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

aco: swap the correct v_mov_b32 if there are two of them

Browse Source 
Previously, this function tried to swap the instruction which is not
v_mov_b32, so that it doesn't introduce any new OPY-only instructions. If
both were v_mov_b32, it swapped Y. Since this makes Y opy-only, this can't
be done if X is also opy-only.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Fixes: 408fa33c0928 ("aco/gfx12: don't use second VALU for VOPD's OPX if there is a WaR")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/13101
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34841>
(cherry picked from commit 9ca71b52aa5bc2eda1f08149f7780e59858ee27b)
This commit is contained in:
Rhys Perry
2025-05-06 16:47:44 +01:00
committed by Eric Engestrom
parent 9c49a2e03f
commit 77e7fd0dee
3 changed files with 15 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.pick_status.json
View File

@@ -104,7 +104,7 @@
"description": "aco: swap the correct v_mov_b32 if there are two of them",
"nominated": true,
"nomination_type": 2,
"resolution": 0,
"resolution": 1,
"main_sha": null,
"because_sha": "408fa33c092810155baac342de90fd712231aa89",
"notes": null

6
src/amd/compiler/aco_scheduler_ilp.cpp
View File

@@ -758,13 +758,17 @@ create_vopd_instruction(const SchedILPContext& ctx, unsigned idx, bool prev_can_
if (x_info.src_banks & y_info.src_banks) {
assert(x_info.is_commutative || y_info.is_commutative);
/* Avoid swapping v_mov_b32 because it will become an OPY-only opcode. */
if (x_info.op == aco_opcode::v_dual_mov_b32 && !y_info.is_commutative) {
if (x_info.op == aco_opcode::v_dual_mov_b32 && y_info.op == aco_opcode::v_dual_mov_b32) {
swap_x = x_info.is_opy_only;
swap_y = !swap_x;
} else if (x_info.op == aco_opcode::v_dual_mov_b32 && !y_info.is_commutative) {
swap_x = true;
x_info.is_opy_only = true;
} else {
swap_x = x_info.is_commutative && x_info.op != aco_opcode::v_dual_mov_b32;
swap_y = y_info.is_commutative && !swap_x;
}
y_info.is_opy_only |= swap_y && y_info.op == aco_opcode::v_dual_mov_b32;
}
if (x_info.is_opy_only) {

9
src/amd/compiler/tests/test_scheduler.cpp
View File

@@ -162,6 +162,7 @@ BEGIN_TEST(vopd_sched.war)
PhysReg reg_v0{256};
PhysReg reg_v1{257};
PhysReg reg_v3{259};
PhysReg reg_v5{261};
//>> p_unit_test 0
//~gfx11! v1: %0:v[1] = v_dual_add_f32 %0:v[3], %0:v[1] :: v1: %0:v[0] = v_dual_mul_f32 %0:v[1], %0:v[3]
@@ -184,6 +185,14 @@ BEGIN_TEST(vopd_sched.war)
bld.vop2(aco_opcode::v_mul_f32, Definition(reg_v1, v1), Operand(reg_v3, v1),
Operand(reg_v1, v1));
/* Test that we swap the right v_mov_b32. */
//>> p_unit_test 2
//~gfx11! v1: %0:v[1] = v_dual_mov_b32 %0:v[5] :: v1: %0:v[0] = v_dual_add_nc_u32 0, %0:v[1]
//~gfx12! v1: %0:v[0] = v_dual_mov_b32 %0:v[1] :: v1: %0:v[1] = v_dual_add_nc_u32 0, %0:v[5]
bld.pseudo(aco_opcode::p_unit_test, Operand::c32(2));
bld.vop1(aco_opcode::v_mov_b32, Definition(reg_v0, v1), Operand(reg_v1, v1));
bld.vop1(aco_opcode::v_mov_b32, Definition(reg_v1, v1), Operand(reg_v5, v1));
finish_schedule_vopd_test();
}
END_TEST