COMPMID-515: L2 Pooling for FP32/FP16 in CL.

Change-Id: I43641fa672f5905ca62edd1f63fc93e0cf7ea382
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85963
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

1 files changed, 33 insertions, 0 deletions

diff --git a/arm_compute/core/NEON/NEMath.inl b/arm_compute/core/NEON/NEMath.inl
index bdd747c4e9..50d85396d4 100644
--- a/arm_compute/core/NEON/NEMath.inl
+++ b/arm_compute/core/NEON/NEMath.inl
@@ -64,6 +64,15 @@ inline float32x4_t vfloorq_f32(float32x4_t val)
     return vbslq_f32(vcgtq_f32(r, val), vsubq_f32(r, CONST_1), r);
 }
 
+inline float32x2_t vinvsqrt_f32(float32x2_t x)
+{
+    float32x2_t sqrt_reciprocal = vrsqrte_f32(x);
+    sqrt_reciprocal             = vmul_f32(vrsqrts_f32(vmul_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
+    sqrt_reciprocal             = vmul_f32(vrsqrts_f32(vmul_f32(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
+
+    return sqrt_reciprocal;
+}
+
 inline float32x4_t vinvsqrtq_f32(float32x4_t x)
 {
     float32x4_t sqrt_reciprocal = vrsqrteq_f32(x);
@@ -73,6 +82,14 @@ inline float32x4_t vinvsqrtq_f32(float32x4_t x)
     return sqrt_reciprocal;
 }
 
+inline float32x2_t vinv_f32(float32x2_t x)
+{
+    float32x2_t recip = vrecpe_f32(x);
+    recip             = vmul_f32(vrecps_f32(x, recip), recip);
+    recip             = vmul_f32(vrecps_f32(x, recip), recip);
+    return recip;
+}
+
 inline float32x4_t vinvq_f32(float32x4_t x)
 {
     float32x4_t recip = vrecpeq_f32(x);
@@ -182,6 +199,14 @@ const std::array<float16x8_t, 8> log_tab_f16 =
     }
 };
 
+inline float16x4_t vinvsqrt_f16(float16x4_t x)
+{
+    float16x4_t sqrt_reciprocal = vrsqrte_f16(x);
+    sqrt_reciprocal             = vmul_f16(vrsqrts_f16(vmul_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
+    sqrt_reciprocal             = vmul_f16(vrsqrts_f16(vmul_f16(x, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal);
+    return sqrt_reciprocal;
+}
+
 inline float16x8_t vinvsqrtq_f16(float16x8_t x)
 {
     float16x8_t sqrt_reciprocal = vrsqrteq_f16(x);
@@ -190,6 +215,14 @@ inline float16x8_t vinvsqrtq_f16(float16x8_t x)
     return sqrt_reciprocal;
 }
 
+inline float16x4_t vinv_f16(float16x4_t x)
+{
+    float16x4_t recip = vrecpe_f16(x);
+    recip             = vmul_f16(vrecps_f16(x, recip), recip);
+    recip             = vmul_f16(vrecps_f16(x, recip), recip);
+    return recip;
+}
+
 inline float16x8_t vinvq_f16(float16x8_t x)
 {
     float16x8_t recip = vrecpeq_f16(x);