This way we ensure to have a single place where these are handled. The
immediate benefit is that now line numbers are always printed out, which
is quite handy.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Unfortunately, it's all a walk in the park. Both, internal code in the
assembler and external shaders (libva) generate registers that trigger
assertions in brw_eu_emit.c's brw_validate().
To fix all that I took the option to be able to emit warning with the -W
flag but still make the assembler generate the same opcodes.
We can fix all this, but it requires validation, something that I cannot
do right now.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Another check (that we hit if we try to use brw_set_src0()). Again,
protect it with the -W option.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
The list of region restrictions in bspec do say that we can't have:
width == 1 && hstrize != 0
We do have plenty of assembly code that don't respect that behaviour. So
let's hide the warning under a -W flag (for now) while we fix things.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Now that we have locations, we can write error() and warn() functions
giving more information about where it's going wrong.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Let's generate location information about the tokens we are parsing.
This can be used to give accurate location when reporting errors and
warnings.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
There's no real need to warn when the same register is declared twice.
Currently the libva driver does do that and this warning makes other
errors really hide in a sea of warnings.
Redefining a register with different parameters is a real error though,
so we should not allow that and error out in that case.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Of course the assertion is there to make sure GRF and MRF have a reg.nr
< 128. To exclude ARF registers, reg.file has be checked, not reg.type
(channel type). Most likely a typo never caught.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
The goal is to use brw_set_src[01](), so let's start by validating the
register we have before generating the opcode.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
A few notes:
I needed to introduce a brw context and compile structs. These are only
used to get which generation we are compiling code for, but eventually
we can use more of the infrastructure.
brw_set_dest() uses the destination register width to program the
instruction execution size.
The assembler can either take subnr in bytes or in number of elements,
so we need a resolve step when setting a brw_reg.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
The goal is to use brw_set_dest(), so let's start by validating the
register we have before generating the opcode.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
One more step on the road to replacing all register-like structures by
struct brw_reg.
Two things in this commit are worth noting:
* As we are using more and more brw_reg, a lot of the field-by-field
assignments can be replaced by 1 assignment which results is a
reduction of code
* As the destination horizontal stride is now stored on 2 bits in
brw_reg, it's not possible to defer the handling of DEFAULT_DSTREGION
(aka (int)-1) when setting the destination operand. It has to be done
when parsing the region and resolve_dst_region() is a helper for that
task.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
More code simplification can be layered on top of that (by using some
brw_* helpers to create registers), that'd be for another commit.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
More code simplification can be layered on top of that (by using some
brw_* helpers to create registers), that'd be for another commit.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
It's time to start converting the emission code in gram.y to use libbrw
infrastructure. Let's start with using brw_reg for declared register.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Until now, the assembler had relocation-related fields added to struct
brw_instruction. This changes the size of the structure and break code
assuming the opcode structure is really 16 bytes, for instance the
emission code in brw_eu_emit.c.
With this commit, we build on the infrastructure that slowly emerged in
the few previous commits to add a relocatable instruction with the
needed fields.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
The output of the parsing is a list of struct brw_program_instruction.
These instructions can be either GEN instructions aka struct
brw_instruction or labels. To make this more explicit we now have a type
to test to determine which instruction we are dealing with.
This will also allow to to pull the relocation bits into struct
brw_program_instruction instead of having them in the structure
representing the opcodes.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Factoring out the code from the grammar will allow us to switch to
using brw_compile in a cleaner way.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
No need to use a brw_program_instruction there as a brw_instruction is
what you really dump anyway, espcially when the plan is to use
brw_compile from Mesa sooner rather than later.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
This function can only be called to resolve subreg_nr in direct mode
(there is an other function for the indirect case) and it makes no sense
to call it with an immediate operand.
Express those facts with asserts and simplify the logic.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Another step towards using struct brw_reg for source and destination
operands.
Instead of having a separate field to store the sub register number of
the address register in indirect access mode, we can reuse the subreg_nr
field that was only used for direct access so far.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
writemask_set gets in the way of switching to using struct brw_reg and
it's possible to derive it from the writemask value.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
To be able to import brw_eu.c and brw_eu_emit.c later on. This could be
used to get the assembler generate compact instructions at some point.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
With the brw_* files imported from mesa.
There are still a few things in that library that needs gen4asm.h, for
instance the GLuint and GLint types. The hope is that eventually libbrw
can be split out in its own directory and shared.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
A lot of the mesa code use struct brw_context to get the GPU generation
and various information. Let's stub this structure and initialize it
ourselves to be able to resuse mesa's code untouched.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
This also add a new brw_compat.h that should help maintaining the
diff between mesa's version and our as small as possible.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
From Mesa. This imports a bit more the of brw_eu* infrastructure (which
is going towards the right direction!) from mesa and the update is quite
a significant improvement over what we had.
I also verified that the changes that were done on the assembler old
version of brw_disasm.c were already supported by the Mesa version, and
indeed they were.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Finally merge both brw_structs.h from mesa. One detail has risen in that
last commit, the msg_control field of data port message descriptors.
Mesa's msg_control field is sometimes split with messages-specific
fields where the assembler (at least for recent generations) exposes the
full msg_control value in the send instruction.
As libva's shaders encodes the full msg_control value in its send
instructions, I've chosen to not take the split msg_control from mesa.
It's absolutely possible to have a patch fixing that divergence at some
later point.
I've also kept a hack introduced with ironlake to not have to rewrite
shaders (that encode msg_control in the text, remember), and thus
creates a another difference with Mesa.
- GLuint msg_control:3;
- GLuint msg_type:3;
+ GLuint msg_control:4;
+ GLuint msg_type:2;
Once again, I've made sure that re-generating libva's shaders don't show
any difference.
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
