Merge branch 'develop' into codegen_hiprtc

include/ck/library/utility/host_tensor_generator.hpp

@@ -81,6 +81,20 @@ struct GeneratorTensor_1<int8_t>
     }
 };
 
+template <>
+struct GeneratorTensor_1<ck::pk_i4_t>
+{
+    int8_t value = 1;
+
+    template <typename... Is>
+    ck::pk_i4_t operator()(Is...)
+    {
+        int t         = value + 8;
+        ck::pk_i4_t r = ((t << 4) + t) & 0xff;
+        return r;
+    }
+};
+
 template <typename T>
 struct GeneratorTensor_2
 {
@@ -121,6 +135,22 @@ struct GeneratorTensor_2<int8_t>
     }
 };
 
+template <>
+struct GeneratorTensor_2<ck::pk_i4_t>
+{
+    int min_value = 0;
+    int max_value = 1;
+
+    template <typename... Is>
+    ck::pk_i4_t operator()(Is...)
+    {
+        int hi        = std::rand() % (max_value - min_value) + min_value + 8;
+        int lo        = std::rand() % (max_value - min_value) + min_value + 8;
+        ck::pk_i4_t r = ((hi << 4) + lo) & 0xff;
+        return r;
+    }
+};
+
 #if defined CK_ENABLE_FP8
 template <>
 struct GeneratorTensor_2<ck::f8_t>

include/ck/tensor/static_tensor.hpp

@@ -167,7 +167,7 @@ struct StaticTensorTupleOfVectorBuffer
     // Idx is for S, not X. Idx should be aligned with X
     template <typename X,
               typename Idx,
-              typename enable_if<has_same_scalar_type<S, X>::value &&
+              typename enable_if<(has_same_scalar_type<S, X>::value || !is_native_type<S>()) &&
                                      is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
                                  bool>::type = false>
     __host__ __device__ constexpr X GetAsType(Idx) const
@@ -201,7 +201,7 @@ struct StaticTensorTupleOfVectorBuffer
     // Idx is for S, not X. Idx should be aligned with X
     template <typename X,
               typename Idx,
-              typename enable_if<has_same_scalar_type<S, X>::value &&
+              typename enable_if<(has_same_scalar_type<S, X>::value || !is_native_type<S>()) &&
                                      is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
                                  bool>::type = false>
     __host__ __device__ constexpr void SetAsType(Idx, X x)

include/ck/tensor_operation/gpu/device/device_base.hpp

@@ -6,6 +6,8 @@
 #ifndef CK_CODE_GEN_RTC
 #include <string>
 #include <sstream>
+#include <regex>
+#include <optional>
 #include "ck/stream_config.hpp"
 #endif
 
@@ -13,6 +15,36 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
+#ifndef CK_CODE_GEN_RTC
+#define GET_OBJECT_NAME_IMLP                                                  \
+    std::optional<std::string> GetObjectName() const override                 \
+    {                                                                         \
+        std::string str = __PRETTY_FUNCTION__;                                \
+        static std::regex obj_name_expr{"<std::string> (.*)::GetObjectName"}; \
+        std::smatch match;                                                    \
+        if(!std::regex_search(str, match, obj_name_expr))                     \
+        {                                                                     \
+            return str;                                                       \
+        }                                                                     \
+        return std::string(match[1]) + ';';                                   \
+    }
+
+#define GET_TEMPLATE_INFO_IMPL                                  \
+    std::optional<std::string> GetTemplateInfo() const override \
+    {                                                           \
+        std::string str = __PRETTY_FUNCTION__;                  \
+        static std::regex template_expr{"\\[(.*)\\]"};          \
+        std::smatch match;                                      \
+        if(!std::regex_search(str, match, template_expr))       \
+        {                                                       \
+            return std::nullopt;                                \
+        }                                                       \
+        return std::string(match[1]);                           \
+    }
+
+#define REGISTER_EXTRA_PRINTING_METHODS GET_OBJECT_NAME_IMLP GET_TEMPLATE_INFO_IMPL
+#endif
+
 #ifndef CK_CODE_GEN_RTC
 struct BaseArgument
 {
@@ -51,6 +83,10 @@ struct BaseOperator
 
     virtual std::string GetTypeIdName() const { return typeid(*this).name(); }
 
+    virtual std::optional<std::string> GetObjectName() const { return std::nullopt; }
+
+    virtual std::optional<std::string> GetTemplateInfo() const { return std::nullopt; }
+
     virtual std::string GetTypeIdHashCode() const
     {
         std::ostringstream oss;

include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp

@@ -89,7 +89,8 @@ struct DeviceBatchedGemmV2MultiD : public BaseOperator
                         index_t BatchStrideE,
                         AElementwiseOperation a_element_op,
                         BElementwiseOperation b_element_op,
-                        CDEElementwiseOperation cde_element_op) = 0;
+                        CDEElementwiseOperation cde_element_op,
+                        index_t KBatch) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };

include/ck/tensor_operation/gpu/device/device_gemm_v2.hpp

@@ -36,6 +36,10 @@ struct DeviceGemmV2 : public BaseOperator
                         CElementwiseOperation c_element_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual bool GetPermuteA()         = 0;
+    virtual bool GetPermuteB()         = 0;
+    virtual ck::index_t GetKPerBlock() = 0;
 };
 
 template <typename ALayout,
@@ -73,6 +77,43 @@ struct DeviceGemmV2R1 : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename BScaleType,
+          typename CDataType,
+          index_t ScaleBlockN,
+          index_t ScaleBlockK,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmV2BScale : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        ck::index_t StrideScaleB,
+                        const void* p_b_scale,
+                        ck::index_t KSplit,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual bool GetPermuteB()         = 0;
+    virtual ck::index_t GetKPerBlock() = 0;
+};
+
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp

@@ -469,7 +469,11 @@ struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout
         {
             return false;
         }
-
+        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> &&
+           arg.Streamk_sel > 0)
+        {
+            return false;
+        }
         if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
                                                        GemmSpec == GemmSpecialization::NKPadding ||
                                                        GemmSpec == GemmSpecialization::MNKPadding ||

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

@@ -64,7 +64,9 @@ template <typename ALayout,
           BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
           BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
           typename ComputeTypeA                       = CDataType,
-          typename ComputeTypeB                       = ComputeTypeA>
+          typename ComputeTypeB                       = ComputeTypeA,
+          bool PermuteA                               = false,
+          bool PermuteB                               = false>
 struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
                                                        BLayout,
                                                        CLayout,
@@ -122,7 +124,9 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
         BlkGemmPipeSched,
         BlkGemmPipelineVer,
         ComputeTypeA,
-        ComputeTypeB>;
+        ComputeTypeB,
+        PermuteA,
+        PermuteB>;
 
     using Argument = typename GridwiseGemm::Argument;
 
@@ -633,6 +637,11 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
         return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
     }
 
+    index_t GetKPerBlock() override { return KPerBlock; }
+
+    bool GetPermuteA() override { return PermuteA; }
+    bool GetPermuteB() override { return PermuteB; }
+
     static auto MakeArgument(const ADataType* p_a,
                              const BDataType* p_b,
                              CDataType* p_c,
@@ -729,6 +738,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
 
         return str.str();
     }
+    REGISTER_EXTRA_PRINTING_METHODS
 };
 
 } // namespace device