[Mesa-dev] [PATCH 4/6] i965/fs: Implement SIMD16 integer multiplies on Gen 7.
Ian Romanick
idr at freedesktop.org
Tue Sep 30 14:26:38 PDT 2014
On 09/28/2014 01:26 PM, Matt Turner wrote:
> ---
> src/mesa/drivers/dri/i965/brw_fs_visitor.cpp | 22 ++++++++++++++++------
> 1 file changed, 16 insertions(+), 6 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
> index e1f5735..e6c34fa 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
> @@ -634,14 +634,24 @@ fs_visitor::visit(ir_expression *ir)
> else
> emit(MUL(this->result, op[0], op[1]));
> } else {
> - if (brw->gen >= 7)
> - no16("SIMD16 explicit accumulator operands unsupported\n");
> -
> struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
>
> - emit(MUL(acc, op[0], op[1]));
> - emit(MACH(reg_null_d, op[0], op[1]));
> - emit(MOV(this->result, fs_reg(acc)));
> + fs_inst *mul = emit(MUL(acc, op[0], op[1]));
> + fs_inst *mach = emit(MACH(reg_null_d, op[0], op[1]));
> + fs_inst *mov = emit(MOV(this->result, fs_reg(acc)));
> +
> + if (brw->gen == 7 && dispatch_width == 16) {
> + mul->force_uncompressed = true;
> + mach->force_uncompressed = true;
> + mov->force_uncompressed = true;
> +
> + mul = emit(MUL(acc, half(op[0], 1), half(op[1], 1)));
> + mul->force_sechalf = true;
> + mach = emit(MACH(reg_null_d, half(op[0], 1), half(op[1], 1)));
> + mach->force_sechalf = true;
> + mov = emit(MOV(half(this->result, 1), fs_reg(acc)));
> + mov->force_sechalf = true;
> + }
Are there a bunch of cases where we double emit things for SIMD16?
Would it make more sense to have a generic function that took a list of
instructions, duplicated them, and did the force_uncompressed /
force_sechalf modification?
> }
> } else {
> emit(MUL(this->result, op[0], op[1]));
>
More information about the mesa-dev
mailing list