fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-08 02:38:04 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

brw: evict memory for workgroup scope in Xe2 and newer

Browse source 
On Xe2 and Xe3, the flushing is necessary due to aliasing of TGM data
in L1 memory (HSD 14020414266). On newer platforms, it is necessary
for proper post-format data conversion handling (HSD 22020984324).

See the Instruction_Fence page (63969) for documentation on the fact
that the threadgroup scope ignores flushes.

Thanks to Francisco Jerez and Kenneth Graunke on their help for this
patch.

v2: restrict the flushing to TGM (Lionel).

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
This commit is contained in:
Paulo Zanoni 2026-03-26 13:16:01 -07:00
parent 4dc4012c4c
commit 82b1afe732
1 changed files with 13 additions and 4 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

17
src/intel/compiler/brw/brw_from_nir.cpp
View file

@ -4614,7 +4614,7 @@ emit_fence(const brw_builder &bld, enum opcode opcode,
static uint32_t
lsc_fence_descriptor_for_intrinsic(const struct intel_device_info *devinfo,
nir_intrinsic_instr *instr)
nir_intrinsic_instr *instr, bool is_tgm)
{
assert(devinfo->has_lsc);
@ -4640,8 +4640,17 @@ lsc_fence_descriptor_for_intrinsic(const struct intel_device_info *devinfo,
flush_type = LSC_FLUSH_TYPE_EVICT;
break;
case SCOPE_WORKGROUP:
scope = LSC_FENCE_THREADGROUP;
flush_type = LSC_FLUSH_TYPE_NONE;
/* On Xe2 and Xe3 we need the eviction due to aliasing of TGM data
* in L1 (HSD 14020414266). On Xe3p we need this due to how data
* post-format conversion happens (HSD 22020984324).
* Also, we have to upgrade the scope to TILE since flush_type is
* ignored for threadgroup fences, which means we'll use the
* values alaready initialized.
*/
if (devinfo->ver < 20 || !is_tgm) {
scope = LSC_FENCE_THREADGROUP;
flush_type = LSC_FLUSH_TYPE_NONE;
}
break;
case SCOPE_SHADER_CALL:
case SCOPE_INVOCATION:
@ -5090,7 +5099,7 @@ brw_from_nir_emit_intrinsic(nir_to_brw_state &ntb,
if (devinfo->has_lsc) {
assert(devinfo->verx10 >= 125);
uint32_t desc =
lsc_fence_descriptor_for_intrinsic(devinfo, instr);
lsc_fence_descriptor_for_intrinsic(devinfo, instr, tgm_fence);
if (ugm_fence) {
fence_regs[fence_regs_count++] =
emit_fence(ubld1, opcode, BRW_SFID_UGM, desc,