Port ClGemmLowpOutputStage operator to new interface

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I96630b45887eaba16ef358b95f3d9ac0b9045157
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5882
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 217 insertions, 54 deletions

diff --git a/Android.bp b/Android.bp
index 31a648679f..3435f02d70 100644
--- a/Android.bp
+++ b/Android.bp
@@ -671,6 +671,7 @@ cc_library_static {
         "src/runtime/gpu/cl/operators/ClFlatten.cpp",
         "src/runtime/gpu/cl/operators/ClFloor.cpp",
         "src/runtime/gpu/cl/operators/ClGemm.cpp",
+        "src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.cpp",
         "src/runtime/gpu/cl/operators/ClLogicalNot.cpp",
         "src/runtime/gpu/cl/operators/ClMul.cpp",
         "src/runtime/gpu/cl/operators/ClPRelu.cpp",
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
index e85f2db8a9..6ec7b71f7d 100644
--- a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
@@ -112,10 +112,8 @@ public:
     void run() override;
 
 private:
-    std::unique_ptr<ICLKernel> _kernel;
-    const ICLTensor           *_input;
-    const ICLTensor           *_bias;
-    ICLTensor                 *_output;
+    struct Impl;
+    std::unique_ptr<Impl> _impl;
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H */
diff --git a/filelist.json b/filelist.json
index 517502f8f8..97f1db0901 100644
--- a/filelist.json
+++ b/filelist.json
@@ -231,6 +231,9 @@
       },
       "GEMMLowp": {
         "files": {
+          "operator": [
+            "src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.cpp"
+          ],
           "kernel": [
             "src/core/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.cpp",
             "src/core/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.cpp",
diff --git a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
index e230e8f2e6..94d4c33fa2 100644
--- a/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpOutputStage.cpp
@@ -23,20 +23,32 @@
  */
 #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
 
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/KernelDescriptors.h"
+#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/CL/CLScheduler.h"
 
-#include "src/core/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel.h"
-#include "src/core/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFloatKernel.h"
-#include "src/core/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.h"
+#include "src/core/CL/ICLKernel.h"
+#include "src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.h"
 
 #include <algorithm>
 
 namespace arm_compute
 {
+struct CLGEMMLowpOutputStage::Impl
+{
+    const ICLTensor                               *src{ nullptr };
+    const ICLTensor                               *bias{ nullptr };
+    ICLTensor                                     *dst{ nullptr };
+    std::unique_ptr<opencl::ClGemmLowpOutputStage> op{ nullptr };
+    ITensorPack                                    run_pack{};
+};
+
 CLGEMMLowpOutputStage::CLGEMMLowpOutputStage()
-    : _kernel(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr)
+    : _impl(std::make_unique<Impl>())
 {
 }
 CLGEMMLowpOutputStage::CLGEMMLowpOutputStage(CLGEMMLowpOutputStage &&) = default;
@@ -52,59 +64,22 @@ void CLGEMMLowpOutputStage::configure(const CLCompileContext &compile_context, c
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
-    _input  = input;
-    _bias   = bias;
-    _output = output;
+    _impl->src  = input;
+    _impl->bias = bias;
+    _impl->dst  = output;
 
-    switch(info.type)
-    {
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
-        {
-            auto k = std::make_unique<opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel>();
-            k->configure(compile_context, input->info(), bias != nullptr ? bias->info() : nullptr, output->info(), &info);
-            _kernel = std::move(k);
-            break;
-        }
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
-        {
-            auto k = std::make_unique<opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleKernel>();
-            k->configure(compile_context, input->info(), bias != nullptr ? bias->info() : nullptr, output->info(), &info);
-            _kernel = std::move(k);
-            break;
-        }
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT:
-        {
-            auto k = std::make_unique<opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFloatKernel>();
-            k->configure(compile_context, input->info(), bias != nullptr ? bias->info() : nullptr, output->info(), &info);
-            _kernel = std::move(k);
-            break;
-        }
-        default:
-            ARM_COMPUTE_ERROR("Unsupported GEMMLowpOutputStage type.");
-    }
+    _impl->op = std::make_unique<opencl::ClGemmLowpOutputStage>();
+    _impl->op->configure(compile_context, input->info(), bias != nullptr ? bias->info() : nullptr, output->info(), info);
+    _impl->run_pack = { { ACL_SRC, _impl->src }, { ACL_BIAS, _impl->bias }, { ACL_DST, _impl->dst } };
 }
 
