darktable and AMD performance in benchmarks

Beware that auto-vectorization may use instructions that require 32 or 64 byte alignment, and will segfault if the operand is not correctly aligned.
If a type or one of its members has __attribute__((aligned(32))) or 64 to encourage vectorization, then it cannot be safely allocated on the heap with malloc, as malloc only ensures 16 byte alignment. There is no compile time warning for this error, only an intermittent crash at runtime.

1 Like