f49ef1410d
- remove patches that are in upstream - remove vdpau as upstream removed it - update version of dependencies - update rust libwrap filename - Update libclc to 22.1 has the 21.1.8 doesn't build on centos stream 9 - Fix python issues with 3.9 (Mesa requires 3.10) - Revert Freedreno tu_autotune to previous C implementation, as C++ implementation - Remove some kmsro driver on x86_64 Resolves: RHEL-135263 Signed-off-by: Jocelyn Falempe <jfalempe@redhat.com>
94 lines
4.2 KiB
Diff
94 lines
4.2 KiB
Diff
From db5315272229f2df847d663aec61c4204627adc4 Mon Sep 17 00:00:00 2001
|
|
From: Jocelyn Falempe <jfalempe@redhat.com>
|
|
Date: Fri, 26 Jun 2026 11:03:15 +0200
|
|
Subject: [PATCH 14/19] Revert "tu/autotune: Disable autotuning for small
|
|
renderpasses by default"
|
|
|
|
This reverts commit bf2777c0130954f969c5a1038a01f94ec2d1ac63.
|
|
---
|
|
docs/drivers/freedreno.rst | 10 +++++-----
|
|
src/freedreno/vulkan/tu_autotune.cc | 19 +++++++------------
|
|
2 files changed, 12 insertions(+), 17 deletions(-)
|
|
|
|
diff --git a/docs/drivers/freedreno.rst b/docs/drivers/freedreno.rst
|
|
index 9b1992fd628..a2318559526 100644
|
|
--- a/docs/drivers/freedreno.rst
|
|
+++ b/docs/drivers/freedreno.rst
|
|
@@ -709,12 +709,12 @@ environment variables:
|
|
GMEM rendering due to less overhead from tiling. This tends to lead to worse
|
|
performance in most cases, so it's only useful for testing.
|
|
|
|
- ``tune_small``
|
|
- Enables tuning for small render passes (those with a small number of draw
|
|
- calls). By default, small RPs always use SYSMEM mode as they generally don't
|
|
- benefit from GMEM rendering due to the overhead of tiling.
|
|
+ ``small_sysmem``
|
|
+ Always chooses SYSMEM rendering if the amount of draw calls in the render pass
|
|
+ is lower than a certain threshold. The benefits of GMEM rendering are less
|
|
+ pronounced in these smaller RPs and SYSMEM rendering tends to win more often.
|
|
|
|
Multiple flags can be combined by separating them with commas, e.g.
|
|
- ``TU_AUTOTUNE_FLAGS=big_gmem,tune_small``.
|
|
+ ``TU_AUTOTUNE_FLAGS=big_gmem,small_sysmem``.
|
|
|
|
If no flags are specified, the default behavior is used.
|
|
\ No newline at end of file
|
|
diff --git a/src/freedreno/vulkan/tu_autotune.cc b/src/freedreno/vulkan/tu_autotune.cc
|
|
index c708c83c044..38d09f5db45 100644
|
|
--- a/src/freedreno/vulkan/tu_autotune.cc
|
|
+++ b/src/freedreno/vulkan/tu_autotune.cc
|
|
@@ -97,8 +97,8 @@ enum class tu_autotune::algorithm : uint8_t {
|
|
|
|
/* Modifier flags, these modify the behavior of the autotuner in a user-defined way. */
|
|
enum class tu_autotune::mod_flag : uint8_t {
|
|
- BIG_GMEM = BIT(1), /* All RPs with >= 10 draws use GMEM. */
|
|
- TUNE_SMALL = BIT(2), /* Try tuning all RPs with <= 5 draws, ignored by default. */
|
|
+ BIG_GMEM = BIT(1), /* All RPs with >= 10 draws use GMEM. */
|
|
+ SMALL_SYSMEM = BIT(2), /* All RPs with <= 5 draws use SYSMEM. */
|
|
};
|
|
|
|
/* Metric flags, for internal tracking of enabled metrics. */
|
|
@@ -198,7 +198,7 @@ struct PACKED tu_autotune::config_t {
|
|
|
|
str += ", Mod Flags: 0x" + std::to_string(mod_flags) + " (";
|
|
MODF_STR(BIG_GMEM);
|
|
- MODF_STR(TUNE_SMALL);
|
|
+ MODF_STR(SMALL_SYSMEM);
|
|
str += ")";
|
|
|
|
str += ", Metric Flags: 0x" + std::to_string(metric_flags) + " (";
|
|
@@ -280,7 +280,7 @@ tu_autotune::get_env_config()
|
|
if (flags_env_str) {
|
|
static const struct debug_control tu_at_flags_control[] = {
|
|
{ "big_gmem", (uint32_t) mod_flag::BIG_GMEM },
|
|
- { "tune_small", (uint32_t) mod_flag::TUNE_SMALL },
|
|
+ { "small_sysmem", (uint32_t) mod_flag::SMALL_SYSMEM },
|
|
{ NULL, 0 }
|
|
};
|
|
|
|
@@ -1256,18 +1256,13 @@ tu_autotune::get_optimal_mode(struct tu_cmd_buffer *cmd_buffer, rp_ctx_t *rp_ctx
|
|
*/
|
|
bool simultaneous_use = cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT;
|
|
|
|
- /* These smaller RPs with few draws are too difficult to create a balanced hash for that can independently identify
|
|
- * them while not being so unique to not properly identify them across CBs. They're generally insigificant outside of
|
|
- * a few edge cases such as during deferred rendering G-buffer passes, as we don't have a good way to deal with those
|
|
- * edge cases yet, we just disable the autotuner for small RPs entirely for now unless TUNE_SMALL is specified.
|
|
- */
|
|
- bool ignore_small_rp = !config.test(mod_flag::TUNE_SMALL) && rp_state->drawcall_count < 5;
|
|
-
|
|
- if (!enabled || simultaneous_use || ignore_small_rp)
|
|
+ if (!enabled || simultaneous_use)
|
|
return default_mode;
|
|
|
|
if (config.test(mod_flag::BIG_GMEM) && rp_state->drawcall_count >= 10)
|
|
return render_mode::GMEM;
|
|
+ if (config.test(mod_flag::SMALL_SYSMEM) && rp_state->drawcall_count <= 5)
|
|
+ return render_mode::SYSMEM;
|
|
|
|
rp_key key(pass, framebuffer, cmd_buffer);
|
|
|
|
--
|
|
2.54.0
|
|
|