On 08/28/2013 01:15 PM, Christian König wrote:
Well, for this discussion let's just assume that we fixed the delay in
the upper layers of the stack and the driver sees the shader code as
soon as the application (if I understood it correctly Vadim has just
volunteered for the job).
No, I'm not really volunteering to implement that. :)
I'm not even sure if it's possible in reasonable time. In fact it was
more like a theoretical discussion about what would be required for the
early compilation in the driver to make sense.
Perhaps I failed to explain it, but actually my point is that while the
compilation is deferred in upper layers and nobody is going to change
this (if it's possible at all), it doesn't make sense to try compiling
early in the driver. I think we might prefer to defer the compilation in
the driver as well - it doesn't make overall situation any worse, but
can make it better by not compiling unused variants at least.
Vadim
Also let's assume that shaders are small and having allot of shader
variants around after they are compiled isn't bad.
In this case the probably best solution is to compile early and try to
make the shaders as state invariant as possible, e.g. don't do
optimizations like getting ride of extra exports for case where we don't
need the alpha test or if it's just a dependency on a boolean then have
both variants covered by the bytecode and use a bit constant to choose
between the two etc...
As a second step the driver should create a optimized version of the
shader in a background thread when we know all the state that is/was
active when the shader is used.
Of course you need a bit of heuristic for this, cause sometimes it is
better to switch between shader variants and other times it is better to
have one variant covering all the different states and just use bit
constants to choose between them.
Just some thoughts on this topic,
Christian.
PS: My mail server is once more driving me nuts, please ignore the extra
copy if you get this mail twice.
Am 28.08.2013 02:07, schrieb Vadim Girlin:
On 08/28/2013 02:59 AM, Marek Olšák wrote:
First, you won't really see any significant continual difference in
frame rate no matter how many shader variants you have unless you are
very CPU-bound. The problem is shader compilation on the first use,
that's where you get a big hiccup. Try Skyrim for example: You have to
first look around and see every object that's around you and get
unpleasant stuttering before you can actually go on and play the game.
Yes, this also Wine's fault that it compiles shaders on the first use
too, but we don't have to be as bad as Wine, do we? Valve also
reported shader recompilations on the first use being a serious issue
with open source drivers.
I perfectly understand that deferred compilation is exactly the
problem that makes the games freeze due to shader compilation on first
use when something new appears on the screen, but I don't think we can
solve this problem in the *driver* by trying to compile early, because
AFAICS currently the shaders are passed to the driver too late anyway,
and this happens not only with wine. E.g. when I run Heaven in a
window with "MESA_GLSL=dump R600_DEBUG=ps,vs", so that I can see
Heaven's window and console output at the same time, what I see is
that most of GL dumps happen while Heaven shows splash screen with
loading progress, but most of the driver's dumps appear on the first
frame and few more times during benchmark. It looks like compilation
is deferred somewhere in the stack before the driver, or am I missing
something?
Vadim
Marek
On Tue, Aug 27, 2013 at 11:52 PM, Vadim Girlin
<vadimgir...@gmail.com> wrote:
On 08/28/2013 12:43 AM, Marek Olšák wrote:
Shader variants are BAD, BAD, BAD. Have you ever played an AAA game
with a Mesa driver that likes to compile shader variants on first use?
It's HORRIBLE.
I don't think that shader variants are bad, but it's definitely bad
when we
are compiling variants that are never used. Currently glxgears
compiles 18
ps/vs shaders. In my branch with initial GS support [1] I switched
handling
of the shaders to deferred compilation, that is, shaders are
compiled only
before the actual draw. I found later that it's not really required
for GS,
but IIRC this change results in only 5 shaders being compiled for
glxgears
instead of 18. It seems most of the useless variants are results of
state
changes between creation of the shader state (initial compilation) and
actual draw call.
I had some concerns about increased overhead with those changes, and
it's
actually noticeable with drawoverhead demo, but I didn't see any
regressions
with a few real apps that I tested, e.g. glxgears even showed slightly
better performance with these changes. Probably I also implemented
it in a
not very optimal way (I was mostly concentrated on GS support) and the
overhead can be reduced.
One more thing is duplicate shaders, I've analyzed shader dumps from
Unigine
Heaven 3.0 some time ago and found that from about 320 compiled
shaders,
only about 180 (50%) were unique, others were duplicates (detected by
comparing the bytecode dumps for them in an automated way), maybe
they had
different shader keys (which still resulted in the same bytecode),
but I
suspect duplicate pipe shaders were also involved. Unfortunately I
didn't
have a time to investigate it more thoroughly since then.
So my point is that we don't really need to eliminate shader
variants, first
we need to eliminate compilation of unused variants and duplicate
shaders.
