There were hacks in EmitCopyBlit before to adjust offsets so that y=0 after
the offsets had been adjusted for a negative pitch. It appears that those
hacks were due to an unclear and surprising aspect of the hardware: inverting
the pitch results in the blit into the specified rectangle being inverted,
without the user needing to adjust y and base offset.
Tested with piglit copytexsubimage test on 915GM and GM965. Should fix
serious performance issues with ETQW and other applications.
We now express arrays in terms of indirect addressing. For example:
dst = a[i];
becomes:
MOV dst, TEMP[1 + TEMP[2].y];
At instruction-emit time indirect addressing is converted into ARL/
ADDR-relative form:
ARL ADDR.x, TEMP[2].y;
MOV dst, TEMP[1 + ADDR.x];
This fixes a number of array-related issues. Arrays of arrays and complex
array/struct nesting works now.
There may be some regressions, but more work is coming.
This is a step toward better array handling code. In particular, when more
than one operand of an instruction uses indirect addressing, we'll need some
temporary instructions and registers. By converting IR storage to instruction
operands all in one place (emit_instruction()) we can be smarter about this.
Also, somewhat better handling of dst register swizzle/writemask handling.
This results in tighter writemasks on some instructions which is good for
SOA execution.
And, cleaner instruction commenting with inst_comment().
Next: remove some more dead code and additional clean-ups...
As far as I can read in the docs, VS threads can be 1:1 with the pairs of
VUE handles allocated for them. Also, G4X can run twice as many threads as
before (though we won't unless the we bump the preferred URB entries for VS).
We were dividing the number of URB entries by two to get number of threads,
which looks suspiciously like a copy'n'paste-o from brw_vs_state.c. Also, the
maximum number of threads is 24, not 12.
The clip thread could potentially deadlock when processing tristrips since
being moved back to dual-thread mode, as the two threads could each have 4 VUEs
referenced and not be able to allocate another one since SF processing
wasn't able to continue (needing 5 entries before it freed 2).
In constrained URB mode, similar deadlock could even have occurred with
polygons (so we cut back max_threads if we can't handle it any primitive type).
This ensures all batchbuffers have a same cliprect mode after calling
_intel_batchbuffer_flush even if there aren't invalid commands in the
current batch buffer. (fix bug#18362).
See bug 18445.
When getting array results, __glXReadReply() always reads a multiple of
four bytes. This can cause writing to invalid memory when 'n' is not a
multiple of four.
Special-case the glAreTexturesResident() functions now.
To fix the bug, we use a temporary buffer that's a multiple of four bytes
in length.
NOTE: this commit also reverts part of commit 919ec22ecf
(glx/x11: Added some #ifdef GLX_DIRECT_RENDERING protection) which
directly edited the indirect.c file rather than the python generator!
I'm not repairing that issue at this time.
