On Wed, Oct 28, 2015 at 3:08 AM, Kenneth Graunke <kenn...@whitecape.org> wrote:
> On Tuesday, October 27, 2015 04:40:19 PM Kristian Høgsberg wrote:
>> On Mon, Oct 12, 2015 at 02:55:32PM -0700, Kenneth Graunke wrote:
>> > Signed-off-by: Kenneth Graunke <kenn...@whitecape.org>
>> > ---
>> > src/mesa/drivers/dri/i965/brw_fs.cpp | 174 ++++++++++
>> > src/mesa/drivers/dri/i965/brw_fs.h | 16 +-
>> > src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 378
>> > ++++++++++++++++++++++
>> > src/mesa/drivers/dri/i965/brw_fs_visitor.cpp | 49 ++-
>> > src/mesa/drivers/dri/i965/brw_vec4_gs_visitor.cpp | 21 ++
>> > 5 files changed, 628 insertions(+), 10 deletions(-)
>> >
>> > diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp
>> > b/src/mesa/drivers/dri/i965/brw_fs.cpp
>> > index dde8c45..778237a 100644
>> > --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
>> > +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
>> > @@ -43,6 +43,7 @@
>> > #include "brw_wm.h"
>> > #include "brw_fs.h"
>> > #include "brw_cs.h"
>> > +#include "brw_vec4_gs_visitor.h"
>> > #include "brw_cfg.h"
>> > #include "brw_dead_control_flow.h"
>> > #include "main/uniforms.h"
>> > @@ -1347,6 +1348,47 @@ fs_visitor::emit_discard_jump()
>> > }
>> >
>> > void
>> > +fs_visitor::emit_gs_thread_end()
>> > +{
>> > + assert(stage == MESA_SHADER_GEOMETRY);
>> > +
>> > + if (gs_compile->control_data_header_size_bits > 0) {
>> > + emit_gs_control_data_bits(this->final_gs_vertex_count);
>> > + }
>> > +
>> > + const fs_builder abld = bld.annotate("thread end");
>> > + fs_inst *inst;
>> > +
>> > + if (gs_compile->prog_data.static_vertex_count != -1) {
>> > + foreach_in_list_reverse(fs_inst, prev, &this->instructions) {
>> > + if (prev->opcode == SHADER_OPCODE_URB_WRITE_SIMD8 ||
>> > + prev->opcode == SHADER_OPCODE_URB_WRITE_SIMD8_MASKED ||
>> > + prev->opcode == SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT ||
>> > + prev->opcode ==
>> > SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT) {
>> > + prev->eot = true;
>> > + return;
>> > + } else if (prev->is_control_flow() || prev->has_side_effects()) {
>> > + break;
>> > + }
>> > + }
>> > + fs_reg hdr = abld.vgrf(BRW_REGISTER_TYPE_UD, 1);
>> > + abld.MOV(hdr, fs_reg(retype(brw_vec8_grf(1, 0),
>> > BRW_REGISTER_TYPE_UD)));
>> > + inst = abld.emit(SHADER_OPCODE_URB_WRITE_SIMD8, reg_undef, hdr);
>> > + inst->mlen = 1;
>> > + } else {
>> > + fs_reg payload = abld.vgrf(BRW_REGISTER_TYPE_UD, 2);
>> > + fs_reg *sources = ralloc_array(mem_ctx, fs_reg, 2);
>> > + sources[0] = fs_reg(retype(brw_vec8_grf(1, 0),
>> > BRW_REGISTER_TYPE_UD));
>> > + sources[1] = this->final_gs_vertex_count;
>> > + abld.LOAD_PAYLOAD(payload, sources, 2, 2);
>> > + inst = abld.emit(SHADER_OPCODE_URB_WRITE_SIMD8, reg_undef, payload);
>> > + inst->mlen = 2;
>> > + }
>> > + inst->eot = true;
>> > + inst->offset = 0;
>> > +}
>> > +
>> > +void
>> > fs_visitor::assign_curb_setup()
>> > {
>> > if (dispatch_width == 8) {
>> > @@ -1550,6 +1592,53 @@ fs_visitor::assign_vs_urb_setup()
>> > }
>> > }
>> >
>> > +void
>> > +fs_visitor::assign_gs_urb_setup()
>> > +{
>> > + assert(stage == MESA_SHADER_GEOMETRY);
>> > +
>> > + const gl_geometry_program *gp = &gs_compile->gp->program;
>> > + brw_vue_prog_data *vue_prog_data = (brw_vue_prog_data *) prog_data;
>> > +
>> > + first_non_payload_grf +=
>> > + 8 * vue_prog_data->urb_read_length * gp->VerticesIn;
>>
>> Where does the 8 * come from here?
