aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/NEReductionOperationKernel.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/NEReductionOperationKernel.cpp')
-rw-r--r--src/core/NEON/kernels/NEReductionOperationKernel.cpp2
1 files changed, 2 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/NEReductionOperationKernel.cpp b/src/core/NEON/kernels/NEReductionOperationKernel.cpp
index 1bfef27d49..67ccc5d736 100644
--- a/src/core/NEON/kernels/NEReductionOperationKernel.cpp
+++ b/src/core/NEON/kernels/NEReductionOperationKernel.cpp
@@ -64,6 +64,7 @@ uint32x4x4_t calculate_index(uint32_t idx, T a, T b, uint32x4x4_t c, ReductionOp
return res;
}
+template <>
uint32x4x4_t calculate_index(uint32_t idx, uint8x16_t a, uint8x16_t b, uint32x4x4_t c, ReductionOperation op, int axis)
{
uint32x4x4_t mask{ { 0 } };
@@ -227,6 +228,7 @@ uint32_t calculate_vector_index(uint32x4x4_t vec_res_idx, uint8x16_t vec_res_val
return (res - 0xFFFFFFFF);
}
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+template <>
uint32x4x4_t calculate_index(uint32_t idx, float16x8_t a, float16x8_t b, uint32x4x4_t c, ReductionOperation op, int axis)
{
uint32x4x2_t mask{ 0 };