KonstantinSeurer
/
mesa
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

intel/compiler: balanced tileY/linear friendly LID order for CS 

Fixes perf regression introduced from tileY LID order for CS
shaders that access both textures and buffers. Walks LIDs in
X-major fashion, but with blocks of height 4. This maps LIDs per
HW thread for SIMD8/16/32 as (2x4/4x4/8x4), which is always good
for tileY resources and usually good for linear resources.

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10733>
This commit is contained in:
Felix DeGrood 2021-05-19 09:50:45 -07:00 committed by Marge Bot
parent c23e2a662a
commit 380fa050f2
1 changed files with 25 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
src/intel/compiler/brw_nir_lower_cs_intrinsics.c
View File

@ -120,6 +120,26 @@ lower_cs_intrinsics_convert_block(struct lower_intrinsics_state *state,
id_x = nir_umod(b, linear, size_x);
id_y = nir_umod(b, nir_udiv(b, linear, size_x), size_y);
local_index = linear;
} else if (!nir->info.cs.local_size_variable &&
nir->info.cs.local_size[1] % 4 == 0) {
/* 1x4 block X-major lid order. Same as X-major except increments in
* blocks of width=1 height=4. Always optimal for tileY and usually
* optimal for linear accesses.
* x = (linear / 4) % size_x
* y = ((linear % 4) + (linear / 4 / size_x) * 4) % size_y
* X,Y ordering will look like: (0,0) (0,1) (0,2) (0,3) (1,0) (1,1)
* (1,2) (1,3) (2,0) ... (size_x-1,3) (0,4) (0,5) (0,6) (0,7) (1,4) ...
*/
const unsigned height = 4;
nir_ssa_def *block = nir_udiv_imm(b, linear, height);
id_x = nir_umod(b, block, size_x);
id_y = nir_umod(b,
nir_iadd(b,
nir_umod(b, linear, nir_imm_int(b, height)),
nir_imul_imm(b,
nir_udiv(b, block, size_x),
height)),
size_y);
} else {
/* Y-major lid order. Optimal for tileY accesses only,
* which are usually images. X,Y ordering will look like:
@ -127,12 +147,15 @@ lower_cs_intrinsics_convert_block(struct lower_intrinsics_state *state,
*/
id_y = nir_umod(b, linear, size_y);
id_x = nir_umod(b, nir_udiv(b, linear, size_y), size_x);
}
id_z = nir_udiv(b, linear, size_xy);
local_id = nir_vec3(b, id_x, id_y, id_z);
if (!local_index) {
local_index = nir_iadd(b, nir_iadd(b, id_x,
nir_imul(b, id_y, size_x)),
nir_imul(b, id_z, size_xy));
}
id_z = nir_udiv(b, linear, size_xy);
local_id = nir_vec3(b, id_x, id_y, id_z);
break;
case DERIVATIVE_GROUP_LINEAR:
/* For linear, just set the local invocation index linearly,