Implement dynamic fusion softmax operator

 - Return aux tensorInfo by get_aux_tensors() at runtime to init the aux
   tensor with the right size.
 - Keep softmax unfusable for this commit
 - Hence, added Tensor3D to template writer arguments declaration, for sake of
   keeping dynamic fusion softmax componenets' kernels matching their cl
   counterparts.

Resolves: COMPMID-5523
Change-Id: I667f39545db925f667036ef448302c79a0330373
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/483924
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8986
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 99 insertions, 4 deletions

diff --git a/arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h b/arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h
index 326880f721..b5af589cd2 100644
--- a/arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h
+++ b/arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022 Arm Limited.
+ * Copyright (c) 2022-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,8 +24,8 @@
 #ifndef ARM_COMPUTE_DYNAMIC_FUSION_RUNTIME_GPU_CL_CLWORKLOADRUNTIME
 #define ARM_COMPUTE_DYNAMIC_FUSION_RUNTIME_GPU_CL_CLWORKLOADRUNTIME
 
+#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/dynamic_fusion/sketch/MemoryDescriptor.h"
-
 #include <map>
 #include <memory>
 
@@ -63,7 +63,7 @@ public:
     Status run(const std::vector<CLTensor *> &tensors);
     /** Get auxiliary tensors of the workload and their memory requirement
      */
-    std::vector<std::pair<CLTensor *, AuxMemoryInfo>> get_auxiliary_tensors();
+    std::vector<std::tuple<CLTensor *, TensorInfo, AuxMemoryInfo>> get_auxiliary_tensors();
 
 private:
     /** Enqueue prepare workload
diff --git a/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h b/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h
index 422edb35f1..155df293bf 100644
--- a/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h
+++ b/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022-2023 Arm Limited.
+ * Copyright (c) 2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.h b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.h
new file mode 100644
index 0000000000..e86ef91e6a
--- /dev/null
+++ b/arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.h
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 2023 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_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPUSOFTMAX
+#define ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPUSOFTMAX
+
+#include "arm_compute/core/ITensorInfo.h"
+#include "arm_compute/dynamic_fusion/sketch/attributes/SoftmaxAttributes.h"
+
+namespace arm_compute
+{
+namespace experimental
+{
+namespace dynamic_fusion
+{
+/** Forward declaration */
+class GpuWorkloadContext;
+class GpuWorkloadSketch;
+
+/** Operator interface. */
+class GpuSoftmax final
+{
+public:
+    /** Attributes are a set of backend-agnostic parameters that define what an operator does */
+    using Attributes = SoftmaxAttributes;
+
+    /** Create an operator and fuse it into the workload sketch.
+     *    @note If @ref validate_op() fails, the creation also fails and may throw an error.
+     *    @note If @ref validate_op() fails, @p sketch remains unchanged and valid.
+     *
+     * Valid data type configurations:
+     * |src            |dst            |
+     * |:--------------|:--------------|
+     * |F16            |F16            |
+     * |F32            |F32            |
+     *
+     * Valid data layouts:
+     * - All
+     *
+     * @param[in,out] sketch     Workload sketch into which the operator will be fused
+     * @param[in]     src        Source tensor info.
+     * @param[in]     dst        Destination tensor info.
+     * @param[in]     attributes Operator attributes
+     */
+    static void create_op(GpuWorkloadSketch &sketch,
+                          ITensorInfo       *src,
+                          ITensorInfo       *dst,
+                          const Attributes &attributes);
+    /** Check if the operator configuration is supported, irrespective of fusion
+     *
+     * @param[in] context    Workload context within which the operator is running
+     * @param[in] src        Source tensor info.
+     * @param[in] dst        Destination tensor info.
+     * @param[in] attributes Operator attributes
+     *
+     * @return Status
+     */
+    static Status is_supported_op(const GpuWorkloadContext &context,
+                                  const ITensorInfo        *src,
+                                  const ITensorInfo        *dst,
+                                  const Attributes         &attributes);
+    /** Validate the operator and check if the its configuration is supported and if it can be fused into the workload sketch.
+     *  Similar to @ref GpuSoftmax::create_op()
+     *
+     * @return a status
+     */
+    static Status validate_op(const GpuWorkloadSketch &sketch,
+                              const ITensorInfo       *src,
+                              const ITensorInfo       *dst,
+                              const Attributes        &attributes);
+};
+} // namespace dynamic_fusion
+} // namespace experimental
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_OPERATORS_GPUSOFTMAX */