COMPMID-1564: Add QASYMM8 on CLPixelwiseMultiplication

Change-Id: I5f719f5b2915c18cd0ca6271db401152112863a6
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/148982
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>

1 files changed, 68 insertions, 0 deletions

diff --git a/src/core/CL/cl_kernels/pixelwise_mul_int.cl b/src/core/CL/cl_kernels/pixelwise_mul_int.cl
index c99a08a583..5b3acb7ae6 100644
--- a/src/core/CL/cl_kernels/pixelwise_mul_int.cl
+++ b/src/core/CL/cl_kernels/pixelwise_mul_int.cl
@@ -32,6 +32,7 @@
 
 #define MUL_OP(x, y, scale, type, size) CONVERT_OP_INT((x) * (y) >> scale, type, size)
 
+#if defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT)
 /** Performs a pixelwise multiplication with integer scale of integer inputs.
  *
  * @attention The inputs and output data types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT:
@@ -85,3 +86,70 @@ __kernel void pixelwise_mul_int(
     // Perform multiplication and store result
     vstore16(MUL_OP(in1_data, in2_data, scale, DATA_TYPE_OUT, 16), 0, (__global DATA_TYPE_OUT *)out.ptr);
 }
+#endif /* defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT) */
+
+#if defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT)
+/** Performs a pixelwise multiplication with float scale of quantized inputs.
+ *
+ * @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, e.g. -DOFFSET_IN1=10
+ * @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, e.g. -DOFFSET_IN2=10
+ * @note The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, e.g. -DOFFSET_OUT=10
+ * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, e.g. -DSCALE_IN1=10
+ * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, e.g. -DSCALE_IN2=10
+ * @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, e.g. -DSCALE_OUT=10
+ * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
+ *
+ * @param[in]  in1_ptr                           Pointer to the source image. Supported data types: U8, S16, F16, F32
+ * @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_stride_z                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in1_step_z                        in1_stride_z * 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, F16, F32
+ * @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_stride_z                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  in2_step_z                        in2_stride_z * 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, 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)
+ * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  out_stride_z                      Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  out_step_z                        out_stride_z * 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                             Float scaling factor. Supported data types: F32
+ */
+__kernel void pixelwise_mul_quantized(
+    TENSOR3D_DECLARATION(in1),
+    TENSOR3D_DECLARATION(in2),
+    TENSOR3D_DECLARATION(out),
+    const float scale)
+{
+    // Get pixels pointer
+    Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
+    Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+
+    // Load data
+    int16 in_a = CONVERT(vload16(0, (__global uchar *)in1.ptr), int16);
+    int16 in_b = CONVERT(vload16(0, (__global uchar *)in2.ptr), int16);
+
+    // Dequantize
+    in_a -= (int16)(int)OFFSET_IN1;
+    in_b -= (int16)(int)OFFSET_IN2;
+    const float16 in1f32 = convert_float16(in_a) * (float16)(float)SCALE_IN1;
+    const float16 in2f32 = convert_float16(in_b) * (float16)(float)SCALE_IN2;
+
+    const float16 qresf32 = (in1f32 * in2f32 * scale) / ((float16)(float)SCALE_OUT) + ((float16)((float16)OFFSET_OUT));
+    const uchar16 res     = convert_uchar16_sat(convert_int16_rte(qresf32));
+
+    // Store result
+    vstore16(res, 0, (__global uchar *)out.ptr);
+}
+#endif /* defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) */
\ No newline at end of file