 Status CLGEMMLowpOutputStage::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo &info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16);
-
-    switch(info.type)
-    {
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
-            return opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel::validate(input, bias, output, &info);
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
-            return opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleKernel::validate(input, bias, output, &info);
-        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT:
-            return opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFloatKernel::validate(input, bias, output, &info);
-        default:
-            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type.");
-    }
+    return opencl::ClGemmLowpOutputStage::validate(input, bias, output, info);
 }
 
 void CLGEMMLowpOutputStage::run()
 {
-    ITensorPack pack{ { ACL_SRC, _input }, { ACL_BIAS, _bias }, { ACL_DST, _output } };
-    CLScheduler::get().enqueue_op(*_kernel, pack, true);
+    _impl->op->run(_impl->run_pack);
 }
 } // namespace arm_compute
diff --git a/src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.cpp b/src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.cpp
new file mode 100644
index 0000000000..3477583c76
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.cpp
@@ -0,0 +1,98 @@
+/*
+ * Copyright (c) 2017-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.h"
+
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
+
+#include "src/core/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleByFloatKernel.h"
+#include "src/core/gpu/cl/kernels/ClGemmLowpQuantizeDownInt32ScaleKernel.h"
+
+namespace arm_compute
+{
+namespace opencl
+{
+void ClGemmLowpOutputStage::configure(const CLCompileContext &compile_context, const ITensorInfo *src, const ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+
+    switch(info.type)
+    {
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
+        {
+            auto k = std::make_unique<opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel>();
+            k->configure(compile_context, src, bias, dst, &info);
+            _kernel = std::move(k);
+            break;
+        }
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
+        {
+            auto k = std::make_unique<opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleKernel>();
+            k->configure(compile_context, src, bias, dst, &info);
+            _kernel = std::move(k);
+            break;
+        }
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT:
+        {
+            auto k = std::make_unique<opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFloatKernel>();
+            k->configure(compile_context, src, bias, dst, &info);
+            _kernel = std::move(k);
+            break;
+        }
+        default:
+            ARM_COMPUTE_ERROR("Unsupported GEMMLowpOutputStage type.");
+    }
+}
+
+Status ClGemmLowpOutputStage::validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo &info)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM16);
+
+    switch(info.type)
+    {
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT:
+            return opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel::validate(src, bias, dst, &info);
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN:
+            return opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleKernel::validate(src, bias, dst, &info);
+        case GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT:
+            return opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFloatKernel::validate(src, bias, dst, &info);
+        default:
+            return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Unsupported GEMMLowpOutputStage type.");
+    }
+}
+
+void ClGemmLowpOutputStage::run(ITensorPack &tensors)
+{
+    const ITensor *src  = tensors.get_const_tensor(ACL_SRC);
+    const ITensor *bias = tensors.get_const_tensor(ACL_BIAS);
+    ITensor       *dst  = tensors.get_tensor(ACL_DST);
+
+    ITensorPack pack{ { ACL_SRC, src }, { ACL_BIAS, bias }, { ACL_DST, dst } };
+    CLScheduler::get().enqueue_op(*_kernel, pack, true);
+}
+} // namespace opencl
+} // namespace arm_compute
diff --git a/src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.h b/src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.h
new file mode 100644
index 0000000000..33b82fcafa
--- /dev/null
+++ b/src/runtime/gpu/cl/operators/ClGemmLowpOutputStage.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2017-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_CL_GEMMLOWP_OUTPUT_STAGE_H
+#define ARM_COMPUTE_CL_GEMMLOWP_OUTPUT_STAGE_H
+
+#include "src/core/gpu/cl/ClCompileContext.h"
+#include "src/runtime/gpu/cl/IClOperator.h"
+
+/** This file contains all available output stages for GEMMLowp on OpenCL.
+ *
+ *  In gemmlowp, the "output stage" is the process that takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyCore),
+ *  and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value.
+ *
+ *  More information about the GEMMLowp output stage can be found at https://github.com/google/gemmlowp/blob/master/doc/output.md
+ */
+
+namespace arm_compute
+{
+namespace opencl
+{
+/** Basic function to execute GEMMLowpQuantizeDown kernels on CL.
+ *
+ *  This function calls the following CL kernels:
+ *
+ * -# @ref opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleKernel
+ * -# @ref opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFloatKernel
+ * -# @ref opencl::kernels::ClGemmLowpQuantizeDownInt32ScaleByFixedPointKernel
+*/
+class ClGemmLowpOutputStage : public IClOperator
+{
+public:
+    /** Constructor */
+    ClGemmLowpOutputStage() = default;
+    /** Initialise the kernel's inputs, output
+     *
+     * Valid data layouts:
+     * - All
+     *
+     * Valid data type configurations:
+     * |src0           |src1          |dst           |
+     * |:--------------|:-------------|:-------------|
+     * |S32            |S32           |QASYMM8       |
+     * |S32            |S32           |QASYMM8_SIGNED|
+     * |S32            |S32           |QSYMM16       |
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  src             Source tensor. Data type supported: S32
+     * @param[in]  bias            Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+     *                             Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p src.
+     * @param[out] dst             Destination tensor. Data type supported: QASYMM8/QASYMM8_SIGNED
+     * @param[in]  info            GEMMLowp output stage metadata.
+     */
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *src, const ITensorInfo *bias, ITensorInfo *dst, const GEMMLowpOutputStageInfo &info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClGemmLowpOutputStage::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, const GEMMLowpOutputStageInfo &info);
+
+    // Inherited methods overridden:
+    void run(ITensorPack &tensors) override;
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_GEMMLOWP_OUTPUT_STAGE_H */