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>
diff --git a/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h b/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
index 095a833..8f66b8a 100644
--- a/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h
@@ -30,7 +30,11 @@
{
class ITensor;
-/** Interface for the dequantization layer kernel. */
+/** Interface for the dequantization layer kernel.
+ *
+ * @note The implementation supports only 3D input tensors
+ *
+ */
class NEDequantizationLayerKernel : public INEKernel
{
public:
@@ -48,12 +52,12 @@
~NEDequantizationLayerKernel() = default;
/** Set input, output, min and max.
*
- * @param[in] input Source tensor. Data types supported: U8.
- * @param[out] output Destination tensor. Data types supported: F32.
- * @param[in] min Minimum value of the input tensor.
- * @param[in] max Maximum value of the input tensor.
+ * @param[in] input Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data type supported: U8.
+ * @param[out] output Destination tensor with the same dimensions of input. Data type supported: F32.
+ * @param[in] min_max Pointer to the tensor with shape [2, batches] which stores the minimum and maximum value for each 3D input tensor.
+ * The dimensions over the second must match the batched dimensions of the input tensor. Data type supported: F32
*/
- void configure(const ITensor *input, ITensor *output, const float *min, const float *max);
+ void configure(const ITensor *input, ITensor *output, const ITensor *min_max);
// Inherited methods overridden:
void run(const Window &window, const ThreadInfo &info) override;
@@ -61,8 +65,7 @@
private:
const ITensor *_input;
ITensor *_output;
- const float *_min;
- const float *_max;
+ const ITensor *_min_max;
};
}
#endif /*__ARM_COMPUTE_NEDEQUANTIZATIONLAYERKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h b/arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h
new file mode 100644
index 0000000..5e01acf
--- /dev/null
+++ b/arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h
@@ -0,0 +1,77 @@
+/*
+ * Copyright (c) 2017 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_NEMINMAXLAYERKERNEL_H__
+#define __ARM_COMPUTE_NEMINMAXLAYERKERNEL_H__
+
+#include "arm_compute/core/NEON/INEKernel.h"
+
+#include <cstdint>
+#include <mutex>
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Interface for the kernel to perform min max search on a 3D tensor. */
+class NEMinMaxLayerKernel : public INEKernel
+{
+public:
+ /** Default constructor */
+ NEMinMaxLayerKernel();
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ NEMinMaxLayerKernel(const NEMinMaxLayerKernel &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ NEMinMaxLayerKernel &operator=(const NEMinMaxLayerKernel &) = delete;
+ /** Allow instances of this class to be moved */
+ NEMinMaxLayerKernel(NEMinMaxLayerKernel &&) = default;
+ /** Allow instances of this class to be moved */
+ NEMinMaxLayerKernel &operator=(NEMinMaxLayerKernel &&) = default;
+ /** Default destructor */
+ ~NEMinMaxLayerKernel() = default;
+
+ /** Initialise the kernel's input and outputs.
+ *
+ * @note output[0] = minimum
+ * @note output[1] = maximum
+ *
+ * @param[in] input Input tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data type supported: F32.
+ * @param[out] output Output tensor with shape [2, batches, ...] which stores the minimum and maximum value for each 3D input tensor.
+ * The dimensions over the second must match the batched dimensions of the input tensor. Data types supported: F32
+ */
+ void configure(const ITensor *input, ITensor *output);
+ /** Resets global minimum and maximum. */
+ void reset();
+
+ // Inherited methods overridden:
+ void run(const Window &window, const ThreadInfo &info) override;
+
+private:
+ void update_min_max(float *out_ptr, float min, float max);
+ const ITensor *_input;
+ ITensor *_output;
+ std::mutex _mtx;
+};
+}
+#endif /* __ARM_COMPUTE_NEMINMAXLAYERKERNEL_H__ */
\ No newline at end of file
diff --git a/arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h b/arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h
index 92cd142..617a2da 100644
--- a/arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h
@@ -30,7 +30,11 @@
{
class ITensor;
-/** Interface for the quantization layer kernel. */
+/** Interface for the quantization layer kernel.
+ *
+ * @note The implementation supports only 3D input tensors
+ *
+ */
class NEQuantizationLayerKernel : public INEKernel
{
public:
@@ -48,12 +52,12 @@
~NEQuantizationLayerKernel() = default;
/** Set the input, output, min and max.
