updated instance list for client example, added different layout example

example/44_elementwise_permute/CMakeLists.txt

@@ -2,3 +2,4 @@ add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4
 add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
 add_example_executable(example_elementwise_permute elementwise_permute.cpp)
 add_example_executable(example_elementwise_permute_3d elementwise_permute_3d.cpp)
+add_example_executable(example_elementwise_permute_3d_ndhwc elementwise_permute_3d_ndhwc.cpp)

example/44_elementwise_permute/elementwise_permute_3d_ndhwc.cpp

+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.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"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwise3dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
+                                                          ck::Tuple<BDataType>, // OutDataTypeTuple
+                                                          PassThrough,          // ElementwiseOp
+                                                          2,                    // NumDim_m, {N, C}
+                                                          2,                    // NumDim_n, {H, W}
+                                                          1,                    // NumDim_k, {D}
+                                                          8,                    // MPerThread
+                                                          8,                    // NPerThread
+                                                          8,                    // KPerThread
+                                                          ck::Sequence<1>,  // InScalarPerVectorSeq
+                                                          ck::Sequence<1>>; // OutScalarPerVectorSeq
+
+template <typename HostTensorA, typename HostTensorB, typename Functor>
+void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
+{
+    for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
+        for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
+            for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
+                for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
+                    for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
+                    {
+                        auto a_val = A_ncdhw(n, c, d, h, w);
+                        functor(B_ndhwc(n, d, h, w, c), a_val);
+                    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    const int N = 4;
+    const int C = 16;
+    const int H = 32;
+    const int W = 5;
+    const int D = 16;
+
+    std::vector<std::size_t> ncdhw = {N, C, D, H, W};
+    std::vector<std::size_t> ndhwc = {N, D, H, W, C};
+    Tensor<ADataType> a(ncdhw);
+    Tensor<BDataType> b(ndhwc);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
+    std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
+    std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, C, H, W, D
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(
+        ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (ncdhw): " << a.mDesc << std::endl;
+    std::cout << "B (ndhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    std::size_t flop = std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
+
+    std::size_t num_btype =
+        sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
+        sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
+
+    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;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        Tensor<BDataType> host_b(ndhwc);
+        host_elementwise4D(host_b, a, PassThrough{});
+
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

library/include/ck/library/tensor_operation_instance/gpu/transpose/device_transpose_instance.hpp

@@ -23,14 +23,9 @@ using S = ck::Sequence<Is...>;
 using device_transpose_f16_instances = std::tuple<
     // FOR 16, 32, 16, 32, 16
     // clang-format off
-    DeviceElementwise3dImpl<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 2, 2, 1, 8, 8, 8, ck::Sequence<8>, ck::Sequence<8>>,
-    DeviceElementwise3dImpl<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 2, 2, 1, 8, 8, 8, ck::Sequence<8>, ck::Sequence<1>>,
-    DeviceElementwise3dImpl<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 2, 2, 1, 8, 8, 8, ck::Sequence<1>, ck::Sequence<8>>,
     DeviceElementwise3dImpl<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 2, 2, 1, 8, 8, 8, ck::Sequence<1>, ck::Sequence<1>>,
     DeviceElementwise3dImpl<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 2, 2, 1, 8, 1, 1, ck::Sequence<1>, ck::Sequence<1>>,
-    DeviceElementwise3dImpl<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 2, 2, 1, 8, 1, 1, ck::Sequence<8>, ck::Sequence<1>>,
     DeviceElementwise3dImpl<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 2, 2, 1, 8, 4, 4, ck::Sequence<1>, ck::Sequence<1>>,
-    DeviceElementwise3dImpl<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 2, 2, 1, 8, 4, 4, ck::Sequence<8>, ck::Sequence<8>>
     // clang-format on
     >;
 
@@ -38,13 +33,8 @@ using device_transpose_f32_instances = std::tuple<
     // for 16, 8, 16, 32, 8 -> test with instances for fp16
     // clang-format off
     DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 4, 4, ck::Sequence<1>, ck::Sequence<1>>,
-    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 4, 4, ck::Sequence<8>, ck::Sequence<1>>,
-    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 4, 4, ck::Sequence<8>, ck::Sequence<8>>,
-    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 8, 4, ck::Sequence<8>, ck::Sequence<8>>,
-    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 8, 8, ck::Sequence<8>, ck::Sequence<8>>,
-    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 8, 8, ck::Sequence<4>, ck::Sequence<8>>,
-    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 8, 8, ck::Sequence<4>, ck::Sequence<4>>,
-    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 8, 8, ck::Sequence<8>, ck::Sequence<4>>,
+    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 8, 4, ck::Sequence<1>, ck::Sequence<1>>,
+    DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 8, 8, ck::Sequence<1>, ck::Sequence<1>>,
     DeviceElementwise3dImpl<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 2, 2, 1, 4, 4, 8, ck::Sequence<8>, ck::Sequence<8>>
     // clang-format on
     >;