COMPMID-1303: CLDepthConvert : Add support for FP32 -> FP16 and FP16 -> FP32 + validate() function
Change-Id: I6808de0254a7c4bca440322cc14b795b3b32465b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/142427
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/src/core/CL/cl_kernels/depth_convert.cl b/src/core/CL/cl_kernels/depth_convert.cl
index 01491ec..611449e 100644
--- a/src/core/CL/cl_kernels/depth_convert.cl
+++ b/src/core/CL/cl_kernels/depth_convert.cl
@@ -24,8 +24,13 @@
#include "helpers.h"
#ifdef SATURATE
+#if defined(IS_DATA_TYPE_FLOAT)
+#define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
+#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
+#else /* defined(IS_DATA_TYPE_FLOAT) */
#define CONVERT_DOWN(x, type) CONVERT_SAT(x, type)
-#else /* SATURATE */
+#endif /* defined(IS_DATA_TYPE_FLOAT) */
+#else /* SATURATE */
#define CONVERT_DOWN(x, type) CONVERT(x, type)
#endif /* SATURATE */
@@ -36,13 +41,13 @@
* @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
* e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
*
- * @param[in] in_ptr Pointer to the source image. Supported data types: U8, U16, S16, U32, S32, F16, F32
+ * @param[in] in_ptr Pointer to the source image. Supported data types: U8/U16/S16/U32/S32/F16/F32
* @param[in] in_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] in_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image
- * @param[out] out_ptr Pointer to the destination image. Supported data types: U8, U16, S16, U32, S32
+ * @param[out] out_ptr Pointer to the destination image. Supported data types: U8/U16/S16/U32/S32/F16/F32
* @param[in] out_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes)
@@ -63,7 +68,12 @@
VEC_DATA_TYPE(DATA_TYPE_IN, 16)
in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr);
+#if defined(IS_DATA_TYPE_FLOAT)
+ const DATA_TYPE_IN scale = (DATA_TYPE_IN)(1 << shift);
+ vstore16(CONVERT_DOWN(in_data / scale, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+#else /* defined(IS_DATA_TYPE_FLOAT) */
vstore16(CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+#endif /* defined(IS_DATA_TYPE_FLOAT) */
}
/** This function performs a up-scaling depth conversion.
@@ -71,13 +81,13 @@
* @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
* e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
*
- * @param[in] in_ptr Pointer to the source image. Supported data types: U8, U16, S16, U32 or S32
+ * @param[in] in_ptr Pointer to the source image. Supported data types: U8/U16/S16/U32/S32/F16/F32
* @param[in] in_stride_x Stride of the source image in X dimension (in bytes)
* @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] in_stride_y Stride of the source image in Y dimension (in bytes)
* @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] in_offset_first_element_in_bytes The offset of the first element in the source image
- * @param[out] out_ptr Pointer to the destination image. Supported data types: U8, U16, S16, U32, S32, F16 or F32
+ * @param[out] out_ptr Pointer to the destination image. Supported data types: U8/U16/S16/U32/S32/F16/F32
* @param[in] out_stride_x Stride of the destination image in X dimension (in bytes)
* @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] out_stride_y Stride of the destination image in Y dimension (in bytes)
@@ -98,5 +108,10 @@
VEC_DATA_TYPE(DATA_TYPE_IN, 16)
in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr);
+#if defined(IS_DATA_TYPE_FLOAT)
+ const DATA_TYPE_OUT scale = (DATA_TYPE_OUT)(1 << shift);
+ vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)) * scale, 0, (__global DATA_TYPE_OUT *)out.ptr);
+#else /* defined(IS_DATA_TYPE_FLOAT) */
vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)) << shift, 0, (__global DATA_TYPE_OUT *)out.ptr);
+#endif /* defined(IS_DATA_TYPE_FLOAT) */
}
diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
index 2f5b246..ffbd295 100644
--- a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
@@ -25,6 +25,7 @@
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/CL/OpenCL.h"
#include "arm_compute/core/Error.h"
@@ -38,64 +39,83 @@
using namespace arm_compute;
-void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift)
+namespace
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16,
- DataType::U16, DataType::U32, DataType::S32);
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16,
- DataType::U16, DataType::U32, DataType::S32, DataType::F32);
- ARM_COMPUTE_ERROR_ON(input == output);
- ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == output->info()->data_type(), "Input and output data types must be different");
- ARM_COMPUTE_ERROR_ON(shift >= 8);
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift)
+{
+ ARM_COMPUTE_UNUSED(policy);
+ ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+ ARM_COMPUTE_RETURN_ERROR_ON(input == output);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16,
+ DataType::U16, DataType::U32, DataType::S32,
+ DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16,
+ DataType::U16, DataType::U32, DataType::S32,
+ DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == output->data_type(), "Input and output data types must be different");
+ ARM_COMPUTE_RETURN_ERROR_ON(shift >= 8);
