From cc5171b85654b9f19a5f52bbe8abea0572ee0163 Mon Sep 17 00:00:00 2001
From: Manuel Bottini <manuel.bottini@arm.com>
Date: Wed, 9 Jan 2019 17:04:39 +0000
Subject: COMPMID-1677: Change ROIPooling layer interface to accept ROIs as
 tensors

Change-Id: If16b572a4d906187b77f32133a72a44316fa74e4
Reviewed-on: https://review.mlplatform.org/490
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
---
 tests/benchmark/fixtures/ROIPoolingLayerFixture.h | 78 ++++++++++++++++++-----
 1 file changed, 61 insertions(+), 17 deletions(-)

(limited to 'tests/benchmark/fixtures')

diff --git a/tests/benchmark/fixtures/ROIPoolingLayerFixture.h b/tests/benchmark/fixtures/ROIPoolingLayerFixture.h
index fa4a5b7044..2c828272de 100644
--- a/tests/benchmark/fixtures/ROIPoolingLayerFixture.h
+++ b/tests/benchmark/fixtures/ROIPoolingLayerFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,37 +39,38 @@ namespace test
 namespace benchmark
 {
 /** Fixture that can be used for NEON and CL */
-template <typename TensorType, typename Function, typename Accessor, typename Array_T, typename ArrayAccessor>
+template <typename TensorType, typename Function, typename AccessorType, typename T>
 class ROIPoolingLayerFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(TensorShape shape, const ROIPoolingLayerInfo pool_info, unsigned int num_rois, DataType data_type, int batches)
+    void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, int batches)
     {
         // Set batched in source and destination shapes
 
         TensorShape shape_dst;
-        shape.set(shape.num_dimensions(), batches);
+        rois_tensor = create_tensor<TensorType>(rois_shape, DataType::U16);
+
+        input_shape.set(input_shape.num_dimensions(), batches);
         shape_dst.set(0, pool_info.pooled_width());
         shape_dst.set(1, pool_info.pooled_height());
-        shape_dst.set(2, shape.z());
-        shape_dst.set(3, num_rois);
+        shape_dst.set(2, input_shape.z());
+        shape_dst.set(3, rois_shape[1]);
 
         // Create tensors
-        src = create_tensor<TensorType>(shape, data_type, 1);
+        src = create_tensor<TensorType>(input_shape, data_type, 1);
         dst = create_tensor<TensorType>(shape_dst, data_type, 1);
 
-        // Create random ROIs
-        std::vector<ROI> rois = generate_random_rois(shape, pool_info, num_rois, 0U);
-        rois_array            = arm_compute::support::cpp14::make_unique<Array_T>(num_rois);
-        fill_array(ArrayAccessor(*rois_array), rois);
-
         // Create and configure function
-        roi_pool.configure(&src, rois_array.get(), &dst, pool_info);
+        roi_pool.configure(&src, &rois_tensor, &dst, pool_info);
 
         // Allocate tensors
+        rois_tensor.allocator()->allocate();
         src.allocator()->allocate();
         dst.allocator()->allocate();
+
+        // Create random ROIs
+        generate_rois(AccessorType(rois_tensor), input_shape, pool_info, rois_shape);
     }
 
     void run()
@@ -89,11 +90,54 @@ public:
         dst.allocator()->free();
     }
 
+protected:
+    template <typename U>
+    void generate_rois(U &&rois, const TensorShape &shape, const ROIPoolingLayerInfo &pool_info, TensorShape rois_shape)
+    {
+        const size_t values_per_roi = rois_shape.x();
+        const size_t num_rois       = rois_shape.y();
+
+        std::mt19937 gen(library->seed());
+        uint16_t    *rois_ptr = static_cast<uint16_t *>(rois.data());
+
+        const float pool_width  = pool_info.pooled_width();
+        const float pool_height = pool_info.pooled_height();
+        const float roi_scale   = pool_info.spatial_scale();
+
+        // Calculate distribution bounds
+        const auto scaled_width  = static_cast<uint16_t>((shape.x() / roi_scale) / pool_width);
+        const auto scaled_height = static_cast<uint16_t>((shape.y() / roi_scale) / pool_height);
+        const auto min_width     = static_cast<uint16_t>(pool_width / roi_scale);
+        const auto min_height    = static_cast<uint16_t>(pool_height / roi_scale);
+
+        // Create distributions
+        std::uniform_int_distribution<int>      dist_batch(0, shape[3] - 1);
+        std::uniform_int_distribution<uint16_t> dist_x1(0, scaled_width);
+        std::uniform_int_distribution<uint16_t> dist_y1(0, scaled_height);
+        std::uniform_int_distribution<uint16_t> dist_w(min_width, std::max(float(min_width), (pool_width - 2) * scaled_width));
+        std::uniform_int_distribution<uint16_t> dist_h(min_height, std::max(float(min_height), (pool_height - 2) * scaled_height));
+
+        for(unsigned int pw = 0; pw < num_rois; ++pw)
+        {
+            const auto batch_idx = dist_batch(gen);
+            const auto x1        = dist_x1(gen);
+            const auto y1        = dist_y1(gen);
+            const auto x2        = x1 + dist_w(gen);
+            const auto y2        = y1 + dist_h(gen);
+
+            rois_ptr[values_per_roi * pw]     = batch_idx;
+            rois_ptr[values_per_roi * pw + 1] = x1;
+            rois_ptr[values_per_roi * pw + 2] = y1;
+            rois_ptr[values_per_roi * pw + 3] = x2;
+            rois_ptr[values_per_roi * pw + 4] = y2;
+        }
+    }
+
 private:
-    TensorType               src{};
-    TensorType               dst{};
-    std::unique_ptr<Array_T> rois_array{};
-    Function                 roi_pool{};
+    TensorType src{};
+    TensorType dst{};
+    TensorType rois_tensor{};
+    Function   roi_pool{};
 };
 } // namespace benchmark
 } // namespace test
-- 
cgit v1.2.1