From c9b7ae7bd53fbd66682bfbbf86064561beec007a Mon Sep 17 00:00:00 2001
From: Tim Besard <tim.besard@gmail.com>
Date: Wed, 7 Oct 2026 07:40:06 +0200
Subject: [PATCH] Avoid trailing_zeros in the strided reduction's offset decode

trailing_zeros becomes SPIR-V's ctz, which spirv2clc can't translate, so
reductions over several non-contiguous dims failed to compile on OpenCL.jl's
OpenCL C path (e.g. pocl with program_backend!(:opencl)). For a power of two,
count_ones(sz - 1) is the same shift.
---
 src/reduce/mapreduce_nd.jl | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/src/reduce/mapreduce_nd.jl b/src/reduce/mapreduce_nd.jl
index bd49707b..f5f9076a 100644
--- a/src/reduce/mapreduce_nd.jl
+++ b/src/reduce/mapreduce_nd.jl
@@ -622,7 +622,8 @@ end
     @inbounds for i in 1:length(reduce_sizes) - 1
         sz = reduce_sizes[i]
         if ispow2(sz)
-            shift = trailing_zeros(sz)
+            # (`trailing_zeros` would be `ctz`, which spirv2clc can't translate to OpenCL C)
+            shift = count_ones(sz - 1)
             r   = tmp & (sz - 1)
             tmp = tmp >> shift
         else