// Check if convertion is supported
- ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U8 && (output->info()->data_type() != DataType::U16 && output->info()->data_type() != DataType::S16
- && output->info()->data_type() != DataType::U32 && output->info()->data_type() != DataType::S32),
- "Only data types supported [in] U8 -> [out] U16, S16, U32, S32");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::U8 && (output->data_type() != DataType::U16 && output->data_type() != DataType::S16
+ && output->data_type() != DataType::U32 && output->data_type() != DataType::S32),
+ "Only data types supported [in] U8 -> [out] U16, S16, U32, S32");
- ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U16 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U32
- && output->info()->data_type() != DataType::S32),
- "Only data types supported [in] U16 -> [out] U8, U32, S32");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::U16 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U32
+ && output->data_type() != DataType::S32),
+ "Only data types supported [in] U16 -> [out] U8, U32, S32");
- ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::S16 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U32
- && output->info()->data_type() != DataType::S32),
- "Only data types supported [in] S16 -> [out] U8, U32, S32");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::S16 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U32
+ && output->data_type() != DataType::S32),
+ "Only data types supported [in] S16 -> [out] U8, U32, S32");
- ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U32 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U16
- && output->info()->data_type() != DataType::S16),
- "Only data types supported [in] U32 -> [out] U8, U16, S16");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::U32 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U16
+ && output->data_type() != DataType::S16),
+ "Only data types supported [in] U32 -> [out] U8, U16, S16");
- ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::S32 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U16
- && output->info()->data_type() != DataType::S16),
- "Only data types supported [in] S32 -> [out] U8, U16, S16");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::S32 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U16
+ && output->data_type() != DataType::S16),
+ "Only data types supported [in] S32 -> [out] U8, U16, S16");
+
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::F16 && output->data_type() != DataType::F32,
+ "Only data types supported [in] F16 -> [out] F32");
+
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::F32 && output->data_type() != DataType::F16,
+ "Only data types supported [in] F32 -> [out] F16");
+
+ // Validate in case of configured output
+ if(output->total_size() > 0)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+ }
+
+ return Status{};
+}
+} // namespace
+
+void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
// Auto initialize output shape if not initialized (We can only auto-configure the shape, datatype must be given)
set_shape_if_empty(*output->info(), input->info()->tensor_shape());
- ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), policy, shift));
// Get data sizes
const size_t input_size = data_size_from_type(input->info()->data_type());
const size_t output_size = data_size_from_type(output->info()->data_type());
- // Construct kernel name and build options
- std::string kernel_name = "convert_depth";
- std::set<std::string> build_opts;
- if(input_size > output_size)
- {
- kernel_name += "_down";
- // Down conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
- build_opts.insert(((policy == ConvertPolicy::WRAP) && !is_data_type_float(input->info()->data_type())) ? "-DWRAP" : "-DSATURATE");
- }
- else
- {
- kernel_name += "_up";
- }
- build_opts.emplace("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type()));
- build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+ // Set build options
+ CLBuildOptions build_opts;
+ build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type()));
+ build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+ // Down conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
+ build_opts.add_option_if(input_size > output_size, ((policy == ConvertPolicy::WRAP) && !is_data_type_float(input->info()->data_type())) ? "-DWRAP" : "-DSATURATE");
+ build_opts.add_option_if(is_data_type_float(input->info()->data_type()), "-DIS_DATA_TYPE_FLOAT");
// Create kernel
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
+ const std::string kernel_name = (input_size > output_size) ? "convert_depth_down" : "convert_depth_up";
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
// Set shift arg
unsigned int idx = 2 * num_arguments_per_2D_tensor(); //Skip the input and output parameters
@@ -105,3 +125,10 @@
constexpr unsigned int num_elems_processed_per_iteration = 16;
ICLSimple2DKernel::configure(input, output, num_elems_processed_per_iteration);
}
+
+Status CLDepthConvertLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift)
+{
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, policy, shift));
+
+ return Status{};
+}