Add small tile size for fp16/fp32 and NN layout (#80)

* add DeviceGemmSplitKXdl

* add file device_gemm_splitk_xdl.hpp

* set c matrix zero

* using atomic

* add all tuning parameter to f32 mkkn

* grid size change to 720

* add tunning parameter for NT

* add tunning parameter for TN

* add tunning parameter for TT

* add m=96tunning parameter

* add lost config

* debug

* fix sweep

* add failed tuning params

* fixed sweep logic

* clean

* add padding to M/N for irr tile size

* clean code

* add element wise operation

* fixed MPerBlock=96

* remove marco for slpitk swtich

* add test

* add new line at the end of device_gemm_xdl_instance.hpp

* remove step hack

* seperate split-k instance files

* add tunning parameters

* change disired grid size to parameters

* remove slice length

* add desiredgridsize parameter to ckProfiler

* add losting file device_gemm_xdl_splitk_instance.hpp

* change desired gride size to kbatch

* format

* format

* clean up

* add selection of device_instances

* clean code

* clean code

* add small tile size in fp16 nn

* test for rocm 4.5

* merge develop

* clean

* clean

* clean

* remove no-use code

* add padding switch to device_gemm_xdl

* add padding switch for ksplit fp32

* clean

* clean

* add files

* rename

* Update profiler.cpp

* format

Co-authored-by: ltqin <letaoqin@amd.com>
Co-authored-by: ltqin <letao.qin@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>

device_operation/include/device_gemm_xdl.hpp

@@ -11,6 +11,7 @@
 #include "tensor_descriptor.hpp"
 #include "tensor_descriptor_helper.hpp"
 #include "gridwise_gemm_xdlops_v2r3.hpp"
+#include "gemm_specialization.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -26,6 +27,7 @@ template <typename ADataType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation,
+          GemmSpecialization_t GemmSpecialization,
           ck::index_t BlockSize,
           ck::index_t MPerBlock,
           ck::index_t NPerBlock,
@@ -77,14 +79,26 @@ struct DeviceGemmXdl
             }
         }();
 
-        const auto a_grid_desc_k0_m_k1 =
-            transform_tensor_descriptor(a_grid_desc_m_k,
-                                        make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
-                                                   make_pass_through_transform(M)),
-                                        make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
 
-        return a_grid_desc_k0_m_k1;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
     }
 
     static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
@@ -104,25 +118,60 @@ struct DeviceGemmXdl
             }
         }();
 
-        const auto b_grid_desc_k0_n_k1 =
-            transform_tensor_descriptor(b_grid_desc_k_n,
-                                        make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
-                                                   make_pass_through_transform(N)),
-                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
-                                        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
 
-        return b_grid_desc_k0_n_k1;
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
     }
 
     static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
     {
-        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
         {
-            return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
         }
-        else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+        else
         {
-            return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
         }
     }
 

device_operation/include/device_gemm_xdl_splitk.hpp

@@ -11,6 +11,7 @@
 #include "tensor_descriptor.hpp"
 #include "tensor_descriptor_helper.hpp"
 #include "gridwise_gemm_xdlops_v2r4.hpp"
+#include "gemm_specialization.hpp"
 
 #ifndef CK_RUN_KERNEL_AND_TIME
 #define CK_RUN_KERNEL_AND_TIME 1
@@ -30,6 +31,7 @@ template <typename ADataType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation,
+          GemmSpecialization_t GemmSpecialization,
           ck::index_t BlockSize,
           ck::index_t MPerBlock,
           ck::index_t NPerBlock,
@@ -88,14 +90,26 @@ struct DeviceGemmXdlSplitK
             make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(M)),
             make_tuple(Sequence<0>{}, Sequence<1>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}));
-        const auto a_grid_desc_kbatch_k0_m_k1 = transform_tensor_descriptor(
-            a_grid_desc_m_kpad,
-            make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
-                       make_pass_through_transform(M)),
-            make_tuple(Sequence<1>{}, Sequence<0>{}),
-            make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
-        return a_grid_desc_kbatch_k0_m_k1;
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
     }
 
     static auto
@@ -122,25 +136,59 @@ struct DeviceGemmXdlSplitK
             make_tuple(Sequence<0>{}, Sequence<1>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}));
 
-        const auto b_grid_desc_kbatch_k0_n_k1 = transform_tensor_descriptor(
-            b_grid_desc_kpad_n,
-            make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
-                       make_pass_through_transform(N)),
-            make_tuple(Sequence<0>{}, Sequence<1>{}),
-            make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
-
-        return b_grid_desc_kbatch_k0_n_k1;
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
     }
 
     static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
     {
-        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
         {
-            return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
         }
-        else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+        else
         {
-            return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
         }
     }
 

device_operation/include/gemm_specialization.hpp

@@ -0,0 +1,17 @@
+#ifndef GEMM_SPECIALIZATION
+#define GEMM_SPECIALIZATION
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum GemmSpecialization_t
+{
+    Default,
+    MNPadding,
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif