IVGCVSW-820: Add QASYMM8 support to NeonActivationLayerKernel

Change-Id: Ic3881e97b4fcbae0ac287a1e010cfc6f0fd8d7d1
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/112139
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

8 files changed, 189 insertions, 29 deletions

diff --git a/arm_compute/core/NEON/NEAsymm.h b/arm_compute/core/NEON/NEAsymm.h
index d227d3ccbe..f0d7439d40 100644
--- a/arm_compute/core/NEON/NEAsymm.h
+++ b/arm_compute/core/NEON/NEAsymm.h
@@ -28,6 +28,12 @@
 
 namespace arm_compute
 {
+using qasymm8x8_t   = uint8x8_t;   /**< 8 bit quantized asymmetric vector with 8 elements */
+using qasymm8x8x2_t = uint8x8x2_t; /**< 8 bit quantized asymmetric vector with 16 elements */
+using qasymm8x8x3_t = uint8x8x3_t; /**< 8 bit quantized asymmetric vector with 24 elements */
+using qasymm8x8x4_t = uint8x8x4_t; /**< 8 bit quantized asymmetric vector with 32 elements */
+using qasymm8x16_t  = uint8x16_t;  /**< 8 bit quantized asymmetric vector with 16 elements */
+
 /** Round to the nearest division by a power-of-two using exponent
  *
  * @note This function calculates the following expression: (x + 2^n -1 ) / 2^n where n = exponent
@@ -38,6 +44,18 @@ namespace arm_compute
  * @return the nearest division by a power-of-two using exponent
  */
 int32x4_t rounding_divide_by_pow2(int32x4_t x, int exponent);
+
+/** Perform a multiply-accumulate on all 16 components of a QASYMM8 vector
+ *
+ * vd*vs + vo
+ *
+ * @param[in] vd Input vector value in QASYMM8 format
+ * @param[in] vs Vector multiplier in F32 format. The multiplier value must be duplicated across all four lanes.
+ * @param[in] vo Vector addend in F32 format. The addend value must be duplicated across all four lanes.
+ *
+ * @return A 16-component vector in QASYMM8 format, saturated to fit
+ */
+uint8x16_t vmlaq_qasymm8(qasymm8x16_t vd, float32x4_t vs, float32x4_t vo);
 } // namespace arm_compute
 #include "arm_compute/core/NEON/NEAsymm.inl"
-#endif // __ARM_COMPUTE_NEASYMM_H__
\ No newline at end of file
+#endif // __ARM_COMPUTE_NEASYMM_H__
diff --git a/arm_compute/core/NEON/NEAsymm.inl b/arm_compute/core/NEON/NEAsymm.inl
index bbce308b35..ce999a5413 100644
--- a/arm_compute/core/NEON/NEAsymm.inl
+++ b/arm_compute/core/NEON/NEAsymm.inl
@@ -30,4 +30,38 @@ inline int32x4_t rounding_divide_by_pow2(int32x4_t x, int exponent)
     const int32x4_t fixed_up_x = vqaddq_s32(x, fixup);
     return vrshlq_s32(fixed_up_x, shift_vec);
 }
-} // namespace arm_compute
\ No newline at end of file
+
+inline qasymm8x16_t vmlaq_qasymm8(qasymm8x16_t vd, float32x4_t vs, float32x4_t vo)
+{
+    // Convert uint8 vectors to uint16 vectors
+    const uint8x8_t vd_low        = vget_low_u8(vd);
+    const uint8x8_t vd_high       = vget_high_u8(vd);
+    uint16x8_t      vd_low_u16x8  = vmovl_u8(vd_low);
+    uint16x8_t      vd_high_u16x8 = vmovl_u8(vd_high);
+    // Convert uint16 vectors to uint32 vectors
+    uint32x4_t A_u32x4 = vmovl_u16(vget_low_u16(vd_low_u16x8));
+    uint32x4_t B_u32x4 = vmovl_u16(vget_high_u16(vd_low_u16x8));
+    uint32x4_t C_u32x4 = vmovl_u16(vget_low_u16(vd_high_u16x8));
+    uint32x4_t D_u32x4 = vmovl_u16(vget_high_u16(vd_high_u16x8));
+    // Convert uint32 vectors to float32 vectors
+    float32x4_t A_f32x4 = vcvtq_f32_u32(A_u32x4);
+    float32x4_t B_f32x4 = vcvtq_f32_u32(B_u32x4);
+    float32x4_t C_f32x4 = vcvtq_f32_u32(C_u32x4);
+    float32x4_t D_f32x4 = vcvtq_f32_u32(D_u32x4);
+    // vd = vd*vs + vo
+    A_f32x4 = vmlaq_f32(vo, A_f32x4, vs);
+    B_f32x4 = vmlaq_f32(vo, B_f32x4, vs);
+    C_f32x4 = vmlaq_f32(vo, C_f32x4, vs);
+    D_f32x4 = vmlaq_f32(vo, D_f32x4, vs);
+    // Convert float32 vectors to uint32 vectors
+    A_u32x4 = vcvtq_u32_f32(A_f32x4);
+    B_u32x4 = vcvtq_u32_f32(B_f32x4);
+    C_u32x4 = vcvtq_u32_f32(C_f32x4);
+    D_u32x4 = vcvtq_u32_f32(D_f32x4);
+    // Convert uint32 vectors to uint16 vectors (with saturation)
+    vd_low_u16x8  = vcombine_u16(vqmovn_u32(A_u32x4), vqmovn_u32(B_u32x4));
+    vd_high_u16x8 = vcombine_u16(vqmovn_u32(C_u32x4), vqmovn_u32(D_u32x4));
+    // convert uint16 vectors to uint8 vectors (with saturation)
+    return vcombine_u8(vqmovn_u16(vd_low_u16x8), vqmovn_u16(vd_high_u16x8));
+}
+} // namespace arm_compute
diff --git a/arm_compute/core/NEON/kernels/NEActivationLayerKernel.h b/arm_compute/core/NEON/kernels/NEActivationLayerKernel.h
index e8c032aaeb..1edda843de 100644
--- a/arm_compute/core/NEON/kernels/NEActivationLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEActivationLayerKernel.h
@@ -26,6 +26,7 @@
 
