From 5c829cab49b0dcffd893065af05f0bf028e7a548 Mon Sep 17 00:00:00 2001 From: Manuel Bottini Date: Tue, 28 Jan 2020 17:25:48 +0000 Subject: COMPMID-3045: CTS failures in ARGMAX/MIN Change-Id: I35276a3d95dc99a7f4dea00e89c8ed206a5f13f1 Signed-off-by: Manuel Bottini Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2669 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Michalis Spyrou --- src/core/CL/cl_kernels/arg_min_max.cl | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) (limited to 'src/core/CL/cl_kernels/arg_min_max.cl') diff --git a/src/core/CL/cl_kernels/arg_min_max.cl b/src/core/CL/cl_kernels/arg_min_max.cl index 06dcc8ddde..104d30d8f3 100644 --- a/src/core/CL/cl_kernels/arg_min_max.cl +++ b/src/core/CL/cl_kernels/arg_min_max.cl @@ -254,10 +254,15 @@ __kernel void arg_min_max_x( barrier(CLK_LOCAL_MEM_FENCE); + // Looking for the next highest power of 2 (maximum value of lsize is 8) + unsigned int middle = lsize - 1; + middle |= middle >> 1; + middle |= middle >> 2; + middle += 1; // Perform parallel reduction - for(unsigned int i = lsize >> 1; i > 0; i >>= 1) + for(unsigned int i = middle; i > 0; i >>= 1) { - if(lid < i) + if( lid < i && lid + i < lsize) { DATA_TYPE tmp0 = *(src_in_row + local_results[lid]); DATA_TYPE tmp1 = *(src_in_row + local_results[lid + i]); -- cgit v1.2.1