16,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,      S<4, 4, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
         DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Col,     Tuple<Row, Col>, Row,     F8,     F8,    Tuple<F32, F32>, BF16,   F32,     F32,  PassThrough, PassThrough, MultiplyMultiply,     GemmSpec,   256,    16,   64,    256, 16,  16,  16,   16,    1,    1,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,     S<4, 4, 1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,

library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt

@@ -9,15 +9,15 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES
         device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_default_instance.cpp
         device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp
         device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
-        # device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp
-        # device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
-        # device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp
+        device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
+        device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
         device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_default_instance.cpp
         device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_kpadding_instance.cpp
         device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_default_instance.cpp
         device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
-        # device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_default_instance.cpp
-        # device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
+        device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_default_instance.cpp
+        device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
         
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_default_instance.cpp
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_kpadding_instance.cpp
@@ -26,15 +26,15 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_default_instance.cpp
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_kpadding_instance.cpp
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
-        # device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_default_instance.cpp
-        # device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_kpadding_instance.cpp
-        # device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_default_instance.cpp
+        device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_kpadding_instance.cpp
+        device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_default_instance.cpp
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_kpadding_instance.cpp
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instance.cpp
         device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp
-        # device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instance.cpp
-        # device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp
+        device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instance.cpp
+        device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp
         )
 
 set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
@@ -59,15 +59,15 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES
           device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_default_instance.cpp
           device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp
           device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
-        #   device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_default_instance.cpp
-        #   device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
-        #   device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
+          device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_default_instance.cpp
+          device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
+          device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
           device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_default_instance.cpp
           device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_kpadding_instance.cpp
           device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_default_instance.cpp
           device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
-        #   device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instance.cpp
-        #   device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
+          device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instance.cpp
+          device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
 
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp
@@ -76,15 +76,15 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_default_instance.cpp
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
-        #   device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp
-        #   device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
-        #   device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
+          device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp
+          device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
+          device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_kpadding_instance.cpp
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_default_instance.cpp
           device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
-        #   device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_default_instance.cpp
-        #   device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
+          device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_default_instance.cpp
+          device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
                 
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_default_instance.cpp
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_kpadding_instance.cpp
@@ -92,15 +92,15 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_default_instance.cpp
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_kpadding_instance.cpp
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_nkpadding_instance.cpp
-        #   device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_default_instance.cpp
-        #   device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instance.cpp
-        #   device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instance.cpp
+          device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_default_instance.cpp
+          device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instance.cpp
+          device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instance.cpp
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_default_instance.cpp
           device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp
-        #   device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_default_instance.cpp
-        #   device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp
+          device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_default_instance.cpp
+          device_gemm_xdl_universal_f8_f8_bf16/device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp
         )
 
  set_source_files_properties(device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")

profiler/src/CMakeLists.txt

@@ -47,7 +47,7 @@ if(GPU_TARGETS MATCHES "gfx9")
   # endif()
   # list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
   # if(GPU_TARGETS MATCHES "gfx94")
-    # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
   #   list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
   # endif()
   # list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
@@ -131,7 +131,7 @@ if(GPU_TARGETS MATCHES "gfx9")
   # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
   # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
   # if(GPU_TARGETS MATCHES "gfx94")
-    # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
   #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
   # endif()
   # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)