add fp16_f8_fp16

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

@@ -268,7 +268,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                 const index_t m_padded = GridwiseGemm::CalculateMPadded(M);
                 const index_t n_padded = GridwiseGemm::CalculateNPadded(N);
                 const index_t k_padded = GridwiseGemm::CalculateKPadded(K, K_BATCH);
-                const index_t k0       = GridwiseGemm::CalculateK0(K, K_BATCH);
+                const index_t k0_padded       = GridwiseGemm::CalculateK0Padded(K, K_BATCH);
 
                 const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, stride_c);
 
@@ -297,7 +297,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                                            m_padded,
                                            n_padded,
                                            k_padded,
-                                           k0,
+                                           k0_padded,
                                            K_BATCH};
 
                 gemm_kernel_args_.emplace_back(
@@ -321,7 +321,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                 auto& karg = gemm_kernel_args_[i].karg_;
 
                 const index_t k_padded = GridwiseGemm::CalculateKPadded(karg.K, K_BATCH);
-                const index_t k0       = GridwiseGemm::CalculateK0(karg.K, K_BATCH);
+                const index_t k0_padded       = GridwiseGemm::CalculateK0Padded(karg.K, K_BATCH);
 
                 const auto c_grid_desc_m_n =
                     GridwiseGemm::MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
@@ -340,7 +340,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                     GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
 
                 karg.KPadded                            = k_padded;
-                karg.K0                                 = k0;
+                karg.K0Padded                                 = k0_padded;
                 karg.k_batch                            = K_BATCH;
                 gemm_kernel_args_[i].block_2_ctile_map_ = grouped_block_2_ctile_map;
                 gemm_kernel_args_[i].block_start_       = block_start;
@@ -362,7 +362,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
     {
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-            index_t K0                       = arg.gemm_kernel_args_[0].karg_.K0;
+            index_t K0                       = arg.gemm_kernel_args_[0].karg_.K0Padded;
             bool all_have_kbatch_gt_one      = arg.gemm_kernel_args_[0].karg_.k_batch > 1;
             bool all_have_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
 
@@ -384,7 +384,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                     throw std::runtime_error(err.str());
                 }
 
-                K0 = karg.K0;
+                K0 = karg.K0Padded;
                 bool not_all_have_main_k0_block_loop_same =
                     all_have_main_k0_block_loop xor GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
                 bool not_all_have_kbatch_value_same = all_have_kbatch_gt_one xor (kbatch > 1);

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

@@ -120,6 +120,21 @@ void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances
                                                   PassThrough,
                                                   PassThrough>>>& instances);
 
+void add_device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F8,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+
+
 template <typename ALayout,
           typename BLayout,
           typename ELayout,
@@ -184,6 +199,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
             }
         }
+	else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+		if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+		    add_device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instances(op_ptrs);
+            }
+        }
         return op_ptrs;
     }
 };

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

@@ -7,4 +7,6 @@ add_instance_library(device_grouped_gemm_instance
    device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
    device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
    device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
+
+   device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp
 )

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp

@@ -37,15 +37,6 @@ using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple
         //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        // Currently AK1 must equal BK1 !
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,

profiler/src/CMakeLists.txt

@@ -40,7 +40,7 @@ set(PROFILER_SOURCES
   #list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
   #list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
   #list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
-  #list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
   #list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
 #endif()
 
@@ -107,7 +107,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance
   #target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
   #target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
   #target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
-  #target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
   #target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
 #endif()
 rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)

profiler/src/profile_grouped_gemm.cpp

@@ -27,6 +27,7 @@ enum struct GemmDataType
     F16_F16_F16,    // 1
     BF16_BF16_BF16, // 2
     INT8_INT8_INT8, // 3
+    F16_F8_F16,    // 4
 };
 
 #define OP_NAME "grouped_gemm"
@@ -56,7 +57,7 @@ int profile_grouped_gemm(int argc, char* argv[])
     {
         std::cout
             << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
-            << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
+            << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: fp16@fp8)\n"
             << "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
             << "                     1: A[m, k] * B[n, k] = C[m, n];\n"
             << "                     2: A[k, m] * B[k, n] = C[m, n];\n"
@@ -169,6 +170,26 @@ int profile_grouped_gemm(int argc, char* argv[])
                                                                                    StrideCs,
                                                                                    kbatch);
     }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::f8_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs,
+                                                                                   kbatch);
+    }
     else
     {
         throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");