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>
For compute shaders, we need to be able to allocate some "high"
registers (r48.x to r55.w). (Possibly these are global to all threads
in a warp?) Add a new register class to handle this.
Signed-off-by: Rob Clark <robdclark@gmail.com>
The instruction encoding allows for more registers, but at least on
a3xx/a4xx they don't actually exist.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Normally the offset in the group would be the same, but not always. For
example, in a sam(w) which only writes the 4th component.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Shuffle things slightly, passing instr-data to ra_name() to reduce the
number of places where we need to add support for array names.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
We'll need to add similar for ir3_instruction, but following the pattern
to use 'id' seems confusing. Let's just go w/ generic 'data' as the
name.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Since i965 is now using make_reg_conflicts_transitive and doesn't need
q-value computations, they are disabled on i965. They are enabled
everywhere else so that they get the old behavior. This reduces the time
spent in eglInitialize() on BDW by around 10-15%.
Reviewed-by: Eric Anholt <eric@anholt.net>
The address and predicate register are special, they don't get assigned
in RA. So do a better job of ignoring them rather than hitting later
asserts.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
It is silly to traverse back to find first instruction that writes part
of a larger "virtual" register many times per instruction (plus per use
as a src to later instructions). Cache this information so we only
figure it out once.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The fanin source could be grouped, for example with shaders like:
VERT
DCL IN[0]
DCL IN[1]
DCL OUT[0], POSITION
DCL OUT[1], GENERIC[9]
DCL SAMP[0]
DCL SVIEW[0], 2D, FLOAT
DCL TEMP[0], LOCAL
0: MOV TEMP[0].xy, IN[1].xyyy
1: MOV TEMP[0].w, IN[1].wwww
2: TXF TEMP[0], TEMP[0], SAMP[0], 2D
3: MOV OUT[1], TEMP[0]
4: MOV OUT[0], IN[0]
5: END
The second arg to the isaml is IN[1].w, so we need to look at the fanin
source to get the correct offset.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
For query_levels, we generate a getinfo with writemask of (z), which RA
will consider as size==3. But we were still generating four fanouts.
Which meant that RA would see it as two different register classes,
depending on the path to definer. Ie. on the getinfo instruction itself
it would see size==3, but when chasing back through the fanouts it would
see size==4.
Easiest way to solve that is to just generate the chain of neighboring
fanouts to have the correct size in the first place.
Note: we may eventually want split_dest() to take start/end or wrmask
instead, since really we only need size==1. But RA is not clever enough
for that, query_levels is not that common, and the other two registers
that get allocated are never used so those register slots can be
immediately re-used. So bunch of work for probably no real gain.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
This shuffles things around to allow the shader to have multiple basic
blocks. We drop the entire CFG structure from nir and just preserve the
blocks. At scheduling we know whether to schedule conditional branches
or unconditional jumps at the end of the block based on the # of block
successors. (Dropping jumps to the following instruction, etc.)
One slight complication is that variables (load_var/store_var, ie.
arrays) are not in SSA form, so we have to figure out where to put the
phi's ourself. For this, we use the predecessor set information from
nir_block. (We could perhaps use NIR's dominance frontier information
to help with this?)
Signed-off-by: Rob Clark <robclark@freedesktop.org>
These belong in the shader, rather than the block. Mostly a lot of
churn and nothing too interesting. But splitting this out from the
rest of ir3_block reshuffling to cut down the noise in the later
patch.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Use standard list_head double-linked list and related iterators,
helpers, etc, rather than weird combo of instruction array and next
pointers depending on stage. Now block has an instrs_list. In
certain stages where we want to remove and re-add to the blocks list
we just use list_replace() to copy the list to a new list_head.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
At least for now.. right now the instruction and instruction list
printing should suffice, and the re-working of ir3_block would require
a lot of changes in that code.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
To simplify RA, assign arrays that are written to first. Since enough
dependency information is in the graph to preserve order of reads and
writes of array, so all SSA names for the array collapse into one, just
assign the entire thing by array-id.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The meta-deref instruction doesn't really do what we need for relative
destination. Instead, since each instruction can reference at most a
single address value, track the dependency on the address register via
instr->address. This lets us express the dependency regardless of
whether it is used for dst and/or src.
