COMPMID-439 - Refactored NEQuantizationLayer and NEQuantizationLayer in order to support 3D input tensors

Change-Id: I03eac2108a30bed56d40dfd52e75577a35d492e0
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85783
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 61 insertions, 34 deletions

diff --git a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
index 3bf2b35a09..70984f0a75 100644
--- a/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDequantizationLayerKernel.cpp
@@ -23,9 +23,9 @@
  */
 #include "arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h"
 
+#include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
 #include "arm_compute/core/Window.h"
@@ -35,16 +35,16 @@
 using namespace arm_compute;
 
 NEDequantizationLayerKernel::NEDequantizationLayerKernel()
-    : _input(nullptr), _output(nullptr), _min(nullptr), _max(nullptr)
+    : _input(nullptr), _output(nullptr), _min_max(nullptr)
 {
 }
 
-void NEDequantizationLayerKernel::configure(const ITensor *input, ITensor *output, const float *min, const float *max)
+void NEDequantizationLayerKernel::configure(const ITensor *input, ITensor *output, const ITensor *min_max)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
-    ARM_COMPUTE_ERROR_ON_NULLPTR(min);
-    ARM_COMPUTE_ERROR_ON_NULLPTR(max);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(min_max);
+    ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() < 3);
 
     // Output tensor auto initialization if not yet initialized
     auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, DataType::F32, 0);
@@ -52,17 +52,20 @@ void NEDequantizationLayerKernel::configure(const ITensor *input, ITensor *outpu
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
 
-    _input  = input;
-    _output = output;
-    _min    = min;
-    _max    = max;
+    _input   = input;
+    _output  = output;
+    _min_max = min_max;
 
     constexpr unsigned int num_elems_processed_per_iteration = 8;
 
     // Configure window
     Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
     AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
-    update_window_and_padding(win, AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration), output_access);
+    AccessWindowStatic     min_max_access(min_max->info(), 0, 0, 2, min_max->info()->dimension(1));
+
+    // Update window and padding
+    update_window_and_padding(win, input_access, output_access, min_max_access);
     output_access.set_valid_region(win, input->info()->valid_region());
 
     INEKernel::configure(win);
@@ -74,31 +77,55 @@ void NEDequantizationLayerKernel::run(const Window &window, const ThreadInfo &in
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
-    Iterator input(_input, window);
-    Iterator output(_output, window);
+    Window window_input_output(window);
+    window_input_output.collapse_if_possible(INEKernel::window(), 3);
+    window_input_output.set(3, Window::Dimension(0, 1, 1));
+
+    Window window_min_max;
+    window_min_max.use_tensor_dimensions(_min_max->info()->tensor_shape());
+    window_min_max.set(Window::DimX, Window::Dimension(0, 1, 1));
+    window_min_max.collapse_if_possible(INEKernel::window(), 1);
 
-    const float32x4_t vmin    = vdupq_n_f32(*_min);
-    const float       range   = *_max - *_min;
-    const float32x4_t scaling = vdupq_n_f32(range / 255.0f);
+    Iterator input(_input, window_input_output);
+    Iterator output(_output, window_input_output);
+    Iterator min_max(_min_max, window_min_max);
 
-    // Uniformly map values to range 8bit integers, i.e. [min, max] -> [0, 255]
-    execute_window_loop(window, [&](const Coordinates & id)
+    execute_window_loop(window_min_max, [&](const Coordinates & id_batch)
     {
-        const uint8x8_t  val_u8       = vld1_u8(reinterpret_cast<uint8_t *>(input.ptr()));
-        const uint16x8_t val_u16      = vmovl_u8(val_u8);
-        const uint32x4_t val_u32_low  = vmovl_u16(vget_low_u16(val_u16));
-        const uint32x4_t val_u32_high = vmovl_u16(vget_high_u16(val_u16));
-        float32x4_t      val_low      = vcvtq_f32_u32(val_u32_low);
-        float32x4_t      val_high     = vcvtq_f32_u32(val_u32_high);
-
-        // Dequantize -> (q / 255.0 * range) + min
-        val_low  = vmulq_f32(val_low, scaling);
-        val_high = vmulq_f32(val_high, scaling);
-        val_low  = vaddq_f32(val_low, vmin);
-        val_high = vaddq_f32(val_high, vmin);
-
-        const float32x4x2_t dequantized = vuzpq_f32(val_low, val_high);
-        vst2q_f32(reinterpret_cast<float *>(output.ptr()), dequantized);
+        // Get the min and max
+        const float min = *(reinterpret_cast<const float *>(min_max.ptr()) + 0);
+        const float max = *(reinterpret_cast<const float *>(min_max.ptr()) + 1);
+
+        const float32x4_t vmin    = vdupq_n_f32(min);
+        const float       range   = max - min;
+        const float32x4_t scaling = vdupq_n_f32(range / 255.0f);
+
+        // Uniformly map values to range 8bit integers, i.e. [min, max] -> [0, 255]
+        execute_window_loop(window_input_output, [&](const Coordinates & id)
+        {
+            // Get the input values
+            const auto input_ptr = reinterpret_cast<const uint8_t *>(input.ptr() + id_batch[1] * _input->info()->strides_in_bytes()[3]);
+
+            const uint8x8_t  val_u8       = vld1_u8(input_ptr);
+            const uint16x8_t val_u16      = vmovl_u8(val_u8);
+            const uint32x4_t val_u32_low  = vmovl_u16(vget_low_u16(val_u16));
+            const uint32x4_t val_u32_high = vmovl_u16(vget_high_u16(val_u16));
+            float32x4_t      val_low      = vcvtq_f32_u32(val_u32_low);
+            float32x4_t      val_high     = vcvtq_f32_u32(val_u32_high);
+
+            // Dequantize -> (q / 255.0 * range) + min
+            val_low  = vmulq_f32(val_low, scaling);
+            val_high = vmulq_f32(val_high, scaling);
+            val_low  = vaddq_f32(val_low, vmin);
+            val_high = vaddq_f32(val_high, vmin);
+
+            const float32x4x2_t dequantized = vuzpq_f32(val_low, val_high);
+
+            // Store the dequantized values
+            auto output_ptr = reinterpret_cast<float *>(output.ptr() + id_batch[1] * _output->info()->strides_in_bytes()[3]);
+            vst2q_f32(output_ptr, dequantized);
+        },
+        input, output);
     },
-    input, output);
-}
+    min_max);
+}
\ No newline at end of file