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>

4 files changed, 90 insertions, 25 deletions

diff --git a/src/core/CL/cl_kernels/fixed_point.h b/src/core/CL/cl_kernels/fixed_point.h
index dcdf840444..b0dab0affc 100644
--- a/src/core/CL/cl_kernels/fixed_point.h
+++ b/src/core/CL/cl_kernels/fixed_point.h
@@ -162,13 +162,34 @@ SUBQ_SAT_IMPL(qs8x16)
 #define SUB_SAT_OP_EXPAND_STR(a, b, type, size) sub_sat_##type##x##size((a), (b))
 #define SUB_SAT_OP_EXPAND(a, b, type, size) SUB_SAT_OP_EXPAND_STR(a, b, type, size)
 
-/** Saturate multiply of two fixed point numbers
-  *
-  * @param[in] type  the actual data type.
-  * @param[in] itype the intermediate data type.
-  *
-  * @return The result of the fixed point multiplication. The result is saturated in case of overflow
-  */
+/* Multiply of two fixed point numbers
+ *
+ * @param[in] type  the actual data type.
+ * @param[in] itype the intermediate data type.
+ *
+ * @return The result of the fixed point multiplication.
+ */
+#define MULQ_IMPL(type, itype)                                                         \
+    inline type mul_##type(type VopA, type VopB, int fixed_point_position)             \
+    {                                                                                  \
+        itype round_val = (itype)(1 << (fixed_point_position - 1));                    \
+        itype res       = CONVERT((VopA), itype) * CONVERT((VopB), itype) + round_val; \
+        return CONVERT((res >> (itype)fixed_point_position), type);                    \
+    }
+
+MULQ_IMPL(qs8x16, qs16x16)
+MULQ_IMPL(qs16x16, qs32x16)
+
+#define MUL_OP_EXPAND_STR(a, b, type, size, position) mul_##type##x##size((a), (b), (position))
+#define MUL_OP_EXPAND(a, b, type, size, position) MUL_OP_EXPAND_STR(a, b, type, size, position)
+
+/* Saturate multiply of two fixed point numbers
+ *
+ * @param[in] type  the actual data type.
+ * @param[in] itype the intermediate data type.
+ *
+ * @return The result of the fixed point multiplication. The result is saturated in case of overflow
+ */
 #define MULQ_SAT_IMPL(type, itype)                                                            \
     inline type mul_sat_##type(type VopA, type VopB, int fixed_point_position)                \
     {                                                                                         \
@@ -179,6 +200,7 @@ SUBQ_SAT_IMPL(qs8x16)
 
 MULQ_SAT_IMPL(qs8x16, qs16x16)
 MULQ_SAT_IMPL(qs16x8, qs32x8)
+MULQ_SAT_IMPL(qs16x16, qs32x16)
 
 #define MUL_SAT_OP_EXPAND_STR(a, b, type, size, position) mul_sat_##type##x##size((a), (b), (position))
 #define MUL_SAT_OP_EXPAND(a, b, type, size, position) MUL_SAT_OP_EXPAND_STR(a, b, type, size, position)
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);
 }
diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
index da417a9020..b95e8fac73 100644
--- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
+++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
@@ -65,12 +65,19 @@ void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const I
     }
 
     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MSG(output->info()->data_type() == DataType::U8 && (input1->info()->data_type() != DataType::U8 || input2->info()->data_type() != DataType::U8),
                              "Output can only be U8 if both inputs are U8");
     ARM_COMPUTE_ERROR_ON_MSG(scale < 0, "Scale cannot be negative. ");
+    if(is_data_type_fixed_point(input1->info()->data_type()))
+    {
+        // All data types must be all QS8 or all QS16
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2, output);
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input1, input2, output);
+        ARM_COMPUTE_ERROR_ON_MSG(scale != 1, "Unsupported scaling factor for QS8/QS16. Scale must be 1.");
+    }
 
     _input1 = input1;
     _input2 = input2;
@@ -96,13 +103,28 @@ void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const I
     if(is_data_type_float(input1->info()->data_type()) || is_data_type_float(input2->info()->data_type()))
     {
         scale_int    = -1;
-        compute_type = (DataType::F32 == input1->info()->data_type() || DataType::F32 == input2->info()->data_type()) ? "float" : "half";
+        compute_type = (input1->info()->data_type() == DataType::F32 || input2->info()->data_type() == DataType::F32) ? "float" : "half";
         data_type    = "DATA_TYPE_FLOAT";
     }
     else
     {
-        compute_type = (DataType::S16 == input1->info()->data_type() || DataType::S16 == input2->info()->data_type()) ? "int" : "ushort";
-        data_type    = "DATA_TYPE_INT";
+        if(input1->info()->data_type() == DataType::S16 || input2->info()->data_type() == DataType::S16)
+        {
+            compute_type = "int";
+        }
+        else if(input1->info()->data_type() == DataType::QS8)
+        {
+            compute_type = "qs8";
+        }
+        else if(input1->info()->data_type() == DataType::QS16)
+        {
+            compute_type = "qs16";
+        }
+        else
+        {
+            compute_type = "ushort";
+        }
+        data_type = "DATA_TYPE_INT";
     }
 
     // Construct kernel name
@@ -113,6 +135,10 @@ void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const I
     std::set<std::string> build_opts;
     build_opts.emplace((overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type())) ? "-DWRAP" : "-DSATURATE");
     build_opts.emplace((rounding_policy == RoundingPolicy::TO_ZERO) ? "-DROUND=_rtz" : "-DROUND=_rte");
+    if(is_data_type_fixed_point(input1->info()->data_type()))
+    {
+        build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input1->info()->fixed_point_position()));
+    }
     build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type()));
     build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type()));
     build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
diff --git a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
index 7c9514723b..c3f61ac94a 100644
--- a/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
+++ b/src/core/NEON/kernels/NEPixelWiseMultiplicationKernel.cpp
@@ -360,7 +360,7 @@ void NEPixelWiseMultiplicationKernel::configure(const ITensor *input1, const ITe
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::S16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MSG(output->info()->data_type() == DataType::U8 && (input1->info()->data_type() != DataType::U8 || input2->info()->data_type() != DataType::U8),
                              "Output can only be U8 if both inputs are U8");
-    if(output->info()->data_type() == DataType::QS8 || input1->info()->data_type() == DataType::QS8 || output->info()->data_type() == DataType::QS8)
+    if(input1->info()->data_type() == DataType::QS8)
     {
         // All data types must be QS8
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2, output);