COMPMID-1785: Support for 4D tensor in CLFlattenLayerKernel

With this patch we are able to dispatch a single GPU job also in case of
batched-flatten

Change-Id: I755e7af29d44b24f67fa04bad3c9b7646e8deefc

2 files changed, 40 insertions, 25 deletions

diff --git a/src/core/CL/cl_kernels/flatten.cl b/src/core/CL/cl_kernels/flatten.cl
index df0f9c4886..02694f709e 100644
--- a/src/core/CL/cl_kernels/flatten.cl
+++ b/src/core/CL/cl_kernels/flatten.cl
@@ -23,12 +23,13 @@
  */
 #include "helpers.h"
 
-#if defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT)
+#if defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(SRC_DEPTH)
 
 /** This opencl kernel flattens the first 3 dimensions of the input tensor
  *
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
- * @note The width and height of the input tensor must be passed at compile time using -DSRC_WIDTH and -DSRC_HEIGHT. e.g. -DSRC_WIDTH=24, -DSRC_HEIGHT=24
+ * @note The width, height and depth of the input tensor must be passed at compile time using -DSRC_WIDTH, -DSRC_HEIGHT and -DSRC_DEPTH. e.g. -DSRC_WIDTH=24, -DSRC_HEIGHT=24, -DSRC_DEPTH=16
+ * @note If the output has 3 dimensions, the 2nd dimension of the output tensor must be passed at compile time using -DDST_DIM1. e.g -DDST_DIM1=3
  *
  * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32
  * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
@@ -37,20 +38,38 @@
  * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
  * @param[in]  src_step_z                        src_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes)
+ * @param[in]  src_step_w                        src_stride_w * number of elements along Y processed per workitem(in bytes)
  * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
  * @param[out] dst_ptr                           Pointer to the destination tensor. Same as @p src_ptr
  * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
  * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
  * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
  */
 __kernel void flatten(
-    TENSOR3D_DECLARATION(src),
-    VECTOR_DECLARATION(dst))
+    TENSOR4D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst))
 {
-    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor4D src = CONVERT_TO_TENSOR4D_STRUCT(src, SRC_DEPTH);
 
-    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) + get_global_id(1) * (int)SRC_WIDTH + get_global_id(2) * (int)(SRC_WIDTH * SRC_HEIGHT)) * sizeof(
-                                     DATA_TYPE);
+    uint c  = get_global_id(2) % SRC_DEPTH; // input feature map
+    uint b0 = get_global_id(2) / SRC_DEPTH; // batch id
+    uint b1 = 0;
+
+#if defined(DST_DIM1)
+    uint b_tmp = b0;
+    b0 = b_tmp % DST_DIM1; // batch id0
+    b1 = b_tmp / DST_DIM1; // batch id1
+#endif // defined(DST_DIM1)
+
+    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes +
+                                 (get_global_id(0) + get_global_id(1) * (uint)SRC_WIDTH + c * (uint)(SRC_WIDTH * SRC_HEIGHT)) * sizeof(DATA_TYPE) +
+                                 b0 * dst_stride_y +
+                                 b1 * dst_stride_z;
 
     *((__global DATA_TYPE *)output_ptr) = *((__global DATA_TYPE *)src.ptr);
 }
diff --git a/src/core/CL/kernels/CLFlattenLayerKernel.cpp b/src/core/CL/kernels/CLFlattenLayerKernel.cpp
index 17189143ef..5c3856839a 100644
--- a/src/core/CL/kernels/CLFlattenLayerKernel.cpp
+++ b/src/core/CL/kernels/CLFlattenLayerKernel.cpp
@@ -90,19 +90,21 @@ void CLFlattenLayerKernel::configure(const ICLTensor *input, ICLTensor *output)
     _input  = input;
     _output = output;
 
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
     build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
+    build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(input->info()->dimension(2)));
+    build_opts.add_option_if(output->info()->num_dimensions() > 2, "-DDST_DIM1=" + support::cpp11::to_string(output->info()->dimension(1)));
 
     // Create kernel
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("flatten", build_opts.options()));
 
-    // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
-
     // Set config_id for enabling LWS tuning
     _config_id = "flatten";
     _config_id += "_";
@@ -131,21 +133,15 @@ void CLFlattenLayerKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
 
-    Window out_window;
-    out_window.use_tensor_dimensions(_output->info()->tensor_shape());
+    Window collapsed_window = window.collapse(ICLKernel::window(), Window::DimZ);
 
-    Window out_slice = out_window.first_slice_window_1D();
-    Window in_slice  = window.first_slice_window_3D();
+    Window output_window;
+    output_window.use_tensor_dimensions(_output->info()->tensor_shape());
 
     // Run kernel
-    do
-    {
-        // Set arguments
-        unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, in_slice);
-        add_1D_tensor_argument(idx, _output, out_slice);
-        enqueue(queue, *this, in_slice, lws_hint());
-    }
-    while(window.slide_window_slice_3D(in_slice) && out_window.slide_window_slice_1D(out_slice));
+    unsigned int idx = 0;
+    add_4D_tensor_argument(idx, _input, collapsed_window);
+    add_3D_tensor_argument(idx, _output, output_window);
+    enqueue(queue, *this, collapsed_window, lws_hint());
 }
 } // namespace arm_compute