*
- * @param[in] input Source tensor. Data types supported: F32.
- * @param[out] output Destination tensor. Data types supported: U8.
- * @param[in] min Pointer to the minimum value of the input tensor.
- * @param[in] max Pointer to the maximum value of the input tensor.
+ * @param[in] input Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: F32.
+ * @param[out] output Destination tensor with the same dimensions of input. Data types supported: U8.
+ * @param[in] min_max Pointer to the tensor with shape [2, batches] which stores the minimum and maximum value for each 3D input tensor.
+ * The dimensions over the second must match the batched dimensions of the input tensor. Data type supported: F32
*/
- void configure(const ITensor *input, ITensor *output, const float *min, const float *max);
+ void configure(const ITensor *input, ITensor *output, const ITensor *min_max);
// Inherited methods overridden:
void run(const Window &window, const ThreadInfo &info) override;
@@ -61,8 +65,7 @@
private:
const ITensor *_input;
ITensor *_output;
- const float *_min;
- const float *_max;
+ const ITensor *_min_max;
};
}
#endif /*__ARM_COMPUTE_NEQUANTIZATIONLAYERKERNEL_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h b/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
index 7cd8360..8985861 100644
--- a/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDequantizationLayer.h
@@ -27,7 +27,6 @@
#include "arm_compute/runtime/IFunction.h"
#include "arm_compute/core/NEON/kernels/NEDequantizationLayerKernel.h"
-#include "arm_compute/runtime/Tensor.h"
#include "arm_compute/core/Types.h"
@@ -37,6 +36,8 @@
/** Basic function to simulate a dequantization layer. This function calls the following NEON kernels:
*
+ * @note The implementation supports only 3D input tensors
+ *
* -# @ref NEDequantizationLayerKernel
*
*/
@@ -47,12 +48,12 @@
NEDequantizationLayer();
/** Configure the kernel.
*
- * @param[in] input Source tensor. Data types supported: U8.
- * @param[out] output Destination tensor. Data types supported: F32.
- * @param[in] min Minimum value of the input tensor.
- * @param[in] max Maximum value of the input tensor.
+ * @param[in] input Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: U8.
+ * @param[out] output Destination tensor with the same dimensions of input. Data type supported: F32.
+ * @param[in] min_max Pointer to the tensor with shape [2, batches] which stores the minimum and maximum value for each 3D input tensor.
+ * The dimensions over the second must match the batched dimensions of the input tensor. Data type supported: F32
*/
- void configure(const ITensor *input, ITensor *output, const float *min, const float *max);
+ void configure(const ITensor *input, ITensor *output, const ITensor *min_max);
// Inherited methods overridden:
void run() override;
diff --git a/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h b/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
index ab189fe..d91b4ad 100644
--- a/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEQuantizationLayer.h
@@ -26,7 +26,7 @@
#include "arm_compute/runtime/IFunction.h"
-#include "arm_compute/core/NEON/kernels/NEMinMaxLocationKernel.h"
+#include "arm_compute/core/NEON/kernels/NEMinMaxLayerKernel.h"
#include "arm_compute/core/NEON/kernels/NEQuantizationLayerKernel.h"
#include "arm_compute/runtime/Tensor.h"
@@ -38,7 +38,9 @@
/** Basic function to simulate a quantization layer. This function calls the following NEON kernels:
*
- * -# @ref NEMinMaxKernel
+ * @note The implementation supports only 3D input tensors
+ *
+ * -# @ref NEMinMaxLayerKernel
* -# @ref NEQuantizationLayerKernel
*
*/
@@ -49,8 +51,8 @@
NEQuantizationLayer();
/** Set the input and output tensors.
*
- * @param[in] input Source tensor. Data types supported: F32
- * @param[out] output Destination tensor. Data types supported: U8
+ * @param[in] input Source tensor with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: F32
+ * @param[out] output Destination tensor with the same dimensions of input. Data types supported: U8
*/
void configure(const ITensor *input, ITensor *output);
@@ -59,9 +61,8 @@
private:
NEQuantizationLayerKernel _quantize_kernel;
- NEMinMaxKernel _min_max_kernel;
- float _min;
- float _max;
+ NEMinMaxLayerKernel _min_max_kernel;
+ Tensor _min_max;
};
}
#endif /* __ARM_COMPUTE_NEQUANTIZATIONLAYER_H__ */