Merge branch 'develop' into amd-develop

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

@@ -11,7 +11,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 namespace ck {
 namespace tensor_operation {
 namespace device {

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

@@ -17,7 +17,7 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-#if defined(__fp16__) && defined(DL_KERNELS)
+#if defined(CK_ENABLE_FP16) && defined(DL_KERNELS)
 void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
@@ -78,7 +78,7 @@ void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(
         DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#if defined(__fp32__) && defined(DL_KERNELS)
+#if defined(CK_ENABLE_FP32) && defined(DL_KERNELS)
 void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
@@ -99,7 +99,7 @@ void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(
         DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#if defined(__int8__) && defined(DL_KERNELS)
+#if defined(CK_ENABLE_INT8) && defined(DL_KERNELS)
 void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
@@ -140,7 +140,7 @@ void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(
         DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
 void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
@@ -161,7 +161,7 @@ void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
         DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
@@ -208,7 +208,7 @@ void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
         instances);
 
 #endif
-#ifdef __bf16__
+#ifdef CK_ENABLE_BF16
 void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
@@ -229,7 +229,7 @@ void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
         DeviceGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#ifdef __fp32__
+#ifdef CK_ENABLE_FP32
 void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
@@ -270,7 +270,7 @@ void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
         DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#ifdef __fp64__
+#ifdef CK_ENABLE_FP64
 void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
     std::vector<std::unique_ptr<
 
@@ -363,7 +363,7 @@ struct DeviceOperationInstanceFactory<
                 add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
             }
         }
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
         else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
                           is_same_v<CDataType, half_t>)
         {
@@ -414,7 +414,7 @@ struct DeviceOperationInstanceFactory<
             }
         }
 #endif
-#ifdef __bf16__
+#ifdef CK_ENABLE_BF16
         else if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
                           is_same_v<CDataType, ck::bhalf_t>)
         {
@@ -440,7 +440,7 @@ struct DeviceOperationInstanceFactory<
             }
         }
 #endif
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
         else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
                           is_same_v<CDataType, int8_t>)
         {

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

@@ -9,7 +9,7 @@
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 namespace ck {
 namespace tensor_operation {
 namespace device {

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

@@ -11,7 +11,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 namespace ck {
 namespace tensor_operation {
 namespace device {

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

@@ -10,7 +10,7 @@
 #include "ck/ck.hpp"
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 namespace ck {
 namespace tensor_operation {
 namespace device {

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

@@ -57,6 +57,46 @@ void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
         DeviceGemmSplitK<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
+void add_device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
 template <typename ADataType,
           typename BDataType,
           typename CDataType,
@@ -136,6 +176,54 @@ struct DeviceOperationInstanceFactory<
                 add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs);
             }
         }
+        else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f8_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<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
 
         return op_ptrs;
     }

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

@@ -16,7 +16,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 void add_device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemmStreamK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&

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

@@ -10,7 +10,7 @@
 #include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 namespace ck {
 namespace tensor_operation {
 namespace device {

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

@@ -16,7 +16,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 // FP16
 void add_device_normalization_rank_2_1_f16_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 2, 1>>>&);
@@ -27,7 +27,7 @@ void add_device_normalization_rank_4_3_f16_instances(
 void add_device_normalization_rank_5_3_f16_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 5, 3>>>&);
 #endif
-#ifdef __fp32__
+#ifdef CK_ENABLE_FP32
 // FP32
 void add_device_normalization_rank_2_1_f32_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, PassThrough, 2, 1>>>&);
@@ -66,7 +66,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
         if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
                      is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16>)
         {
@@ -84,7 +84,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal
             }
         }
 #endif
-#ifdef __fp32__
+#ifdef CK_ENABLE_FP32
         if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> &&
                      is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32>)
         {

library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp

@@ -22,7 +22,7 @@ static constexpr auto WindowRank = 3;
 
 static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
 static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 // FP16
 void add_device_pool3d_fwd_ndhwc_f16_instances(
     std::vector<std::unique_ptr<
@@ -37,7 +37,7 @@ void add_device_pool3d_fwd_ndhwc_index_f16_instances(
     std::vector<std::unique_ptr<
         DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NDHWC, NDHWC, MaxOp, true>>>&);
 #endif
-#ifdef __fp32__
+#ifdef CK_ENABLE_FP32
 // FP32
 void add_device_pool3d_fwd_ndhwc_f32_instances(
     std::vector<std::unique_ptr<
@@ -84,7 +84,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFw
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
         if constexpr(is_same_v<InLayout, NDHWC> && is_same_v<OutLayout, NDHWC>)
         {
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
             if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
                          is_same_v<IndexDataType, I32>)
             {
@@ -98,7 +98,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFw
                 }
             }
 #endif
-#ifdef __fp32__
+#ifdef CK_ENABLE_FP32
             if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
                          is_same_v<IndexDataType, I32>)
             {

