diff options
author | Daniel Schürmann <daniel@schuermann.dev> | 2021-02-05 18:25:18 +0100 |
---|---|---|
committer | Dylan Baker <dylan.c.baker@intel.com> | 2021-02-23 10:13:31 -0800 |
commit | 17aac7926a8bf06873fc8b9b5a03a3508131714f (patch) | |
tree | ea8d6ee3bf7df6c66a005cfa516ae2050daf22f5 | |
parent | a0b7f2318517124778a489e4ae0b970f89a677d8 (diff) |
aco: fix shared VGPR allocation on RDNA2
VGPRs are now allocated in blocks of 8 normal
or 16 shared VGPRs, respectively.
Fixes: 14a5021aff661a26d76f330fec55d400d35443a8 ('aco/gfx10: Refactor of GFX10 wave64 bpermute.')
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8921>
(cherry picked from commit bacc3b36f55ed1f0fbb8a7efdfb072a0f3ee4ee1)
-rw-r--r-- | .pick_status.json | 2 | ||||
-rw-r--r-- | src/amd/compiler/aco_instruction_selection.cpp | 5 |
2 files changed, 4 insertions, 3 deletions
diff --git a/.pick_status.json b/.pick_status.json index c9e47ea2620..52c59a1ed07 100644 --- a/.pick_status.json +++ b/.pick_status.json @@ -5836,7 +5836,7 @@ "description": "aco: fix shared VGPR allocation on RDNA2", "nominated": true, "nomination_type": 1, - "resolution": 0, + "resolution": 1, "master_sha": null, "because_sha": "14a5021aff661a26d76f330fec55d400d35443a8" }, diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp index 9824324aa0a..142c75c2b48 100644 --- a/src/amd/compiler/aco_instruction_selection.cpp +++ b/src/amd/compiler/aco_instruction_selection.cpp @@ -196,8 +196,9 @@ static Temp emit_bpermute(isel_context *ctx, Builder &bld, Temp index, Temp data /* GFX10 wave64 mode: emulate full-wave bpermute */ if (!ctx->has_gfx10_wave64_bpermute) { ctx->has_gfx10_wave64_bpermute = true; - ctx->program->config->num_shared_vgprs = 8; /* Shared VGPRs are allocated in groups of 8 */ - ctx->program->vgpr_limit -= 4; /* We allocate 8 shared VGPRs, so we'll have 4 fewer normal VGPRs */ + /* Shared VGPRs are allocated in groups of 8/16 */ + ctx->program->config->num_shared_vgprs = ctx->program->chip_class >= GFX10_3 ? 16 : 8; + ctx->program->vgpr_limit -= ctx->program->chip_class >= GFX10_3 ? 8 : 4; } Temp index_is_lo = bld.vopc(aco_opcode::v_cmp_ge_u32, bld.def(bld.lm), Operand(31u), index); |