COMPMID-3862: Add support QASYMM8 LEAKY RELU activation

- LEAKY RELU activation is supported for QASYMM8 data type
- vquantize on NEON side has been modified to match with
  other backends (OpenCL and reference)

Change-Id: I194631225c8d4f3cc96027d64812ec2be2b4328a
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4593
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 26 insertions, 59 deletions

diff --git a/src/core/NEON/NEAsymm.h b/src/core/NEON/NEAsymm.h
index 70d48d5835..9b92a865d0 100644
--- a/src/core/NEON/NEAsymm.h
+++ b/src/core/NEON/NEAsymm.h
@@ -25,6 +25,7 @@
 #define ARM_COMPUTE_NEASYMM_H
 
 #include "src/core/NEON/NEMath.h"
+#include "src/core/NEON/wrapper/intrinsics/intrinsics.h"
 #include <arm_neon.h>
 
 namespace arm_compute
@@ -647,35 +648,39 @@ inline int8x8_t vquantize_signed(const float32x4x2_t &qv, const UniformQuantizat
     return vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
 }
 
-/** Quantize a neon vector holding 16 floating point values.
- *
- * @param[in] qv Input values to be quantized.
- * @param[in] qi Quantization information to be used in the computation.
- *
- * @return A neon vector holding the quantized values
- */
-inline uint8x16_t vquantize(const float32x4x4_t &qv, const UniformQuantizationInfo &qi)
+inline int32x4x4_t vquantize_internal(const float32x4x4_t &qv, float scale, int32_t offset)
 {
-    const float       scale     = qi.scale;
-    const int         offset    = qi.offset;
-    const float32x4_t voffset   = vdupq_n_f32(offset);
+    const int32x4_t   voffset   = vdupq_n_s32(offset);
     const float32x4_t vinvscale = vdupq_n_f32(1.f / scale);
     const int32x4x4_t rf =
     {
         {
 #ifdef __aarch64__
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),
+            vaddq_s32(vcvtaq_s32_f32(vmulq_f32(qv.val[0], vinvscale)), voffset),
+            vaddq_s32(vcvtaq_s32_f32(vmulq_f32(qv.val[1], vinvscale)), voffset),
+            vaddq_s32(vcvtaq_s32_f32(vmulq_f32(qv.val[2], vinvscale)), voffset),
+            vaddq_s32(vcvtaq_s32_f32(vmulq_f32(qv.val[3], vinvscale)), voffset),
 #else  //__aarch64__
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),
+            vaddq_s32(vcvtq_s32_f32(vmulq_f32(qv.val[0], vinvscale)), voffset),
+            vaddq_s32(vcvtq_s32_f32(vmulq_f32(qv.val[1], vinvscale)), voffset),
+            vaddq_s32(vcvtq_s32_f32(vmulq_f32(qv.val[2], vinvscale)), voffset),
+            vaddq_s32(vcvtq_s32_f32(vmulq_f32(qv.val[3], vinvscale)), voffset),
 #endif //__aarch64__
         }
     };
+    return rf;
+}
+
+/** Quantize a neon vector holding 16 floating point values.
+ *
+ * @param[in] qv Input values to be quantized.
+ * @param[in] qi Quantization information to be used in the computation.
+ *
+ * @return A neon vector holding the quantized values
+ */
+inline uint8x16_t vquantize(const float32x4x4_t &qv, const UniformQuantizationInfo &qi)
+{
+    auto            rf = vquantize_internal(qv, qi.scale, qi.offset);
     const uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
     const uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
     return vcombine_u8(pa, pb);
@@ -690,26 +695,7 @@ inline uint8x16_t vquantize(const float32x4x4_t &qv, const UniformQuantizationIn
  */
 inline int8x16_t vquantize_signed(const float32x4x4_t &qv, const UniformQuantizationInfo &qi)
 {
-    const float       scale     = qi.scale;
-    const int         offset    = qi.offset;
-    const float32x4_t voffset   = vdupq_n_f32(offset);
-    const float32x4_t vinvscale = vdupq_n_f32(1.f / scale);
-    const int32x4x4_t rf =
-    {
-        {
-#ifdef __aarch64__
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),
-#else  //__aarch64__
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),
-#endif //__aarch64__
-        }
-    };
+    auto           rf = vquantize_internal(qv, qi.scale, qi.offset);
     const int8x8_t pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
     const int8x8_t pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
     return vcombine_s8(pa, pb);
@@ -724,26 +710,7 @@ inline int8x16_t vquantize_signed(const float32x4x4_t &qv, const UniformQuantiza
  */
 inline uint16x8x2_t vquantize_qasymm16(const float32x4x4_t &qv, const UniformQuantizationInfo &qi)
 {
-    const float       scale     = qi.scale;
-    const int         offset    = qi.offset;
-    const float32x4_t voffset   = vdupq_n_f32(offset);
-    const float32x4_t vinvscale = vdupq_n_f32(1.f / scale);
-    const int32x4x4_t rf =
-    {
-        {
-#ifdef __aarch64__
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),
-            vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),
-#else  //__aarch64__
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),
-            vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),
-#endif //__aarch64__
-        }
-    };
+    auto             rf = vquantize_internal(qv, qi.scale, qi.offset);
     const uint16x8_t pa = vcombine_u16(vqmovun_s32(rf.val[0]), vqmovun_s32(rf.val[1]));
     const uint16x8_t pb = vcombine_u16(vqmovun_s32(rf.val[2]), vqmovun_s32(rf.val[3]));
     return { pa, pb };