Merge branch 'develop' into feature/check-window-lengths

example/15_grouped_gemm/grouped_gemm_multiple_d_xdl_fp16.cpp

@@ -63,7 +63,7 @@ using DeviceGemmInstance =
 //######|        |        |         |        |      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|
 //######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,    S<4, 64, 1>,     S<1, 0, 2>,      S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>;
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,    S<4, 64, 1>,     S<1, 0, 2>,      S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,       S<4,4,4>>;
 // clang-format on
 
 struct ProblemSize final
@@ -92,9 +92,10 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
     auto group_count = problem_size.group_count;
 
     using KernelArguments = ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDs>;
+    using GemmDesc        = ck::tensor_operation::device::GemmDesc;
 
     // GEMM shape
-    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<GemmDesc> gemm_descs;
     std::vector<KernelArguments> ggemm_kargs;
     std::vector<void*> p_Cs;
     std::vector<const void*> p_As;

example/26_contraction/README.md

@@ -7,14 +7,3 @@
 #arg3: time kernel (0=no, 1=yes)
 ./bin/example_contraction_bilinear_xdl_fp32 1 1 1
 ```
-
-Result (MI100 @ dynammic freq, 46TFlops peak FP32)
-```
-a_ms_ks: dim 4, lengths {30, 128, 32, 64}, strides {524288, 4096, 128, 1}
-b_ks_ns: dim 4, lengths {32, 64, 32, 64}, strides {128, 1, 524288, 4096}
-c_ms_ns: dim 4, lengths {30, 128, 32, 64}, strides {524288, 4096, 128, 1}
-launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
-Warm up 1 time
-Start running 10 times...
-Perf: 0.843286 ms, 38.1985 TFlops, 94.5014 GB/s, DeviceContractionMultipleD_Xdl_CShuffle<256, 256, 128, 16, 4, 4>
-```

example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp

@@ -83,14 +83,14 @@ using DeviceOpInstanceKKNN =
                                                                                   2,
                                                                                   4,
                                                                                   4,
-                                                                                  true,
+                                                                                  false,
                                                                                   S<4, 32, 1>,
                                                                                   S<1, 0, 2>,
                                                                                   S<1, 0, 2>,
                                                                                   2,
                                                                                   4,
                                                                                   4,
-                                                                                  true,
+                                                                                  false,
                                                                                   1,
                                                                                   1,
                                                                                   S<1, 64, 1, 2>,

example/30_grouped_conv_fwd_multiple_d/README.md

@@ -16,15 +16,3 @@ Following arguments (depending on number of spatial dims):
 ./bin/example_grouped_conv_fwd_bias_relu_add_xdl_fp16 1 1 1
 ```
 
-Result (MI100)
-```
-in: dim 5, lengths {1, 128, 192, 71, 71}, strides {192, 967872, 1, 13632, 192}
-wei: dim 5, lengths {1, 256, 192, 3, 3}, strides {442368, 1728, 1, 576, 192}
-bias: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
-residual: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
-out: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 331776, 1, 9216, 256}
-launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
-Warm up 1 time
-Start running 10 times...
-Perf: 1.55981 ms, 94.0927 TFlops, 213.868 GB/s, DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<256, 128, 256, 16, Default>
-```

example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp

@@ -2,6 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common_wmma.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 // kernel data types
 using InKernelDataType       = FP16;
@@ -23,4 +24,14 @@ using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
 
 #include "run_grouped_conv_fwd_bias_relu_add_wmma_example.inc"
 
-int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv);
+}

example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp

@@ -2,6 +2,7 @@
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common_wmma.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 // kernel data types
 using InKernelDataType       = I8;
@@ -23,4 +24,14 @@ using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
 
 #include "run_grouped_conv_fwd_bias_relu_add_wmma_example.inc"
 
-int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16.cpp

@@ -27,6 +27,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -163,4 +164,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
 
 #include "run_batched_gemm_scale_softmax_gemm_permute_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_wmma_fp16.cpp

@@ -27,6 +27,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -285,4 +286,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
 
 #include "run_batched_gemm_scale_softmax_gemm_permute_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/cross_attention_forward_wmma_fp16.cpp

