COMPMID-409: Add support for QS8 and QS16 CLPixelWiseMultiplication.

Change-Id: I7f66d49d746ba9fb6e726ccab83d3a97b8ddef80
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78491
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

1 files changed, 28 insertions, 11 deletions

diff --git a/src/core/CL/cl_kernels/pixelwise_mul_int.cl b/src/core/CL/cl_kernels/pixelwise_mul_int.cl
index e6dfd3043d..a407a3264e 100644
--- a/src/core/CL/cl_kernels/pixelwise_mul_int.cl
+++ b/src/core/CL/cl_kernels/pixelwise_mul_int.cl
@@ -23,12 +23,28 @@
  */
 #include "helpers.h"
 
-#ifdef SATURATE
-#define CONVERT_OP_INT_STR(x, type) (convert_##type##_sat(x))
-#else /* SATURATE */
-#define CONVERT_OP_INT_STR(x, type) (convert_##type(x))
-#endif /* SATURATE */
-#define CONVERT_OP_INT(x, type) CONVERT_OP_INT_STR(x, type)
+#if defined(FIXED_POINT_POSITION)
+
+#include "fixed_point.h"
+
+#if defined(SATURATE)
+#define MUL_OP(x, y, scale, type, size) MUL_SAT_OP_EXPAND((x), (y), type, size, FIXED_POINT_POSITION)
+#else // SATURATE
+#define MUL_OP(x, y, scale, type, size) MUL_OP_EXPAND((x), (y), type, size, FIXED_POINT_POSITION)
+#endif // SATURATE
+
+#else // FIXED_POINT_POSITION
+
+#if defined(SATURATE)
+#define CONVERT_OP_INT_STR(x, type, size) (convert_##type##size##_sat(x))
+#else // SATURATE
+#define CONVERT_OP_INT_STR(x, type, size) (convert_##type##size(x))
+#endif // SATURATE
+#define CONVERT_OP_INT(x, type, size) CONVERT_OP_INT_STR(x, type, size)
+
+#define MUL_OP(x, y, scale, type, size) CONVERT_OP_INT((x) * (y) >> scale, type, size)
+
+#endif // FIXED_POINT_POSITION
 
 /** Performs a pixelwise multiplication with integer scale of integer inputs.
  *
@@ -36,26 +52,27 @@
  * e.g. -DDATA_TYPE_IN1=uchar -DDATA_TYPE_IN2=ushort -DDATA_TYPE_OUT=short
  * @attention The data_type of the intermediate result of the multiplication should passed as well using -DDATA_TYPE_RES.
  * e.g. If one of inputs is S16 -DDATA_TYPE_RES=int should be passed else -DDATA_TYPE_RES=short.
+ * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3
  *
- * @param[in]  in1_ptr                           Pointer to the source image. Supported data types: U8, S16
+ * @param[in]  in1_ptr                           Pointer to the source image. Supported data types: U8/QS8/QS16/S16
  * @param[in]  in1_stride_x                      Stride of the source image in X dimension (in bytes)
  * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  in1_stride_y                      Stride of the source image in Y dimension (in bytes)
  * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source image
- * @param[in]  in2_ptr                           Pointer to the source image. Supported data types: U8, S16
+ * @param[in]  in2_ptr                           Pointer to the source image. Supported data types: same as @p in1_ptr
  * @param[in]  in2_stride_x                      Stride of the source image in X dimension (in bytes)
  * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  in2_stride_y                      Stride of the source image in Y dimension (in bytes)
  * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  in2_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, S16
+ * @param[out] out_ptr                           Pointer to the destination image. Supported data types: same as @p in1_ptr
  * @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)
  * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination image
- * @param[in]  scale                             Integer scaling factor. Supported data types: S32
+ * @param[in]  scale                             Integer scaling factor. Supported data types: S32 (ignored for QS8 and QS16 as the assumption is scale = 1).
  */
 __kernel void pixelwise_mul_int(
     IMAGE_DECLARATION(in1),
@@ -75,5 +92,5 @@ __kernel void pixelwise_mul_int(
     in2_data = CONVERT(vload16(0, (__global DATA_TYPE_IN2 *)in2.ptr), VEC_DATA_TYPE(DATA_TYPE_RES, 16));
 
     // Perform multiplication and store result
-    vstore16(CONVERT_OP_INT(((in1_data * in2_data) >> scale), VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+    vstore16(MUL_OP(in1_data, in2_data, scale, DATA_TYPE_OUT, 16), 0, (__global DATA_TYPE_OUT *)out.ptr);
 }