>
> Vertex data is read from the URB in 256-bit (HWord) units - two vec4
> slots at a time. The fact that this happens to be the size of a
> register is quite confusing, but not entirely relevant :)
Ah yes, *urb* read length is a number of vec4s. I was thinking that
urb_read_length was a number of regs. I think this part of the spec
confused me:
"This is followed by GRFs containing Vertex 0 Element 1 (if it
exists), and so on, up to the number of Vertex 0 elements specified by
Vertex URB Read Length."
but of course, there it's understood that URB Read Length is pairs or vec4s.
> Data read from the URB is expanded into the SIMD8 data layout, where
> a vec4 takes up 4 registers. So when we read a 256-bit URB block,
> those two vec4s end up taking up 8 registers.
>
> See vec4_gs_visitor::setup_varying_inputs(), which multiplies by 2
> since in SIMD4x2 mode a vec4 takes up a whole register, so 2 vec4s
> take up 2 registers.
>
> Another way of explaining it: a SIMD4x2 GS processes 2 completely
> unrelated primitives in each half. Data comes from two VUEs.
>
> VUE entry for primitive 0: [<A0.x A0.y A0.z A0.w>, <B0.x B0.y B0.z B0.w>,
> ...]
> VUE entry for primitive 1: [<A1.x A1.y A1.z A1.w>, <B1.x B1.y B1.z B1.w>,
> ...]
>
> One URB access reads 256-bits, but splats that across 2 registers:
>
> r10 = A1.x A1.y A1.z A1.w | A0.x A0.y A0.z A0.w
> r11 = B1.x B1.y B1.z B1.w | B0.x B0.y B0.z B0.w
>
> Similarly, SIMD8 shaders process 8 independent primitives per thread.
> The data comes from 8 different VUEs:
>
> VUE entry for primitive 0: [<A0.x A0.y A0.z A0.w>, <B0.x B0.y B0.z B0.w>,
> ...]
> VUE entry for primitive 1: [<A1.x A1.y A1.z A1.w>, <B1.x B1.y B1.z B1.w>,
> ...]
> VUE entry for primitive 2: [<A2.x A2.y A2.z A2.w>, <B2.x B2.y B2.z B2.w>,
> ...]
> VUE entry for primitive 3: [<A3.x A3.y A3.z A3.w>, <B3.x B3.y B3.z B3.w>,
> ...]
> VUE entry for primitive 4: [<A4.x A4.y A4.z A4.w>, <B4.x B4.y B4.z B4.w>,
> ...]
> VUE entry for primitive 5: [<A5.x A5.y A5.z A5.w>, <B5.x B5.y B5.z B5.w>,
> ...]
> VUE entry for primitive 6: [<A6.x A6.y A6.z A6.w>, <B6.x B6.y B6.z B6.w>,
> ...]
> VUE entry for primitive 7: [<A7.x A7.y A7.z A7.w>, <B7.x B7.y B7.z B7.w>,
> ...]
>
> One URB access reads 256-bits, but splats that across 8 registers:
>
> r10 = A7.x A6.x A5.x A4.x A3.x A2.x A1.x A0.x
> r11 = A7.y A6.y A5.y A4.y A3.y A2.y A1.y A0.y
> r12 = A7.z A6.z A5.z A4.z A3.z A2.z A1.z A0.z
> r13 = A7.w A6.w A5.w A4.w A3.w A2.w A1.w A0.w
> r14 = B7.x B6.x B5.x B4.x B3.x B2.x B1.x B0.x
> r15 = B7.y B6.y B5.y B4.y B3.y B2.y B1.y B0.y
> r16 = B7.z B6.z B5.z B4.z B3.z B2.z B1.z B0.z
> r17 = B7.w B6.w B5.w B4.w B3.w B2.w B1.w B0.w
>
> I hope that helps! This was a good question.
This was a good answer.
thanks,
Kristian
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
