From: Arvind Sankar > Sent: 25 October 2020 23:54 ... > > That's odd, the BLEND loop is about 20 instructions. > > I wouldn't expect unrolling to help - unless you manage > > to use 16 registers for the active W[] values. > > > > I am not sure about what's going on inside the hardware, but even with > a straightforward assembly version that just reads out of memory the way > the calculation is specified, unrolling the BLEND loop 8x speeds up the > performance by 7-8%. > > The compiler is actually pretty bad here, just translating everything > into assembler with no attempt to optimize anything gets a 10-12% > speedup over the C version. I'm not seeing anything particularly stupid. The loop body (excluding loop control) is 23 instructions. Doubles to 46 if I unroll once. Unrolling 4 times does save a couple of instructions per iteration. The only horrid part of the code is the long dependency chain at the end when the values get xor'ed together. gcc is very bad at that, it converts (a + b) + (c + d) to (((a + b) + c) + d) which takes an extra clock. Unrolling 4 times gives almost all the gain. But it really shouldn't be needed at all. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
