COMPMID-403:Add support for 7x7 pooling on CL.

Change-Id: I3c2c8d7e8e61d7737170cb1568900ce4ac337068
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78181
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>

3 files changed, 99 insertions, 16 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 15a5d90835..3070d4817e 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -219,6 +219,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "pixelwise_mul_int", "pixelwise_mul_int.cl" },
     { "pooling_layer_2", "pooling_layer.cl" },
     { "pooling_layer_3", "pooling_layer.cl" },
+    { "pooling_layer_7", "pooling_layer.cl" },
     { "remap_nearest_neighbour", "remap.cl" },
     { "remap_bilinear", "remap.cl" },
     { "reshape_to_columns", "convolution_layer.cl" },
diff --git a/src/core/CL/cl_kernels/pooling_layer.cl b/src/core/CL/cl_kernels/pooling_layer.cl
index 1902df9b7d..6bdb174235 100644
--- a/src/core/CL/cl_kernels/pooling_layer.cl
+++ b/src/core/CL/cl_kernels/pooling_layer.cl
@@ -41,7 +41,6 @@ float calculate_avg_scale(const int pool_size, const int upper_bound_w, const in
 
 /** Performs a pooling function of pool size equal to 2.
  *
- * @note Pooling stride must be passed using -DPOOL_STRIDE e.g -DPOOL_STRIDE=2. Supported strides are 1,2,3
  * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32;
  * @note In case of average pooling -DPOOL_AVG must be provided otherwise max pooling will be performed.
  *
@@ -88,7 +87,7 @@ __kernel void pooling_layer_2(
     data0         = POOL_OP(data0, data1);
     DATA_TYPE res = POOL_OP(data0.s0, data0.s1);
 
-    // Divide by 4 in case of average pooling
+    // Divide by pool region in case of average pooling
 #ifdef POOL_AVG
     res *= calculate_avg_scale(2, max_dims.x, max_dims.y, paddings.x, paddings.y, strides.x, strides.y);
 #endif
@@ -99,7 +98,6 @@ __kernel void pooling_layer_2(
 
 /** Performs a pooling function of pool size equal to 3.
  *
- * @note Pooling stride must be passed using -DPOOL_STRIDE e.g -DPOOL_STRIDE=2. Supported strides are 1,2,3
  * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32;
  * @note In case of average pooling -DPOOL_AVG must be provided otherwise max pooling will be performed.
  *
@@ -149,7 +147,7 @@ __kernel void pooling_layer_3(
     data0         = POOL_OP(data0, data2);
     DATA_TYPE res = POOL_OP(POOL_OP(data0.s0, data0.s1), data0.s2);
 
