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>

4 files changed, 103 insertions, 15 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index a7a95b8b94..391a0bb6d5 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -332,6 +332,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "permute_3201", "permute.cl" },
     { "pixelwise_mul_float", "pixelwise_mul_float.cl" },
     { "pixelwise_mul_int", "pixelwise_mul_int.cl" },
+    { "pixelwise_mul_quantized", "pixelwise_mul_int.cl" },
     { "pooling_layer_2", "pooling_layer.cl" },
     { "pooling_layer_3", "pooling_layer.cl" },
     { "pooling_layer_optimized_3", "pooling_layer.cl" },
diff --git a/src/core/CL/cl_kernels/pixelwise_mul_float.cl b/src/core/CL/cl_kernels/pixelwise_mul_float.cl
index f4f36a02af..9fa540e946 100644
--- a/src/core/CL/cl_kernels/pixelwise_mul_float.cl
+++ b/src/core/CL/cl_kernels/pixelwise_mul_float.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -30,6 +30,7 @@
 #endif /* SATURATE */
 #define CONVERT_OP_FLOAT(x, type, round) CONVERT_OP_FLOAT_STR(x, type, round)
 
+#if defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT)
 /** Performs a pixelwise multiplication with float scale of either integer or float 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:
@@ -93,3 +94,4 @@ __kernel void pixelwise_mul_float(
     // Store result
     vstore16(res, 0, (__global DATA_TYPE_OUT *)out.ptr);
 }
+#endif /* defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(DATA_TYPE_RES) && defined(DATA_TYPE_OUT) */
\ No newline at end of file
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
diff --git a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
index 4ca2ef8aa3..286b94ebdc 100644
--- a/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
+++ b/src/core/CL/kernels/CLPixelWiseMultiplicationKernel.cpp
@@ -51,9 +51,9 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2,
     ARM_COMPUTE_UNUSED(rounding_policy);
 
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input2);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(scale < 0, "Scale cannot be negative.");
 
     const TensorShape &out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape());
@@ -64,7 +64,7 @@ Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2,
     if(output->total_size() > 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(output);
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QASYMM8, DataType::S16, DataType::F16, DataType::F32);
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_type() == DataType::U8 && (input1->data_type() != DataType::U8 || input2->data_type() != DataType::U8),
                                         "Output can only be U8 if both inputs are U8");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), "Wrong shape for output");
@@ -168,27 +168,44 @@ void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const I
         data_type = "DATA_TYPE_INT";
     }
 
+    const bool is_quantized = is_data_type_quantized_asymmetric(input1->info()->data_type());
+
     // Construct kernel name
     std::string kernel_name = "pixelwise_mul";
-    kernel_name += (scale_int >= 0) ? "_int" : "_float";
+    if(!is_data_type_quantized(output->info()->data_type()))
+    {
+        kernel_name += (scale_int >= 0) ? "_int" : "_float";
+    }
 
     // Set kernel build options
-    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");
-    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()));
-    build_opts.emplace("-DDATA_TYPE_RES=" + compute_type);
-    build_opts.emplace("-D" + data_type);
+    CLBuildOptions build_opts;
+    if(is_quantized)
+    {
+        build_opts.add_option("-DOFFSET_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().offset));
+        build_opts.add_option("-DOFFSET_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().offset));
+        build_opts.add_option("-DOFFSET_OUT=" + support::cpp11::to_string(output->info()->quantization_info().offset));
+        build_opts.add_option("-DSCALE_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().scale));
+        build_opts.add_option("-DSCALE_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().scale));
+        build_opts.add_option("-DSCALE_OUT=" + support::cpp11::to_string(output->info()->quantization_info().scale));
+        kernel_name += "_quantized";
+    }
+    else
+    {
+        build_opts.add_option_if_else(overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type()), "-DWRAP", "-DSATURATE");
+        build_opts.add_option_if_else(rounding_policy == RoundingPolicy::TO_ZERO, "-DROUND=_rtz", "-DROUND=_rte");
+        build_opts.add_option("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type()));
+        build_opts.add_option("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type()));
+        build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+        build_opts.add_option("-DDATA_TYPE_RES=" + compute_type);
+    }
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Set scale argument
     unsigned int idx = 3 * num_arguments_per_3D_tensor(); //Skip the inputs and output parameters
 
-    if(scale_int >= 0)
+    if(scale_int >= 0 && !is_quantized)
     {
         _kernel.setArg(idx++, scale_int);
     }