On Tue, Jul 30, 2024 at 08:33:13PM +0530, Nitin Gote wrote:
> Wa_14019789679 implementation for MTL, ARL and DG2.
>
> v2: Corrected condition
>
> v3:
> - Fix indentation (Jani Nikula)
> - dword size should be 0x1 and
> initialize dword to 0 instead of MI_NOOP (Tejas)
> - Use IS_GFX_GT_IP_RANGE() (Tejas)
>
> Bspec: 47083
>
> Signed-off-by: Nitin Gote <nitin.r.gote@xxxxxxxxx>
> ---
> drivers/gpu/drm/i915/gt/intel_gpu_commands.h | 1 +
> drivers/gpu/drm/i915/gt/intel_workarounds.c | 14 +++++++++++++-
> 2 files changed, 14 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
> index 2bd8d98d2110..7eaf7eddd25b 100644
> --- a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
> +++ b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
> @@ -220,6 +220,7 @@
> #define GFX_OP_DESTBUFFER_INFO ((0x3<<29)|(0x1d<<24)|(0x8e<<16)|1)
> #define GFX_OP_DRAWRECT_INFO ((0x3<<29)|(0x1d<<24)|(0x80<<16)|(0x3))
> #define GFX_OP_DRAWRECT_INFO_I965 ((0x7900<<16)|0x2)
> +#define CMD_3DSTATE_MESH_CONTROL ((0x3 << 29) | (0x3 << 27) | (0x0 << 24) | (0x77 << 16) | (0x1))
>
> #define XY_CTRL_SURF_INSTR_SIZE 5
> #define MI_FLUSH_DW_SIZE 3
> diff --git a/drivers/gpu/drm/i915/gt/intel_workarounds.c b/drivers/gpu/drm/i915/gt/intel_workarounds.c
> index 09a287c1aedd..d942707381be 100644
> --- a/drivers/gpu/drm/i915/gt/intel_workarounds.c
> +++ b/drivers/gpu/drm/i915/gt/intel_workarounds.c
> @@ -974,7 +974,12 @@ int intel_engine_emit_ctx_wa(struct i915_request *rq)
> if (ret)
> return ret;
>
> - cs = intel_ring_begin(rq, (wal->count * 2 + 2));
> + if (IS_GFX_GT_IP_RANGE(rq->engine->gt, IP_VER(12, 70), IP_VER(12, 74)) ||
> + IS_DG2(rq->i915))
> + cs = intel_ring_begin(rq, (wal->count * 2 + 4));
The 3DSTATE_MESH_CONTROL instruction is 3 dwords long (one header dword,
two body dwords) so going from +2 up to +4 doesn't cover it completely.
You'd need +5 to cover the entire instruction, but I think we want the
output here to remain qword-aligned, so you probably actually want +6.
> + else
> + cs = intel_ring_begin(rq, (wal->count * 2 + 2));
> +
> if (IS_ERR(cs))
> return PTR_ERR(cs);
>
> @@ -1004,6 +1009,13 @@ int intel_engine_emit_ctx_wa(struct i915_request *rq)
> }
> *cs++ = MI_NOOP;
>
> + /* Wa_14019789679 */
> + if (IS_GFX_GT_IP_RANGE(rq->engine->gt, IP_VER(12, 70), IP_VER(12, 74)) ||
> + IS_DG2(rq->i915)) {
> + *cs++ = CMD_3DSTATE_MESH_CONTROL;
> + *cs++ = 0;
As noted above, the body is two dwords (64-bits), so we want to emit one
more empty dword here to finish the mesh control command, and then an
MI_NOOP after it to preserve qword alignment.
Matt
> + }
> +
> intel_uncore_forcewake_put__locked(uncore, fw);
> spin_unlock(&uncore->lock);
> intel_gt_mcr_unlock(wal->gt, flags);
> --
> 2.25.1
>
--
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation
