[Mesa-dev] [PATCH 14/15] i965/vs: Always store pull constant offsets in GRFs on Gen8.

[Kenneth Graunke]

On 11/30/2013 10:40 AM, Paul Berry wrote:
> On 12 November 2013 17:51, Kenneth Graunke <kenneth at whitecape.org
> <mailto:kenneth at whitecape.org>> wrote:
> 
>     We need to SEND from a GRF, and we can only obtain those prior to
>     register allocation.
> 
>     This allows us to do pull constant loads without the MRF hack.
> 
>     Signed-off-by: Kenneth Graunke <kenneth at whitecape.org
>     <mailto:kenneth at whitecape.org>>
>     ---
>      src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp | 13 ++++++++++++-
>      1 file changed, 12 insertions(+), 1 deletion(-)
> 
>     diff --git a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
>     b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
>     index a036e2d..5f0d0b4 100644
>     --- a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
>     +++ b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
>     @@ -1582,7 +1582,13 @@ vec4_visitor::visit(ir_expression *ir)
>            src_reg surf_index =
>               src_reg(prog_data->base.binding_table.ubo_start +
>     uniform_block->value.u[0]);
>            if (const_offset_ir) {
>     -         offset = src_reg(const_offset / 16);
>     +         if (brw->gen >= 8) {
>     +            /* Put the offset in a GRF; we can't SEND from
>     immediates. */
>     +            offset = src_reg(this, glsl_type::int_type);
>     +            emit(MOV(dst_reg(offset), src_reg(const_offset / 16)));
>     +         } else {
>     +            offset = src_reg(const_offset / 16);
>     +         }
>            } else {
>               offset = src_reg(this, glsl_type::uint_type);
>               emit(SHR(dst_reg(offset), op[1], src_reg(4)));
>     @@ -2983,6 +2989,11 @@
>     vec4_visitor::get_pull_constant_offset(vec4_instruction *inst,
>            }
> 
>            return index;
>     +   } else if (brw->gen >= 8) {
>     +      /* Put the offset in a GRF; we can't SEND from immediates. */
>     +      src_reg offset = src_reg(this, glsl_type::int_type);
>     +      emit_before(inst, MOV(dst_reg(offset), src_reg(reg_offset)));
>     +      return offset;
>         } else {
>            int message_header_scale = brw->gen < 6 ? 16 : 1;
>            return src_reg(reg_offset * message_header_scale);
>     --
>     1.8.3.2
> 
> 
> Looking at the docs, it appears to me that SENDs from immediates have
> never been allowed.  Is it possible that they have been working just by
> luck?  Maybe we should make this change apply regardless of brw->gen.

Not by luck...by MRFs.

(Gen4-7) vec4_generator handles the VS_OPCODE_PULL_CONSTANT_LOAD opcode
by emitting MOVs to a MRF (or fake MRF on Gen7).  It's treated as a
OWORD_DUAL_BLOCK_READ, and is never headerless.

On Gen8, I chose to make both VS_OPCODE_PULL_CONSTANT_LOAD and
VS_OPCODE_PULL_CONSTANT_LOAD_GEN7 use the same method (the ld sampler
message), which doesn't require a header, and thus is easy to send from
a GRF.  We just can't use immediates.

Looking at the situation now...it seems pretty stupid that we use
VS_OPCODE_PULL_CONSTANT_LOAD on Gen7+ at all.  Data port reads are
ridiculously slow on Ivybridge, and still pretty bad on Haswell.  And we
use them for UBOs.  We should probably convert those, which would
require doing this code on Gen7 as well.