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

freedreno/ir3: adjust condition for when to use ldib

Browse Source 
We have to use it any time that the image is writable. Otherwise writes
from the same invocation won't have posted into the texture cache.

See: https://gitlab.freedesktop.org/mesa/mesa/-/issues/5629
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13807>
This commit is contained in:
Ilia Mirkin
2021-11-15 23:41:59 -05:00
committed by Marge Bot
parent 011ea32585
commit aa93896156
3 changed files with 3 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/freedreno/ci/freedreno-a530-fails.txt
-3
View File
@@ -195,9 +195,6 @@ KHR-GLES3.packed_pixels.varied_rectangle.rgba32f,Fail
KHR-GLES3.packed_pixels.varied_rectangle.rgba32i,Fail
KHR-GLES3.packed_pixels.varied_rectangle.rgba32ui,Fail
# "[31] Check failed. Received: [1,0,0,2] instead of: [5,0,0,2]"
KHR-GLES31.core.shader_image_load_store.basic-glsl-misc-fs,Fail
glx@glx-make-current,Crash
glx@glx-multi-window-single-context,Fail
glx@glx-query-drawable-glx_fbconfig_id-window,Fail
src/freedreno/ci/freedreno-a630-flakes.txt
-7
View File
@@ -109,13 +109,6 @@ KHR-GLES31.core.tessellation_shader.tessellation_shader_tc_barriers.barrier_guar
# looks like a cache flushing issue, and it does sometimes pass.
bypass-dEQP-GLES31.functional.blend_equation_advanced.msaa.*
# Testcase was mostly fixed in 23f7e06cd8d40569f8bfabde9c01d1597573abef, but has
# flaked in CI since then:
# " [775] Check failed. Received: [3,0,0,2] instead of: [5,0,0,2]
# [806] Check failed. Received: [3,0,0,2] instead of: [5,0,0,2]
# ..."
KHR-GLES31.core.shader_image_load_store.basic-glsl-misc-fs
# Flakes, all seen since merge of:
# https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12258
# Failures seen so far in different flakes:
src/freedreno/ir3/ir3_compiler_nir.c
+3 -3
View File
@@ -1293,10 +1293,10 @@ static void
emit_intrinsic_load_image(struct ir3_context *ctx, nir_intrinsic_instr *intr,
struct ir3_instruction **dst)
{
/* Coherent accesses have to go directly to memory, rather than through
* ISAM's texture cache (which isn't coherent with image stores).
/* If the image can be written, must use LDIB to retrieve data, rather than
* through ISAM (which uses the texture cache and won't get previous writes).
*/
if (nir_intrinsic_access(intr) & ACCESS_COHERENT && ctx->compiler->gen >= 5) {
if (!(nir_intrinsic_access(intr) & ACCESS_NON_WRITEABLE) && ctx->compiler->gen >= 5) {
ctx->funcs->emit_intrinsic_load_image(ctx, intr, dst);
return;
}