The foreach_ssa_src{_n} iterator macros now also iterates the address
register so, at least in SSA form, the address register behaves as an
additional virtual src to the instruction. Which is pretty much what
we want, as far as scheduling/etc.
TODO:
For now, the foreach_src{_n} iterators are unchanged. We could wrap
the address in an ir3_register and make the foreach_src_{_n} iterators
behave the same way. But that seems unnecessary at this point, since
we mainly care about the address dependency when in SSA form.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
I remembered that we are using c99.. which makes some sugary iterator
macros easier. So introduce iterator macros to iterate all src
registers and all SSA src instructions. The _n variants also return
the src #, since there are a handful of places that need this.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
We can't (or don't know how to) turn this off. But it can end up being
stored to a higher reg # than what the shader uses, leading to
corruption.
Also we currently aren't clever enough to turn off frag_coord/frag_face
if the input is dead-code, so just fixup max_reg/max_half_reg. Re-org
this a bit so both vp and fp reg footprint fixup are called by a common
fxn used also by ir3_cmdline. Also add a few more output lines for
ir3_cmdline to make it easier to see what is going on.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Handle TEMP[ADDR[]] src registers by generating a fanin to group array
elements, similarly to how texture fetch instructions work.
NOTE:
For all the scalar instructions generated for a single tgsi vector
operation which uses an array src (or possibly even uses the same array
as multiple srcs), re-use the same fanin node. Since a vector operation
operates on all components at the same time, it should never see more
than one version of the same array.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Group inputs/outputs, in addition to fanin/fanout, as they must also
exist in sequential scalar registers. This lets us simplify RA by
working in terms of neighbor groups.
NOTE: has the slight problem that it can't optimize out mov's for things
like:
MOV OUT[n], IN[m]
To avoid this, instead of trying to figure out what mov's we can
eliminate, we first remove all mov's prior to grouping, and then
re-insert mov's as needed while grouping inputs/outputs/fanins.
Eventually we'd prefer the frontend to not insert extra mov's in the
first place (so we don't have to bother removing them). This is the
plan for an eventual NIR based frontend, so separate out the instr
grouping (which will still be needed for NIR frontend) from the mov
elimination (which won't).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Shaders like:
FRAG
PROPERTY FS_COLOR0_WRITES_ALL_CBUFS 1
DCL IN[0], GENERIC[0], PERSPECTIVE
DCL OUT[0], COLOR
DCL SAMP[0]
DCL TEMP[0], LOCAL
IMM[0] FLT32 { 0.0000, 1.0000, 0.0000, 0.0000}
0: TEX TEMP[0], IN[0].xyyy, SAMP[0], 2D
1: MOV OUT[0], IMM[0].xyxx
2: END
cause unhappyness. They have an IN[], but once this is compiled the
useless TEX instruction goes away. Leaving a varying that is never
fetched, which makes the hw unhappy.
In the process fix a signed vs unsigned compare. If the vertex shader
has max_reg=-1, MAX2() vs an unsigned would not give the desired result.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Triggered by shaders like:
FRAG
PROPERTY FS_COLOR0_WRITES_ALL_CBUFS 1
DCL OUT[0], COLOR
DCL CONST[0]
DCL TEMP[0..2], LOCAL
0: IF CONST[0].xxxx :0
1: MOV TEMP[0], TEMP[1]
2: ELSE :0
3: MOV TEMP[0], TEMP[2]
4: ENDIF
5: MOV OUT[0], TEMP[0]
6: END
not really a sane shader, although driver segfaulting is probably
not the appropriate response.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Move the bits we want to share between generations from fd3_program to
ir3_shader. So overall structure is:
fdN_shader_stateobj -> ir3_shader -> ir3_shader_variant -> ir3
|- ...
\- ir3_shader_variant -> ir3
So the ir3_shader becomes the topmost generation neutral object, which
manages the set of variants each of which generates, compiles, and
assembles it's own ir.
There is a bit of additional renaming to s/fd3_compiler/ir3_compiler/,
etc.
Keep the split between the gallium level stateobj and the shader helper
object because it might be a good idea to pre-compute some generation
specific register values (ie. anything that is independent of linking).
Signed-off-by: Rob Clark <robclark@freedesktop.org>
2014-07-25 13:29:28 -04:00
Renamed from src/gallium/drivers/freedreno/a3xx/ir3_ra.c (Browse further)