This struct allows us to report:
- accurate max point size/line width.
- accurate texel and texture gather offsets
- vertex/geometry limits.
Signed-off-by: Dave Airlie <airlied@redhat.com>
My gcc doesn't figure out that dims >= 1 (seems reasonable), and doesn't
notice that ddmax is used from the same no_rho_opt as its initialization.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Flush a resource's previous write_batch synchronously. Because a
resource's associated batches are not updated until after the flush
thread submits rendering to the kernel, this was causing a bit of
confusion in the following loop. This fixes a bug that appeared with
recent stk.
Perhaps we need to re-work things a bit to clear out dependent patches
in the ctx's thread and use a fence to deal with the period between
when a flush is queued and when it is submitted to the kernel. But
this will do until time permits a larger refactor.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Because of loops, we can't schedule all of a block's predecessors first.
Instead just assume that the result consumed in a block was written far
enough away in all paths into a block. And do an intra-block scheduling
pass to figure out if there are any cases where we need to insert extra
nop's. This works out better than always assuming the worst case (ie.
that a value live into a block was written in the last instruction in
the predecessor block).
Signed-off-by: Rob Clark <robdclark@gmail.com>
Normally false-deps are not something to consider, since they mostly
exist for delay-slot related reasons:
* barriers
* ordering writes after read
* SSBO/image access ordering
The exception is a false-dependency on an array store.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Previously we didn't handle flow control in legalize, and instead just
set (ss)(sy) on the first instruction in every block. Which isn't very
clever.
Instead, consider output state of all predecessor blocks, so we only
set a sync bit if needed for any possible path leading into a block.
Because of loops, we can't require that all successor blocks are
legalized before a given block, so instead run in a loop until results
converge.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Maybe there is a better way for this.. where it comes useful is "array"
loads, which end up as a false-dep for a later array store.
If all the uses of an array load are CP'd into their consumer, it still
leaves the dangling array load, leading to funny things like:
mov.u32u32 r5.y, r0.y
mov.u32u32 r5.y, r0.z
Signed-off-by: Rob Clark <robdclark@gmail.com>
Generally seems to do worse on instruction count and register usage,
according to shader-db. But shader-db also doesn't do a very good job
of weighting loop bodies, so that might not be totally valid.
So add an env variable to enable GCM pass for easier experimentation.
Signed-off-by: Rob Clark <robdclark@gmail.com>
There are more useful nir passes added since initial conversion to nir.
But ir3 was never updated to use them.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Agressively lowering all if/else to selects in some extreme cases
results in much higher register pressure. Using peephole select instead
with a modest threshold speeds up alu2 4x!
16 seems like a good limit, low enough to help alu2 but not too low that
it penalizes everything else. With a bit better scheduling of the
instruction that moves a value into a predicate register, we might be
able to lower this limit a bit more in the future, but since we need 6
cycles from the move to predicate register to predicated branch, that
puts some sort of lower bound on how far we can lower this threshold.
Signed-off-by: Rob Clark <robdclark@gmail.com>
nir's from_ssa pass is much better at avoiding inserting extra moves
than our logic is. And lowering phi webs to regs just treats anything
involved in a phi web as an array of length=1. Which with previous
array related fixes in RA/etc ends up working out quite well. This cuts
down on extra instructions and also helps with register pressure.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Makes it easier to compare values seen in-game (where there are many
shaders) to cmdline standalone compiler.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Seems to be there since a3xx, but we always lowered fsat. But we can
shave some instructions, especially in shaders that use lots of
clamp(foo, 0.0, 1.0) by not lowering fsat.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Use livein state of other blocks to extend liverange of arrays when they
are still needed by successor blocks.
Signed-off-by: Rob Clark <robdclark@gmail.com>
When eliminating movs, the instruction that is now directly using the
src of the mov has the same scheduling order constraints as the original
mov instruction.
Signed-off-by: Rob Clark <robdclark@gmail.com>
The number of bits depends on generation. But printing negative values
with a5xx encoding (largest size) but compiling for a3xx or a4xx, would
result in negative values printed as large positive values.
I guess in practice huge negative branch offsets aren't likely (and if
that is the case, the shader is probably too big to grok by reading the
assembly). So just print using smallest bitfield size.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Clover now checks PIPE_SHADER_CAP_SUPPORTED_IRS for native support instead.
This change indirectly enables NIR support for compute shaders
on radeonsi.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
PIPE_SHADER_CAP_PREFERRED_IR was conflicting with PIPE_SHADER_IR_NIR
for compute shaders, so we let clover pick the one it wants to use.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Include the Mesa version and detail about the platform.
Signed-off-by: Mark Thompson <sw@jkqxz.net>
Reviewed-by: Christian König <christian.koenig@amd.com>
It is present from libva 2.1 (VAAPI 1.1.0 or higher).
Signed-off-by: Mark Thompson <sw@jkqxz.net>
Reviewed-by: Christian König <christian.koenig@amd.com>
When NIR is enabled, TGSI must not be used. When NIR is disabled, TGSI
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
v2: * Check whether the node src and dst registers are NULL before using
them.
* fix a type in the commit message.
Two cases are handled with this patch:
1. If copy propagation tries to eliminated a move from a relative
array access then it could optimize
MOV R1, ARRAY[RELADDR_1]
MOV R2, ARRAY[RELADDR_2]
OP2 R3, R1 R2
into
OP2 R3, ARRAY[RELADDR_1], ARRAY[RELADDR_2]
which is forbidden, because there is only one address register available.
2. When MULADD(x,a,MUL(x,c)) is handled
MUL TMP, R1, ARRAY[RELADDR_1]
MULLADD R3, R1, ARRAY[RELADDR_2], TMP
by folding this into
ADD TMP, ARRAY[RELADDR_2], ARRAY[RELADDR_1]
MUL R3, R1, TMP
which is also forbidden.
Test for these cases and reject the optimization if a forbidden combination
of relative access would be created.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103142
Signed-off-by: Gert Wollny <gw.fossdev@gmail.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
On r600 we use the scratch mem with read/read_ind, in that case
sb should track the rw_gpr as a dst instead of a src.
This stops the whole shader being optimised out.
Signed-off-by: Dave Airlie <airlied@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