library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp

@@ -11,7 +11,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
 namespace ck {
 namespace tensor_operation {
 namespace device {

library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp

@@ -11,7 +11,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
 namespace ck {
 namespace tensor_operation {
 namespace device {

library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp

@@ -11,7 +11,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
 namespace ck {
 namespace tensor_operation {
 namespace device {

library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp

@@ -11,7 +11,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
 namespace ck {
 namespace tensor_operation {
 namespace device {

library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp

@@ -11,7 +11,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
 namespace ck {
 namespace tensor_operation {
 namespace device {

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp

@@ -89,13 +89,13 @@ void add_device_reduce_instance_blockwise(
 {
     static_for<0, std::tuple_size<reduce_configuration_1_instances_blockwise>::value, 1>{}(
         [&](auto i) {
-            using cfg1 = remove_cvref_t<decltype(
-                std::get<i.value>(reduce_configuration_1_instances_blockwise{}))>;
+            using cfg1 = remove_cvref_t<decltype(std::get<i.value>(
+                reduce_configuration_1_instances_blockwise{}))>;
 
             static_for<0, std::tuple_size<reduce_configuration_2_instances_blockwise>::value, 1>{}(
                 [&](auto j) {
-                    using cfg2 = remove_cvref_t<decltype(
-                        std::get<j.value>(reduce_configuration_2_instances_blockwise{}))>;
+                    using cfg2 = remove_cvref_t<decltype(std::get<j.value>(
+                        reduce_configuration_2_instances_blockwise{}))>;
 
                     using ReduceOpInstance =
                         DeviceReduceMultiBlock<InDataType,

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp

@@ -90,14 +90,14 @@ void add_device_reduce_instance_multiblock_atomic_add(
     static_for<0,
                std::tuple_size<reduce_configuration_1_instances_multiblock_atomic_add>::value,
                1>{}([&](auto i) {
-        using cfg1 = remove_cvref_t<decltype(
-            std::get<i.value>(reduce_configuration_1_instances_multiblock_atomic_add{}))>;
+        using cfg1 = remove_cvref_t<decltype(std::get<i.value>(
+            reduce_configuration_1_instances_multiblock_atomic_add{}))>;
 
         static_for<0,
                    std::tuple_size<reduce_configuration_2_instances_multiblock_atomic_add>::value,
                    1>{}([&](auto j) {
-            using cfg2 = remove_cvref_t<decltype(
-                std::get<j.value>(reduce_configuration_2_instances_multiblock_atomic_add{}))>;
+            using cfg2 = remove_cvref_t<decltype(std::get<j.value>(
+                reduce_configuration_2_instances_multiblock_atomic_add{}))>;
 
             using ReduceOpInstance = DeviceReduceMultiBlock<InDataType,
                                                             AccDataType,

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp

@@ -77,8 +77,8 @@ void add_device_reduce_instance_threadwise(
 
     static_for<0, std::tuple_size<reduce_configuration_2_instances_threadwise>::value, 1>{}(
         [&](auto j) {
-            using cfg2 = remove_cvref_t<decltype(
-                std::get<j.value>(reduce_configuration_2_instances_threadwise{}))>;
+            using cfg2 = remove_cvref_t<decltype(std::get<j.value>(
+                reduce_configuration_2_instances_threadwise{}))>;
 
             using ReduceOpInstance = DeviceReduceThreadWise<InDataType,
                                                             AccDataType,

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

@@ -40,7 +40,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceSoftma
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
         if constexpr(std::is_same_v<InDataType, F16> && std::is_same_v<AccDataType, F32> &&
                      std::is_same_v<OutDataType, F16>)
         {
@@ -66,7 +66,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceSoftma
             }
         }
 #endif
-#ifdef __fp32__
+#ifdef CK_ENABLE_FP32
         if constexpr(std::is_same_v<InDataType, F32> && std::is_same_v<AccDataType, F32> &&
                      std::is_same_v<OutDataType, F32>)
         {