-    // Divide by 4 in case of average pooling
+    // Divide by pool region in case of average pooling
 #ifdef POOL_AVG
     res *= calculate_avg_scale(3, max_dims.x, max_dims.y, paddings.x, paddings.y, strides.x, strides.y);
 #endif
@@ -157,3 +155,88 @@ __kernel void pooling_layer_3(
     // Store result
     *(__global DATA_TYPE *)output.ptr = res;
 }
+
+/** Performs a pooling function of pool size equal to 7.
+ *
+ * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32;
+ * @note In case of average pooling -DPOOL_AVG must be provided otherwise max pooling will be performed.
+ *
+ * @param[in]  input_ptr                            Pointer to the source image. Supported data types: F16, F32
+ * @param[in]  input_stride_x                       Stride of the source image in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source image in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source image
+ * @param[out] output_ptr                           Pointer to the destination image. Supported data types: F16, F32
+ * @param[in]  output_stride_x                      Stride of the destination image in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination image
+ * @param[in]  max_dims                             The maximum index that can be accessed in x and y dimension (width + pad)
+ * @param[in]  strides                              The pooling operation strides in each dimension
+ * @param[in]  paddings                             The pooling operation paddings in each dimension
+ */
+__kernel void pooling_layer_7(
+    TENSOR3D_DECLARATION(input),
+    TENSOR3D_DECLARATION(output)
+#ifdef POOL_AVG
+    ,
+    int2 max_dims, int2 strides, int2 paddings
+#endif
+)
+{
+    // Get pixels pointer
+    Tensor3D input  = CONVERT_TO_TENSOR3D_STRUCT(input);
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+
+    // Load data
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    data0 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0));
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    data1 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0));
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    data2 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0));
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    data3 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 3, 0));
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    data4 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 4, 0));
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    data5 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 5, 0));
+    VEC_DATA_TYPE(DATA_TYPE, 8)
+    data6 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 6, 0));
+
+    // Pool operation of all rows
+    data0 = POOL_OP(data0, data1);
+    data2 = POOL_OP(data2, data3);
+    data4 = POOL_OP(data4, data5);
+    data0 = POOL_OP(data0, data2);
+    data4 = POOL_OP(data4, data6);
+    data0 = POOL_OP(data0, data4);
+
+    // Set last element
+#ifdef POOL_AVG
+    data0.s7 = 0;
+#else
+    data0.s7 = data0.s6;
+#endif
+
+    // Reduce result
+    VEC_DATA_TYPE(DATA_TYPE, 4)
+    reduce4 = POOL_OP(data0.s0123, data0.s4567);
+    VEC_DATA_TYPE(DATA_TYPE, 2)
+    reduce2       = POOL_OP(reduce4.s01, reduce4.s23);
+    DATA_TYPE res = POOL_OP(reduce2.s0, reduce2.s1);
+
+    // Divide by pool region in case of average pooling
+#ifdef POOL_AVG
+    res *= calculate_avg_scale(7, max_dims.x, max_dims.y, paddings.x, paddings.y, strides.x, strides.y);
+#endif
+
+    // Store result
+    *(__global DATA_TYPE *)output.ptr = res;
+}
diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
index dc5ae4ec7a..7648025caa 100644
--- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp
+++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
@@ -65,10 +65,13 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     std::tie(pool_pad_x, pool_pad_y)       = pad_stride_info.pad();
     std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride();
 
+    static const std::set<int> supported_pool_sizes = { 2, 3, 7 };
+    ARM_COMPUTE_UNUSED(supported_pool_sizes);
+
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_ERROR_ON(2 != pool_size && 3 != pool_size);
+    ARM_COMPUTE_ERROR_ON(supported_pool_sizes.find(pool_size) == supported_pool_sizes.end());
     ARM_COMPUTE_ERROR_ON(pool_pad_x >= pool_size || pool_pad_y >= pool_size);
 
     // Check output dimensions
@@ -82,10 +85,11 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     ARM_COMPUTE_UNUSED(pooled_h);
     ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pooled_w) || (output->info()->dimension(1) != pooled_h));
 
-    const int input_width   = input->info()->dimension(0);
-    const int input_height  = input->info()->dimension(1);
-    const int upper_bound_w = ((pooled_w - 1) * pool_stride_x - pool_pad_x + pool_size) - input_width;
-    const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height;
+    const int num_elements_read_per_iteration = (pool_size == 7) ? 8 : pool_size;
+    const int input_width                     = input->info()->dimension(0);
+    const int input_height                    = input->info()->dimension(1);
+    const int upper_bound_w                   = ((pooled_w - 1) * pool_stride_x - pool_pad_x + num_elements_read_per_iteration) - input_width;
+    const int upper_bound_h                   = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height;
 
     // Set instance variables
     _input              = input;
@@ -138,17 +142,12 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     }
 
     // Configure kernel window
-    const unsigned int num_elems_processed_per_iteration = 1;
-
-    Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
-
+    const unsigned int     num_elems_processed_per_iteration = 1;
+    Window                 win                               = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
     AccessWindowStatic     input_access(input->info(), -pool_pad_x, -pool_pad_y, input_width + _border_size.right, input_height + _border_size.bottom);
     AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-
     update_window_and_padding(win, input_access, output_access);
-
     output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
     ICLKernel::configure(win);
 }