@@ -27,6 +27,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -71,7 +72,7 @@ static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecial
 #define CK_MHA_USE_WAVE_1
 #define CK_MHA_USE_WAVE_2
 #define CK_MHA_USE_WAVE_4
-#define CK_MHA_USE_WAVE_8
+//#define CK_MHA_USE_WAVE_8
 using DeviceMHAFactory = 
     std::tuple<
 #ifdef CK_MHA_USE_WAVE_1
@@ -277,10 +278,10 @@ using DeviceMHAFactory =
             S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
             // CShuffleBlockTransfer MN
             1, 1, S<1, 64, 1, 2>, 8,             
-            MaskingSpec>,
+            MaskingSpec>
 #endif
 #ifdef CK_MHA_USE_WAVE_8
-        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
+        ,ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
             NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
             ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
             AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
@@ -351,4 +352,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
 
 #include "run_cross_attention_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/grouped_query_attention_forward_wmma_fp16.cpp

@@ -28,6 +28,7 @@ Example is GQA-4
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -299,4 +300,14 @@ using ReferenceGemm1Instance =
 
 #include "run_grouped_query_attention_forward_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/multi_query_attention_forward_wmma_fp16.cpp

@@ -26,6 +26,7 @@ Shazeer, Noam. “Fast Transformer Decoding: One Write-Head Is All You Need.”
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -284,4 +285,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm_
 
 #include "run_multi_query_attention_forward_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/32_batched_gemm_scale_softmax_gemm/self_attention_forward_wmma_fp16.cpp

@@ -27,6 +27,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_n = Softmax(A_g_m_k * B0_g
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -71,7 +72,7 @@ static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecial
 #define CK_MHA_USE_WAVE_1
 #define CK_MHA_USE_WAVE_2
 #define CK_MHA_USE_WAVE_4
-#define CK_MHA_USE_WAVE_8
+//#define CK_MHA_USE_WAVE_8
 using DeviceMHAFactory = 
     std::tuple<
 #ifdef CK_MHA_USE_WAVE_1
@@ -277,10 +278,10 @@ using DeviceMHAFactory =
             S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
             // CShuffleBlockTransfer MN
             1, 1, S<1, 64, 1, 2>, 8,             
-            MaskingSpec>,
+            MaskingSpec>
 #endif
 #ifdef CK_MHA_USE_WAVE_8
-        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
+        ,ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
             NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
             ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
             AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,
@@ -329,4 +330,14 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
 
 #include "run_self_attention_wmma.inc"
 
-int main(int argc, char* argv[]) { return run(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run(argc, argv);
+}

example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp

@@ -3,6 +3,7 @@
 
 #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp"
 #include "common.hpp"
+#include "ck/host_utility/device_prop.hpp"
 
 using OutDataType      = FP16;
 using WeiDataType      = FP16;
@@ -31,4 +32,14 @@ using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDat
 
 #include "run_grouped_conv_bwd_data_example.inc"
 
-int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    bool is_supported = ck::is_gfx11_supported();
+    if(!is_supported)
+    {
+        std::cout << "WARNING: wmma example not supported on the platform " << ck::get_device_name()
+                  << std::endl;
+        return 0;
+    }
+    return run_grouped_conv_bwd_data_example(argc, argv);
+}

