[Mesa-dev] [PATCH 4/9] i965/fs: Get rid of all remaining algebraic optimizations for floats

[Matt Turner]

On Thu, Mar 17, 2016 at 10:21 AM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> ---
>  src/mesa/drivers/dri/i965/brw_fs.cpp | 74 +++++++++---------------------------
>  1 file changed, 18 insertions(+), 56 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
> index 365231e..9f303d5 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
> @@ -2097,28 +2097,27 @@ fs_visitor::opt_algebraic()
>     bool progress = false;
>
>     foreach_block_and_inst(block, fs_inst, inst, cfg) {
> -      switch (inst->opcode) {
> -      case BRW_OPCODE_MOV:
> -         if (inst->src[0].file != IMM)
> -            break;
> -
> -         if (inst->saturate) {
> -            if (inst->dst.type != inst->src[0].type)
> -               assert(!"unimplemented: saturate mixed types");
> +      /* We don't want to do any float algebraic optimizations in the backend
> +       * because they aren't bit-for-bit safe and the backend doesn't know
> +       * when a value is declared precise or invariant.

Does this fix anything? I have a hard time believing single-precision
floating-point addition is more or less precise on the GPU than it is
on the CPU. Perhaps modulo denorm flushing rules.