Wow, I solved this so annoying problem. It is not enough to have single barrier inside the cycle. While one wavefront might be doing "update weighted sum" in iteration 1 the other at the same time might already "fill local weight_buffer and input_buffer" in iteration 2 thus ruining the work of the 1st wavefront.
While now I see that the solution looks obvious and the error is silly, it was not obvious at all until I found (almost accidently) the solution. The cycle itself visually looked... so nice and compact that I assumed that I need to take care about syncronization inside cycle only, as if cycle will take care of itself.
I am attaching the fixed code.
__kernel __attribute__((reqd_work_group_size(16, 16, 1))) void ConvolutionRegister( const __global float * restrict input, __global float * restrict output, const __global float * restrict weights, const __global int * restrict weights_offsets, const __global float * restrict biases ) { __local float input_buffer[IN_SIZE]; __local float weight_buffer[W_SIZE]; float sum = 0.0F; for(uint input_feature_map_id = 0; input_feature_map_id < INPUT_FEATURE_MAP_COUNT; input_feature_map_id++) { const int weights_offset = weights_offsets[input_feature_map_id]; // fill local weight_buffer and input_buffer // ... // end of fill local weight_buffer and input_buffer barrier(CLK_LOCAL_MEM_FENCE); // update weighted sum // ... // end of update weighted sum barrier(CLK_LOCAL_MEM_FENCE); } // write result to output feature map output[some_index] = sum; }