Also we might want to consider offloading of the compilation to
separate
thread(s) and caching of shader binaries between runs.
Vadim
[1] http://cgit.freedesktop.org/~vadimg/mesa/log/?h=r600-geom-shaders
What the patch does is probably the right solution. At least
alpha-test state changes don't cause shader recompilation and
re-binding, which also negatively affects performance. Ideally we
shouldn't depend on the framebuffer state at all, but we need to
emulate the TGSI property FS_COLOR0_WRITES_ALL_CBUFS. I think we
should always be fine with key.nr_cbufs forced to 8 for any shader
without that property. I expect app developers to do the right thing
and not write outputs they don't need.
Marek
On Tue, Aug 27, 2013 at 9:00 PM, Roland Scheidegger
<srol...@vmware.com>
wrote:
Not that I'm qualified to review r600 code, but couldn't you create
different shader variants depending on whether you need alpha
test? At
least I would assume shader exports aren't free.
Roland
Am 27.08.2013 19:56, schrieb Vadim Girlin:
We need to export at least one color if the shader writes it,
even when nr_cbufs==0.
Signed-off-by: Vadim Girlin <vadimgir...@gmail.com>
---
Tested on evergreen with multiple combinations of backends - no
regressions,
fixes some tests:
default - fixes fb-alphatest-nocolor and
fb_alphatest-nocolor-ff
default+sb - fixes fb-alphatest-nocolor and
fb_alphatest-nocolor-ff
llvm - fixes about 25 tests related to depth/stencil
llvm+sb - fixes about 300 tests (llvm's depth/stencil
issues and
regressions cased by reordering of exports in sb)
With this patch, there are no regressions with default+sb vs
default.
There is one regression with llvm+sb vs llvm -
fs-texturegrad-miplevels,
AFAICS it's a problem with llvm backend uncovered by sb -
SET_GRADIENTS_V/H
instructions are not placed in the same TEX clause with
corresponding
SAMPLE_G.
src/gallium/drivers/r600/r600_shader.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/src/gallium/drivers/r600/r600_shader.c
b/src/gallium/drivers/r600/r600_shader.c
index 300b5c4..f7eab76 100644
--- a/src/gallium/drivers/r600/r600_shader.c
+++ b/src/gallium/drivers/r600/r600_shader.c
@@ -918,6 +918,7 @@ static int r600_shader_from_tgsi(struct
r600_screen
*rscreen,
unsigned opcode;
int i, j, k, r = 0;
int next_pos_base = 60, next_param_base = 0;
+ int max_color_exports = MAX2(key.nr_cbufs, 1);
/* Declarations used by llvm code */
bool use_llvm = false;
bool indirect_gprs;
@@ -1130,7 +1131,7 @@ static int r600_shader_from_tgsi(struct
r600_screen *rscreen,
radeon_llvm_ctx.face_gpr = ctx.face_gpr;
radeon_llvm_ctx.r600_inputs = ctx.shader->input;
radeon_llvm_ctx.r600_outputs = ctx.shader->output;
- radeon_llvm_ctx.color_buffer_count =
MAX2(key.nr_cbufs ,
1);
+ radeon_llvm_ctx.color_buffer_count =
max_color_exports;
radeon_llvm_ctx.chip_class = ctx.bc->chip_class;
radeon_llvm_ctx.fs_color_all =
shader->fs_write_all &&
(rscreen->chip_class >= EVERGREEN);
radeon_llvm_ctx.stream_outputs = &so;
@@ -1440,7 +1441,7 @@ static int r600_shader_from_tgsi(struct
r600_screen *rscreen,
case TGSI_PROCESSOR_FRAGMENT:
if (shader->output[i].name ==
TGSI_SEMANTIC_COLOR) {
/* never export more colors than the
number of CBs */
- if (shader->output[i].sid >=
key.nr_cbufs)
{
+ if (shader->output[i].sid >=
max_color_exports) {
/* skip export */
j--;
continue;
@@ -1450,7 +1451,7 @@ static int r600_shader_from_tgsi(struct
r600_screen *rscreen,
output[j].type =
V_SQ_CF_ALLOC_EXPORT_WORD0_SQ_EXPORT_PIXEL;
shader->nr_ps_color_exports++;
if (shader->fs_write_all &&
(rscreen->chip_class >= EVERGREEN)) {
- for (k = 1; k <
key.nr_cbufs; k++)
{
+ for (k = 1; k <
max_color_exports;
k++) {
j++;
memset(&output[j], 0,
sizeof(struct r600_bytecode_output));
output[j].gpr =
shader->output[i].gpr;
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
