diff options
| author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2018-08-02 12:02:48 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:54 +0000 |
| commit | 6b4e604aa2f8598bce37abe76654b87a41ababb7 (patch) | |
| tree | 8e65c4144d77573dcd0022d4326c7c8e1b042675 /src/core/CL/cl_kernels/depth_convert.cl | |
| parent | 088c2b038ea2de67e90f8a7179c129133127f854 (diff) | |
| download | ComputeLibrary-6b4e604aa2f8598bce37abe76654b87a41ababb7.tar.gz | |
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>
Diffstat (limited to 'src/core/CL/cl_kernels/depth_convert.cl')
| -rw-r--r-- | src/core/CL/cl_kernels/depth_convert.cl | 25 |
1 files changed, 20 insertions, 5 deletions
diff --git a/src/core/CL/cl_kernels/depth_convert.cl b/src/core/CL/cl_kernels/depth_convert.cl index 01491ec1b7..611449e614 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 @@ __kernel void convert_depth_down( 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 @@ __kernel void convert_depth_down( * @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 @@ __kernel void convert_depth_up( 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) */ } |