example/46_gemm_add_multiply/README.md

@@ -8,19 +8,3 @@
 #arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
 ./bin/example_gemm_add_multiply_dl_fp16 1 1 1
 ```
-
-Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
-```
-a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
-b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
-d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
-d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-arg.a_grid_desc_k0_m0_m1_k1_{2048, 3840, 2}
-arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
-arg.e_grid_desc_m_n_{ 3840, 4096}
-launch_and_time_kernel: grid_dim {960, 1, 1}, block_dim {256, 1, 1}
-Warm up 1 time
-Start running 10 times...
-Perf: 3.99904 ms, 32.22 TFlops, 31.9913 GB/s, DeviceGemmMultipleD_Dl<256, 128, 128, 16, 2, 4, 4, 1>
-```

example/59_grouped_gemm_multi_ABD/CMakeLists.txt

 add_custom_target(example_grouped_gemm_xdl_multi_abd)
 
 add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16 grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp)
-add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16)
+add_example_dependencies(example_grouped_gemm_xdl_multi_abd example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16)
 
 add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8 grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp)
-add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8)
+add_example_dependencies(example_grouped_gemm_xdl_multi_abd example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8)

example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using A0DataType       = F8;
+using A1DataType       = F32;
+using B0DataType       = F8;
+using B1DataType       = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using EDataType        = F16;
+using ComputeDataType  = F8;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct Multiply
+{
+    __host__ __device__ constexpr void
+    operator()(ck::f8_t& a, const ck::f8_t& a0, const float& a1) const
+    {
+        a = ck::type_convert<ck::half_t>(ck::type_convert<float>(a0) * a1);
+    }
+};
+
+using AElementOp   = Multiply;
+using BElementOp   = Multiply;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultipleABD_Xdl_CShuffle<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ck::Tuple<A0DataType, A1DataType>,
+    ck::Tuple<B0DataType, B1DataType>,
+    AccDataType,
+    CShuffleDataType,
+    ck::Tuple<>,
+    EDataType,
+    AElementOp,
+    BElementOp,
+    CDEElementOp,
+    GemmSpec,
+    1,
+    256,
+    256,
+    128,
+    32,
+    8,
+    8,
+    32,
+    32,
+    4,
+    2,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    1,
+    8,
+    1,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    1,
+    1,
+    1,
+    S<1, 32, 1, 8>,
+    8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A0[M0, M1, K0, K1]
+    std::vector<ck::index_t> a0_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
+    // A1[M1, K1] -> A1[M0, M1, K0, K1]
+    std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 1, 0};
+    // B0[N0, N1, K0, K1]
+    std::vector<ck::index_t> b0_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b0_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
+    // B1[N0, N1, K0, K1]
+    std::vector<ck::index_t> b1_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b1_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<A0DataType> a0_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+    Tensor<A1DataType> a1_ms_ks(a1_ms_ks_lengths, a1_ms_ks_strides);
+    Tensor<B0DataType> b0_ns_ks(b0_ns_ks_lengths, b0_ns_ks_strides);
+    Tensor<B1DataType> b1_ns_ks(b1_ns_ks_lengths, b1_ns_ks_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a0_ms_ks: " << a0_ms_ks.mDesc << std::endl;
+    std::cout << "a1_ms_ks: " << a1_ms_ks.mDesc << std::endl;
+
+    std::cout << "b0_ns_ks: " << b0_ns_ks.mDesc << std::endl;
+    std::cout << "b1_ns_ks: " << b1_ns_ks.mDesc << std::endl;
+
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
+        b0_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_ns_ks.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    default:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
+        b0_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        b1_ns_ks.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(A1DataType) * a1_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_ms_ks.mData.data());
+    a1_device_buf.ToDevice(a1_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_ns_ks.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(
+        std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
+                                   a1_device_buf.GetDeviceBuffer()},
+        std::array<const void*, 2>{b0_device_buf.GetDeviceBuffer(),
+                                   b1_device_buf.GetDeviceBuffer()},
+        std::array<const void*, 0>{},
+        e_device_buf.GetDeviceBuffer(),
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_lengths, a1_ms_ks_lengths},
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_strides, a1_ms_ks_strides},
+        std::array<std::vector<ck::index_t>, 2>{b0_ns_ks_lengths, b1_ns_ks_lengths},
+        std::array<std::vector<ck::index_t>, 2>{b0_ns_ks_strides, b1_ns_ks_strides},
+        std::array<std::vector<ck::index_t>, 0>{},
+        std::array<std::vector<ck::index_t>, 0>{},
+        e_ms_ns_lengths,
+        e_ms_ns_strides,
+        a_element_op,
+        b_element_op,
+        PassThrough{});
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_contraction with the specified compilation parameters does "
+            "not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    if(time_kernel)
+    {
+        ck::index_t M =
+            ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+        ck::index_t N = ck::accumulate_n<ck::index_t>(
+            e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+        ck::index_t K = ck::accumulate_n<ck::index_t>(
+            a0_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+        std::size_t flop = std::size_t(2) * M * N * K;
+        std::size_t num_btype =
+            sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + +sizeof(EDataType) * M * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s" << std::endl;
