On Tue, Oct 16, 2018 at 4:57 PM Sagar Ghuge <sagar.gh...@intel.com> wrote:
>
Let's describe a little of why we're doing this and how we found it.
If I recall correctly, we set a NOP (XYZW) swizzle on 3-src
instructions, except we set an XXXX swizzle on LRP. Is that correct?
> Signed-off-by: Sagar Ghuge <sagar.gh...@intel.com>
> ---
> src/intel/compiler/brw_eu_emit.c | 36 ++++++++++++++++++++++++--------
> 1 file changed, 27 insertions(+), 9 deletions(-)
>
> diff --git a/src/intel/compiler/brw_eu_emit.c
> b/src/intel/compiler/brw_eu_emit.c
> index 0cbc682ebc..a6b45fcb1a 100644
> --- a/src/intel/compiler/brw_eu_emit.c
> +++ b/src/intel/compiler/brw_eu_emit.c
> @@ -765,31 +765,49 @@ brw_alu3(struct brw_codegen *p, unsigned opcode, struct
> brw_reg dest,
> brw_inst_set_3src_a16_dst_writemask(devinfo, inst, dest.writemask);
>
> assert(src0.file == BRW_GENERAL_REGISTER_FILE);
> - brw_inst_set_3src_a16_src0_swizzle(devinfo, inst, src0.swizzle);
> brw_inst_set_3src_a16_src0_subreg_nr(devinfo, inst,
> get_3src_subreg_nr(src0));
> brw_inst_set_3src_src0_reg_nr(devinfo, inst, src0.nr);
> brw_inst_set_3src_src0_abs(devinfo, inst, src0.abs);
> brw_inst_set_3src_src0_negate(devinfo, inst, src0.negate);
> - brw_inst_set_3src_a16_src0_rep_ctrl(devinfo, inst,
> - src0.vstride ==
> BRW_VERTICAL_STRIDE_0);
> +
> + /* RepCtrl field in Source or Destination Operand table in BDW Bspec
> + * says:
> + * "ChanSel does not apply when Replicate Control is set."
> + */
> + if (src0.vstride == BRW_VERTICAL_STRIDE_0)
> + brw_inst_set_3src_a16_src0_rep_ctrl(devinfo, inst, true);
> + else
> + brw_inst_set_3src_a16_src0_swizzle(devinfo, inst, src0.swizzle);
I see that the brw_vec1_reg() function uses an XXXX swizzle. That
indeed determines what swizzle scalar sources (i.e., those with
rep_ctrl set) will have.
I would rather we always set the swizzle (so as to not complicate the
code here) and instead figure out why the brw_reg's swizzle field is
sometimes XYZW fix that.
I modified brw_disasm.c locally to always print the swizzle on 3-src
operands and using the piglit tests
generated_tests/spec/glsl-1.10/execution/built-in-functions/fs-mix-vec2-vec2-vec2.shader_test
(for LRP)
tests/spec/arb_gpu_shader5/execution/built-in-functions/fs-fma.shader_test
(for MAD)
I cannot confirm that the swizzles are different between MAD and LRP.
What am I missing?
lrp(8) g4<1>F g2.4<0,1,0>.xF g2.2<0,1,0>.xF
g2.0<0,1,0>.xF { align16 1Q };
mad(8) g4<1>F g3.0<0,1,0>.xF g2.4<0,1,0>.xF
g2.0<0,1,0>.xF { align16 1Q };
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
