diff options
| author | Nathan West <nathan.west@okstate.edu> | 2014-10-28 05:36:57 -0500 |
|---|---|---|
| committer | Nathan West <nathan.west@okstate.edu> | 2014-10-28 05:36:57 -0500 |
| commit | 5d1c29c4491a2fcefc69857a7f2f557f6b8eff57 (patch) | |
| tree | 8405d5d00bf6a30d517a4ea771bf304bd04c994f /volk | |
| parent | eba094a3010ca326da747390be51da8326fff0c1 (diff) | |
volk: fix memory overrun/corruption in neon binary_slicer_8i
Diffstat (limited to 'volk')
| -rw-r--r-- | volk/kernels/volk/volk_32f_binary_slicer_8i.h | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/volk/kernels/volk/volk_32f_binary_slicer_8i.h b/volk/kernels/volk/volk_32f_binary_slicer_8i.h index 17b48d7510..ae4420b6e1 100644 --- a/volk/kernels/volk/volk_32f_binary_slicer_8i.h +++ b/volk/kernels/volk/volk_32f_binary_slicer_8i.h @@ -221,7 +221,7 @@ volk_32f_binary_slicer_8i_neon(int8_t* cVector, const float* aVector, int8_t* cPtr = cVector; const float* aPtr = aVector; unsigned int number = 0; - unsigned int n8points = num_points / 8; + unsigned int n16points = num_points / 16; float32x4x2_t input_val0, input_val1; float32x4_t zero_val; @@ -237,7 +237,7 @@ volk_32f_binary_slicer_8i_neon(int8_t* cVector, const float* aVector, // TODO: this is a good candidate for asm because the vcombines // can be eliminated simply by picking dst registers that are // adjacent. - for(number = 0; number < n8points; number++) { + for(number = 0; number < n16points; number++) { input_val0 = vld2q_f32(aPtr); input_val1 = vld2q_f32(aPtr+8); @@ -269,12 +269,12 @@ volk_32f_binary_slicer_8i_neon(int8_t* cVector, const float* aVector, res_u8.val[1] = vand_u8(one, res_u8.val[1]); vst2_u8((unsigned char*)cPtr, res_u8); - cPtr += 8; - aPtr += 8; + cPtr += 16; + aPtr += 16; } - for(number = n8points * 8; number < num_points; number++) { + for(number = n16points * 16; number < num_points; number++) { if(*aPtr++ >= 0) { *cPtr++ = 1; } |