COMPMID-2775 [NE] add support for QASYMM8_SIGNED to SoftmaxLayer

Change-Id: Ic46d4143929c8c9b548355d85c78542faf25d612
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2376
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 50 insertions, 12 deletions

diff --git a/arm_compute/core/NEON/NEColorConvertHelper.inl b/arm_compute/core/NEON/NEColorConvertHelper.inl
index 62c6eb5aea..7145d6f206 100644
--- a/arm_compute/core/NEON/NEColorConvertHelper.inl
+++ b/arm_compute/core/NEON/NEColorConvertHelper.inl
@@ -83,7 +83,7 @@ inline void rgb_to_u8_conversion(const uint8x16x3_t &in, uint8x16_t &out)
                                                       rgb2u8_red_coef, rgb2u8_green_coef, rgb2u8_blue_coef);
 
     //Conversion from 1(Greyscale) 4 floats to 1(Greyscale) 4 uint8s
-    arm_compute::convert_float32x4x4_to_unit8x16(out_float32, out);
+    arm_compute::convert_float32x4x4_to_uint8x16(out_float32, out);
 }
 
 inline void rgb_to_yuv_calculation(const float32x4_t &rvec, const float32x4_t &gvec, const float32x4_t &bvec,
@@ -214,12 +214,12 @@ inline void rgb_to_yuv_conversion(uint8x16x3_t &vec_top, uint8x16x3_t &vec_botto
                                fyvec_bottom.val[i], fuvec_bottom.val[i], fvvec_bottom.val[i]);
     }
 
-    arm_compute::convert_float32x4x4_to_unit8x16(fyvec_top, vec_top.val[0]);
-    arm_compute::convert_float32x4x4_to_unit8x16(fuvec_top, vec_top.val[1]);
-    arm_compute::convert_float32x4x4_to_unit8x16(fvvec_top, vec_top.val[2]);
-    arm_compute::convert_float32x4x4_to_unit8x16(fyvec_bottom, vec_bottom.val[0]);
-    arm_compute::convert_float32x4x4_to_unit8x16(fuvec_bottom, vec_bottom.val[1]);
-    arm_compute::convert_float32x4x4_to_unit8x16(fvvec_bottom, vec_bottom.val[2]);
+    arm_compute::convert_float32x4x4_to_uint8x16(fyvec_top, vec_top.val[0]);
+    arm_compute::convert_float32x4x4_to_uint8x16(fuvec_top, vec_top.val[1]);
+    arm_compute::convert_float32x4x4_to_uint8x16(fvvec_top, vec_top.val[2]);
+    arm_compute::convert_float32x4x4_to_uint8x16(fyvec_bottom, vec_bottom.val[0]);
+    arm_compute::convert_float32x4x4_to_uint8x16(fuvec_bottom, vec_bottom.val[1]);
+    arm_compute::convert_float32x4x4_to_uint8x16(fvvec_bottom, vec_bottom.val[2]);
 }
 
 inline void store_rgb_to_nv12(const uint8x16_t &rvec_top, const uint8x16_t &gvec_top, const uint8x16_t &bvec_top,
@@ -298,9 +298,9 @@ inline void store_rgb_to_yuv4(const uint8x16_t &rvec, const uint8x16_t &gvec, co
     }
 
     uint8x16_t yvec, uvec, vvec;
-    arm_compute::convert_float32x4x4_to_unit8x16(fyvec, yvec);
-    arm_compute::convert_float32x4x4_to_unit8x16(fuvec, uvec);
-    arm_compute::convert_float32x4x4_to_unit8x16(fvvec, vvec);
+    arm_compute::convert_float32x4x4_to_uint8x16(fyvec, yvec);
+    arm_compute::convert_float32x4x4_to_uint8x16(fuvec, uvec);
+    arm_compute::convert_float32x4x4_to_uint8x16(fvvec, vvec);
 
     vst1q_u8(out_y, yvec);
     vst1q_u8(out_u, uvec);
diff --git a/arm_compute/core/NEON/NEMath.h b/arm_compute/core/NEON/NEMath.h
index aa3054306c..54f8252250 100644
--- a/arm_compute/core/NEON/NEMath.h
+++ b/arm_compute/core/NEON/NEMath.h
@@ -165,6 +165,14 @@ int32_t rounding_divide_by_pow2(int32_t x, int exponent);
  */
 float32x4x4_t convert_uint8x16_to_float32x4x4(const uint8x16_t &in);
 
