OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

intel/compiler: Micro optimize regions_overlap

Browse Source 
On my Ice Lake laptop (using a locked CPU speed and other measures to
prevent thermal throttling, etc.) using a release build, improves
performance of compiling shaders from batman_arkham_city_goty.foz by
-1.09% ± 0.084% (n = 5, pooled s = 0.354471)

Reduces the size of a release build by 26k.

   text	   data	    bss	    dec	    hex	filename
23163641 400720	 231360	23795721	16b1809	before/lib64/dri/iris_dri.so
23137264 400720	 231360	23769344	16ab100	after/lib64/dri/iris_dri.so

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22299>
This commit is contained in:
Ian Romanick
2023-03-13 19:47:07 -07:00
committed by Marge Bot
parent 7873edee6e
commit 78ee74de4a
1 changed files with 34 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
src/intel/compiler/brw_ir_fs.h
View File

@@ -212,6 +212,31 @@ reg_padding(const fs_reg &r)
return (MAX2(1, stride) - 1) * type_sz(r.type); return (MAX2(1, stride) - 1) * type_sz(r.type);
} }
/* Do not call this directly. Call regions_overlap() instead. */
static inline bool
regions_overlap_MRF(const fs_reg &r, unsigned dr, const fs_reg &s, unsigned ds)
{
if (r.nr & BRW_MRF_COMPR4) {
fs_reg t = r;
t.nr &= ~BRW_MRF_COMPR4;
/* COMPR4 regions are translated by the hardware during decompression
* into two separate half-regions 4 MRFs apart from each other.
*
* Note: swapping s and t in this parameter list eliminates one possible
* level of recursion (since the s in the called versions of
* regions_overlap_MRF can't be COMPR4), and that makes the compiled
* code a lot smaller.
*/
return regions_overlap_MRF(s, ds, t, dr / 2) ||
regions_overlap_MRF(s, ds, byte_offset(t, 4 * REG_SIZE), dr / 2);
} else if (s.nr & BRW_MRF_COMPR4) {
return regions_overlap_MRF(s, ds, r, dr);
}
return !((r.nr * REG_SIZE + r.offset + dr) <= (s.nr * REG_SIZE + s.offset) ||
(s.nr * REG_SIZE + s.offset + ds) <= (r.nr * REG_SIZE + r.offset));
}
/** /**
* Return whether the register region starting at \p r and spanning \p dr * Return whether the register region starting at \p r and spanning \p dr
* bytes could potentially overlap the register region starting at \p s and * bytes could potentially overlap the register region starting at \p s and
@@ -220,22 +245,17 @@ reg_padding(const fs_reg &r)
static inline bool static inline bool
regions_overlap(const fs_reg &r, unsigned dr, const fs_reg &s, unsigned ds) regions_overlap(const fs_reg &r, unsigned dr, const fs_reg &s, unsigned ds)
{ {
if (r.file == MRF && (r.nr & BRW_MRF_COMPR4)) { if (r.file != s.file)
fs_reg t = r; return false;
t.nr &= ~BRW_MRF_COMPR4;
/* COMPR4 regions are translated by the hardware during decompression
* into two separate half-regions 4 MRFs apart from each other.
*/
return regions_overlap(t, dr / 2, s, ds) ||
regions_overlap(byte_offset(t, 4 * REG_SIZE), dr / 2, s, ds);
} else if (s.file == MRF && (s.nr & BRW_MRF_COMPR4)) { if (r.file == VGRF) {
return regions_overlap(s, ds, r, dr); return r.nr == s.nr &&
!(r.offset + dr <= s.offset || s.offset + ds <= r.offset);
} else { } else if (r.file != MRF) {
return reg_space(r) == reg_space(s) && return !(reg_offset(r) + dr <= reg_offset(s) ||
!(reg_offset(r) + dr <= reg_offset(s) ||
reg_offset(s) + ds <= reg_offset(r)); reg_offset(s) + ds <= reg_offset(r));
} else {
return regions_overlap_MRF(r, dr, s, ds);
} }
} }