 #include "arm_compute/core/FixedPoint.h"
 #include "arm_compute/core/NEON/INEKernel.h"
+#include "arm_compute/core/QAsymm8.h"
 
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 #include <arm_fp16.h>
@@ -105,6 +106,12 @@ private:
      * @param[in] window Region on which to execute the kernel
      */
     template <ActivationLayerInfo::ActivationFunction F, typename T>
+    typename std::enable_if<std::is_same<T, qasymm8_t>::value, void>::type activation(const Window &window);
+    /** Function to apply an activation function on a tensor.
+     *
+     * @param[in] window Region on which to execute the kernel
+     */
+    template <ActivationLayerInfo::ActivationFunction F, typename T>
     typename std::enable_if<std::is_same<T, qint16_t>::value, void>::type activation(const Window &window);
 
 private:
diff --git a/arm_compute/core/QAsymm8.h b/arm_compute/core/QAsymm8.h
new file mode 100644
index 0000000000..2fa4029807
--- /dev/null
+++ b/arm_compute/core/QAsymm8.h
@@ -0,0 +1,33 @@
+/*
+ * 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_QASYMM8_H__
+#define __ARM_COMPUTE_QASYMM8_H__
+
+#include "arm_compute/core/Rounding.h"
+#include <cstdint>
+
+namespace arm_compute
+{
+using qasymm8_t = uint8_t; /**< 8 bit quantized asymmetric scalar value */
+}
+#include "arm_compute/core/QAsymm8.inl"
+#endif /* __ARM_COMPUTE_QASYMM8_H__ */
diff --git a/arm_compute/core/QAsymm8.inl b/arm_compute/core/QAsymm8.inl
new file mode 100644
index 0000000000..611d68eb23
--- /dev/null
+++ b/arm_compute/core/QAsymm8.inl
@@ -0,0 +1,41 @@
+/*
+ * 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.
+ */
+#include <cmath>
+#include <limits>
+
+namespace arm_compute
+{
+inline qasymm8_t sqcvt_qasymm8_f32(float value, float scale, int offset, RoundingPolicy rounding_policy = RoundingPolicy::TO_NEAREST_UP)
+{
+    int quantized = arm_compute::round(value / scale, rounding_policy) + offset;
+    quantized     = std::max(0, std::min(quantized, 255));
+    return quantized;
+}
+
+inline float scvt_f32_qasymm8(qasymm8_t value, float scale, int offset)
+{
+    float dequantized = (static_cast<int>(value) - offset) * scale;
+    return dequantized;
+}
+}
diff --git a/arm_compute/core/Rounding.h b/arm_compute/core/Rounding.h
new file mode 100644
index 0000000000..f95058c567
--- /dev/null
+++ b/arm_compute/core/Rounding.h
@@ -0,0 +1,46 @@
+/*
+ * 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_ROUNDING_H__
+#define __ARM_COMPUTE_ROUNDING_H__
+
+namespace arm_compute
+{
+/** Rounding method */
+enum class RoundingPolicy
+{
+    TO_ZERO,         /**< Truncates the least significand values that are lost in operations. */
+    TO_NEAREST_UP,   /**< Rounds to nearest value; half rounds away from zero */
+    TO_NEAREST_EVEN, /**< Rounds to nearest value; half rounds to nearest even */
+};
+
+/** Return a rounded value of x. Rounding is done according to the rounding_policy.
+ *
+ * @param[in] x               Float value to be rounded.
+ * @param[in] rounding_policy Policy determining how rounding is done.
+ *
+ * @return Rounded value of the argument x.
+ */
+int round(float x, RoundingPolicy rounding_policy);
+}
+#endif /*__ARM_COMPUTE_ROUNDING_H__ */
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index beaec143ef..538449b40a 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -25,10 +25,13 @@
 #define __ARM_COMPUTE_TYPES_H__
 