+/** Converts from int8x16 to float32x4x4_t
+ *
+ * @param[in] in Vector of int8 to be converted
+ *
+ * @return Converted vector of float
+ */
+float32x4x4_t convert_int8x16_to_float32x4x4(const int8x16_t &in);
+
 /** Converts from two float32x4x3_t to just one uint8x8x3_t
  *
  * @param[in]  in1 First input vector of float to be converted
@@ -178,7 +186,14 @@ void convert_float32x4x3_to_uint8x8x3(const float32x4x3_t &in1, const float32x4x
  * @param[in]  in  Vector of float to be converted
  * @param[out] out Converted vector of uint8 to store the result
  */
-void convert_float32x4x4_to_unit8x16(const float32x4x4_t &in, uint8x16_t &out);
+void convert_float32x4x4_to_uint8x16(const float32x4x4_t &in, uint8x16_t &out);
+
+/** Converts from float32x4x4_t to just one int8x16_t
+ *
+ * @param[in]  in  Vector of float to be converted
+ * @param[out] out Converted vector of uint8 to store the result
+ */
+void convert_float32x4x4_to_int8x16(const float32x4x4_t &in, int8x16_t &out);
 
 /** Calculate sine.
  *
diff --git a/arm_compute/core/NEON/NEMath.inl b/arm_compute/core/NEON/NEMath.inl
index 179f1b6299..5d8b82c281 100644
--- a/arm_compute/core/NEON/NEMath.inl
+++ b/arm_compute/core/NEON/NEMath.inl
@@ -331,6 +331,20 @@ inline float32x4x4_t convert_uint8x16_to_float32x4x4(const uint8x16_t &in)
     return out;
 }
 
+inline float32x4x4_t convert_int8x16_to_float32x4x4(const int8x16_t &in)
+{
+    float32x4x4_t out;
+
+    const auto tmp1 = vmovl_s8(vget_low_s8(in));
+    out.val[0]      = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp1)));
+    out.val[1]      = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp1)));
+
+    const auto tmp2 = vmovl_s8(vget_high_s8(in));
+    out.val[2]      = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp2)));
+    out.val[3]      = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp2)));
+    return out;
+}
+
 inline void convert_float32x4x3_to_uint8x8x3(const float32x4x3_t &in1, const float32x4x3_t &in2, uint8x8x3_t &out)
 {
     out.val[0] = vqmovn_u16(vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in1.val[0])),
@@ -341,7 +355,7 @@ inline void convert_float32x4x3_to_uint8x8x3(const float32x4x3_t &in1, const flo
                                          vqmovn_u32(vcvtq_u32_f32(in2.val[2]))));
 }
 
-inline void convert_float32x4x4_to_unit8x16(const float32x4x4_t &in, uint8x16_t &out)
+inline void convert_float32x4x4_to_uint8x16(const float32x4x4_t &in, uint8x16_t &out)
 {
     const auto low = vcombine_u16(vqmovn_u32(vcvtq_u32_f32(in.val[0])),
                                   vqmovn_u32(vcvtq_u32_f32(in.val[1])));
@@ -350,6 +364,15 @@ inline void convert_float32x4x4_to_unit8x16(const float32x4x4_t &in, uint8x16_t
     out = vcombine_u8(vqmovn_u16(low), vqmovn_u16(high));
 }
 
+inline void convert_float32x4x4_to_int8x16(const float32x4x4_t &in, int8x16_t &out)
+{
+    const auto low = vcombine_s16(vqmovn_s32(vcvtq_s32_f32(in.val[0])),
+                                  vqmovn_s32(vcvtq_s32_f32(in.val[1])));
+    const auto high = vcombine_s16(vqmovn_s32(vcvtq_s32_f32(in.val[2])),
+                                   vqmovn_s32(vcvtq_s32_f32(in.val[3])));
+    out = vcombine_s8(vqmovn_s16(low), vqmovn_s16(high));
+}
+
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 /** Exponent polynomial coefficients */
 /** Logarithm polynomial coefficients */