COMPMID-431 Port CLDepthConvert to use 8-bit and 16-bit fixed point

Change-Id: Iedea9e985427e6242f34a5362615f79c0526d5bd
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79786
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

1 files changed, 42 insertions, 7 deletions

diff --git a/src/core/CL/cl_kernels/depth_convert.cl b/src/core/CL/cl_kernels/depth_convert.cl
index 3a1c7ca2c5..a9b7284c83 100644
--- a/src/core/CL/cl_kernels/depth_convert.cl
+++ b/src/core/CL/cl_kernels/depth_convert.cl
@@ -23,24 +23,47 @@
  */
 #include "helpers.h"
 
+#if defined(FIXED_POINT_POSITION)
+
+#include "fixed_point.h"
+
+#ifdef SATURATE
+#define CONVERT_DOWN(x, in_type, out_type, fixed_point_position) CONVERT_DOWN1_SAT(x, in_type, out_type, fixed_point_position)
+#define CONVERT_DOWN1_SAT(x, in_type, out_type, fixed_point_position) convert_##out_type##_##in_type##_sat(x, fixed_point_position)
+#else /* SATURATE */
+#define CONVERT_DOWN(x, in_type, out_type, fixed_point_position) CONVERT_DOWN1(x, in_type, out_type, fixed_point_position)
+#define CONVERT_DOWN1(x, in_type, out_type, fixed_point_position) convert_##out_type##_##in_type(x, fixed_point_position)
+#endif /* SATURATE */
+
+#define CONVERT_UP(x, in_type, out_type, fixed_point_position) CONVERT_UP1(x, in_type, out_type, fixed_point_position)
+#define CONVERT_UP1(x, in_type, out_type, fixed_point_position) convert_##out_type##_##in_type(x, fixed_point_position)
+
+#else /* FIXED_POINT_POSITION */
+
 #ifdef SATURATE
 #define CONVERT_DOWN(x, type) CONVERT_SAT(x, type)
 #else /* SATURATE */
 #define CONVERT_DOWN(x, type) CONVERT(x, type)
 #endif /* SATURATE */
 
+#define CONVERT_UP(x, type) CONVERT(x, type)
+
+#endif /* FIXED_POINT_POSITION */
+
 /** This function performs a down-scaling depth conversion.
  *
  * @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
+ * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3
+ *
+ * @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 or S32
+ * @param[out] out_ptr                           Pointer to the destination image. Supported data types: QS8, U8, QS16, U16, S16, U32, S32
  * @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)
@@ -60,7 +83,12 @@ __kernel void convert_depth_down(
     // Load data
     VEC_DATA_TYPE(DATA_TYPE_IN, 16)
     in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr);
+
+#if defined(FIXED_POINT_POSITION)
+    vstore16(CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_IN, 16), VEC_DATA_TYPE(DATA_TYPE_OUT, 16), FIXED_POINT_POSITION), 0, (__global DATA_TYPE_OUT *)out.ptr);
+#else  /* FIXED_POINT_POSITION */
     vstore16(CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+#endif /* FIXED_POINT_POSITION */
 }
 
 /** This function performs a up-scaling depth conversion.
@@ -68,13 +96,15 @@ __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
+ * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3
+ *
+ * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8, QS8, U16, S16, QS16, U32 or S32
  * @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 or S32
+ * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8, U16, S16, U32, S32, F16 or 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)
@@ -92,7 +122,12 @@ __kernel void convert_depth_up(
     Image out = CONVERT_TO_IMAGE_STRUCT(out);
 
     // Load data
-    VEC_DATA_TYPE(DATA_TYPE_OUT, 16)
-    in_data = CONVERT(vload16(0, (__global DATA_TYPE_IN *)in.ptr), VEC_DATA_TYPE(DATA_TYPE_OUT, 16));
-    vstore16(in_data << shift, 0, (__global DATA_TYPE_OUT *)out.ptr);
+    VEC_DATA_TYPE(DATA_TYPE_IN, 16)
+    in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr);
+
+#if defined(FIXED_POINT_POSITION)
+    vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_IN, 16), VEC_DATA_TYPE(DATA_TYPE_OUT, 16), FIXED_POINT_POSITION), 0, (__global DATA_TYPE_OUT *)out.ptr);
+#else  /* FIXED_POINT_POSITION */
+    vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)) << shift, 0, (__global DATA_TYPE_OUT *)out.ptr);
+#endif /* FIXED_POINT_POSITION */
 }