 #include "arm_compute/core/Coordinates.h"
+#include "arm_compute/core/QAsymm8.h"
+#include "arm_compute/core/Rounding.h"
 #include "arm_compute/core/Strides.h"
 #include "arm_compute/core/TensorShape.h"
 #include "support/Half.h"
 
+#include <cmath>
 #include <cstddef>
 #include <cstdint>
 #include <string>
@@ -102,17 +105,6 @@ constexpr float SCALE_PYRAMID_HALF = 0.5f;
 /* Constant value used to indicate a ORB scaled pyramid */
 constexpr float SCALE_PYRAMID_ORB = 8.408964152537146130583778358414e-01;
 
-/** Rounding method */
-enum class RoundingPolicy
-{
-    TO_ZERO,         /**< Truncates the least significand values that are lost in operations. */
-    TO_NEAREST_UP,   /**< Rounds to nearest value; half rounds away from zero */
-    TO_NEAREST_EVEN, /**< Rounds to nearest value; half rounds to nearest even */
-};
-
-//forward declare round function
-int round(float, RoundingPolicy);
-
 /** Quantization settings (used for QASYMM8 data type) */
 struct QuantizationInfo
 {
@@ -140,20 +132,17 @@ struct QuantizationInfo
     int   offset; /**< offset */
 
     /** Quantizes a value using the scale/offset in this QuantizationInfo */
-    uint8_t quantize(float value, RoundingPolicy rounding_policy) const
+    qasymm8_t quantize(float value, RoundingPolicy rounding_policy) const
     {
         ARM_COMPUTE_ERROR_ON_MSG(scale == 0, "QuantizationInfo::quantize: scale == 0");
-        int quantized = arm_compute::round(value / scale, rounding_policy) + offset;
-        quantized     = std::max(0, std::min(quantized, 255));
-        return quantized;
+        return sqcvt_qasymm8_f32(value, scale, offset, rounding_policy);
     }
 
     /** Dequantizes a value using the scale/offset in this QuantizationInfo */
-    float dequantize(uint8_t value) const
+    float dequantize(qasymm8_t value) const
     {
         ARM_COMPUTE_ERROR_ON_MSG(scale == 0, "QuantizationInfo::dequantize: scale == 0");
-        float dequantized = (static_cast<int>(value) - offset) * scale;
-        return dequantized;
+        return scvt_f32_qasymm8(value, scale, offset);
     }
 
     /** Indicates whether this QuantizationInfo has valid settings or not */
diff --git a/arm_compute/core/Utils.h b/arm_compute/core/Utils.h
index 9397d507f8..f78add13f9 100644
--- a/arm_compute/core/Utils.h
+++ b/arm_compute/core/Utils.h
@@ -25,6 +25,7 @@
 #define __ARM_COMPUTE_UTILS_H__
 
 #include "arm_compute/core/Error.h"
+#include "arm_compute/core/Rounding.h"
 #include "arm_compute/core/Types.h"
 
 #include <algorithm>
@@ -62,15 +63,6 @@ constexpr auto DIV_CEIL(S val, T m) -> decltype((val + m - 1) / m)
     return (val + m - 1) / m;
 }
 
-/** Return a rounded value of x. Rounding is done according to the rounding_policy.
- *
- * @param[in] x               Float value to be rounded.
- * @param[in] rounding_policy Policy determining how rounding is done.
- *
- * @return Rounded value of the argument x.
- */
-int round(float x, RoundingPolicy rounding_policy);
-
 /** Returns the arm_compute library build information
  *
  * Contains the version number and the build options used to build the library