+    }
+
+    if(do_verification)
+    {
+
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        Tensor<A0DataType> a_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+
+        for(size_t m0 = 0; m0 < a_ms_ks.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < a_ms_ks.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t k0 = 0; k0 < a_ms_ks.mDesc.GetLengths()[2]; ++k0)
+                {
+                    for(size_t k1 = 0; k1 < a_ms_ks.mDesc.GetLengths()[3]; ++k1)
+                    {
+                        a_element_op(a_ms_ks(m0, m1, k0, k1),
+                                     a0_ms_ks(m0, m1, k0, k1),
+                                     a1_ms_ks(m0, m1, k0, k1));
+                    }
+                }
+            }
+        }
+
+        Tensor<B0DataType> b_ns_ks(b0_ns_ks_lengths, b0_ns_ks_strides);
+
+        for(size_t n0 = 0; n0 < b_ns_ks.mDesc.GetLengths()[0]; ++n0)
+        {
+            for(size_t n1 = 0; n1 < b_ns_ks.mDesc.GetLengths()[1]; ++n1)
+            {
+                for(size_t k0 = 0; k0 < b_ns_ks.mDesc.GetLengths()[2]; ++k0)
+                {
+                    for(size_t k1 = 0; k1 < b_ns_ks.mDesc.GetLengths()[3]; ++k1)
+                    {
+                        b_element_op(b_ns_ks(n0, n1, k0, k1),
+                                     b0_ns_ks(n0, n1, k0, k1),
+                                     b1_ns_ks(n0, n1, k0, k1));
+                    }
+                }
+            }
+        }
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      A0DataType,
+                                                                      B0DataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      ComputeDataType,
+                                                                      PassThrough,
+                                                                      PassThrough>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
+        auto ref_argument = ref_op.MakeArgument(
+            a_ms_ks, b_ns_ks, c_ms_ns_host_result, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/62_convnd_activ/CMakeLists.txt

 add_subdirectory(binary)
+add_subdirectory(convinvscale)
+add_subdirectory(convscale)
+add_subdirectory(convscale_relu)
 add_subdirectory(multi_AB)
 add_subdirectory(unary)
 

example/62_convnd_activ/convinvscale/CMakeLists.txt

+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+      add_custom_target(example_convnd_activ_xdl_convinvscale)
+      add_example_executable(example_convnd_fwd_xdl_convinvscale_fp8 convnd_fwd_xdl_convinvscale_fp8.cpp)
+      add_example_dependencies(example_convnd_activ_xdl_convinvscale example_convnd_fwd_xdl_convinvscale_fp8)
+      set(target 1)
+ endif()
+endforeach()
\ No newline at end of file

example/62_convnd_activ/convinvscale/convnd_fwd_convinvscale_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ConvInvscale = ck::tensor_operation::element_wise::ConvInvscale;
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetFlops(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths,
+         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths,
+         const std::size_t& ds_size)
+{
+    // G * N * C * <output spatial lengths product> * (2 * K * <filter spatial lengths product> +
+    // <number of scale factors>)
+    ck::index_t G = weights_lengths[0];
+    ck::index_t N = output_lengths[1];
+    ck::index_t K = weights_lengths[1];
+    ck::index_t C = weights_lengths[2];
+
+    return G * N * C *
+           std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           (static_cast<std::size_t>(2) * K *
+                std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
+                                std::end(weights_lengths),
+                                static_cast<std::size_t>(1),
+                                std::multiplies<>()) +
+            ds_size);
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<CShuffleDataType> c(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // random scale values
+    float scale_in  = float(std::rand()) / float(RAND_MAX);
+    float scale_wei = float(std::rand()) / float(RAND_MAX);
+    float scale_out = float(std::rand()) / float(RAND_MAX);
+
+    // initialize out_element_op for each iteration
+    const auto out_element_op = OutElementOp{scale_in, scale_wei, scale_out};
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t ds_size   = 3; // 3 element-wise scale multipliers
+    std::size_t flop      = GetFlops<NDimSpatial>(e_g_n_k_wos_lengths, b_g_k_c_xs_lengths, ds_size);
+    std::size_t num_btype = conv_param.GetInputByte<InDataType>() +
+                            conv_param.GetWeightByte<WeiDataType>() + sizeof(float) +
+                            sizeof(float) + sizeof(float) + conv_param.GetOutputByte<OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     CShuffleDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     PassThrough>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        out_host.ForEach([&](auto&, auto idx) { out_element_op(out_host(idx), c(idx)); });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device,
+                                    out_host,
+                                    "Error: incorrect results!",
+                                    get_rtol<OutDataType>(),
+                                    get_atol<OutDataType>());
+    }
+
+    return true;
+}

example/62_convnd_activ/convinvscale/convnd_fwd_xdl_convinvscale_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_convinvscale_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = ck::f8_t;
+using WeiDataType      = ck::f8_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using DsDataType       = ck::Tuple<>;
+using OutDataType      = ck::f8_t;
+using AComputeDataType = ck::f8_t;
+using BComputeDataType = ck::f8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ConvInvscale;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        DsLayout,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8,
+        AComputeDataType,
+        BComputeDataType>;
+
+#include "run_convnd_fwd_convinvscale_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }