From ce093143ec7b554edefc533c90e45c80946cde51 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 19 Jun 2017 16:11:53 +0100 Subject: 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 Tested-by: Kaizen Reviewed-by: Moritz Pflanzer --- src/core/CL/cl_kernels/pooling_layer.cl | 91 +++++++++++++++++++++++++++++++-- 1 file changed, 87 insertions(+), 4 deletions(-) (limited to 'src/core/CL/cl_kernels/pooling_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; +} -- cgit v1.2.1