funkemunky/packages
1
0
Fork 0
mirror of https://github.com/terrapkg/packages.git synced 2026-05-31 09:01:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(mesa): backport AMD GPU artifact fix, cleanup and attribute patches (#3943) (#3944)

Browse Source 
* fix(mesa): backport AMD GPU artifact fix, cleanup and attribute patches

* cleanup and reorder

* Fix patch credit

Signed-off-by: Gilver <rockgrub@disroot.org>

---------

Signed-off-by: Gilver <rockgrub@disroot.org>
Co-authored-by: Gilver <rockgrub@disroot.org>
(cherry picked from commit 985de89ca8)

Co-authored-by: Antheas Kapenekakis <5252246+antheas@users.noreply.github.com>
This commit is contained in:
Raboneko
2025-03-15 10:38:07 -07:00
committed by GitHub
parent 1852f4d3e6
commit 7d0c06ebbc
7 changed files with 584 additions and 308 deletions
Download Patch File Download Diff File Expand all files Collapse all files
anda/lib/mesa/12633.patch
-26
View File
@@ -1,26 +0,0 @@
commit d0279e717ee740746f0770c5a9870d752108e756 (HEAD -> makepkg)
Author: Maarten Lankhorst <maarten.lankhorst@intel.com>
Date: Mon Feb 17 14:55:29 2025 -0800
HACK: drm/xe gamescope fix
Workaround gamescope DRM issues on Xe by allocating ANV_BO_ALLOC_SCANOUT
when using VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT
Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4292#note_2784316
Signed-off-by: Matthew Schwartz <matthew.schwartz@linux.dev>
diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
index 1884932bbc7..cbc1b4aad87 100644
--- a/src/intel/vulkan/anv_device.c
+++ b/src/intel/vulkan/anv_device.c
@@ -1533,6 +1533,9 @@ VkResult anv_AllocateMemory(
dedicated_info->image != VK_NULL_HANDLE) {
ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
+ if (image->vk.tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT)
+ alloc_flags |= ANV_BO_ALLOC_SCANOUT;
+
/* Apply implicit sync to be compatible with clients relying on
* implicit fencing. This matches the behavior in iris i915_batch
* submit. An example client is VA-API (iHD), so only dedicated
anda/lib/mesa/13a3f9a972324a72dd507e09ac975b969e6c88e0.patch
-34
View File
@@ -1,34 +0,0 @@
From 13a3f9a972324a72dd507e09ac975b969e6c88e0 Mon Sep 17 00:00:00 2001
From: Hans-Kristian Arntzen <post@arntzen-software.no>
Date: Tue, 25 Feb 2025 12:43:17 +0100
Subject: [PATCH] radv: Always set 0 dispatch offset for indirect CS.
Fixes severe glitching in Avowed.
Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Natalie Vock <natalie.vock@gmx.de>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33732>
---
src/amd/vulkan/radv_cmd_buffer.c | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
index d8c0c6810524b..d85c0659cb04e 100644
--- a/src/amd/vulkan/radv_cmd_buffer.c
+++ b/src/amd/vulkan/radv_cmd_buffer.c
@@ -11722,6 +11722,10 @@ radv_emit_dispatch_packets(struct radv_cmd_buffer *cmd_buffer, const struct radv
dispatch_initiator |= S_00B800_USE_THREAD_DIMENSIONS(1);
}
+ /* Indirect CS does not support offsets in the API. Must program this in case there have been
+ * preceding 1D RT dispatch or vkCmdDispatchBase. */
+ dispatch_initiator |= S_00B800_FORCE_START_AT_000(1);
+
if (grid_size_offset) {
if (device->load_grid_size_from_user_sgpr) {
assert(pdev->info.gfx_level >= GFX10_3);
--
GitLab
anda/lib/mesa/237d8799be3afe9a1e7ca9156a5d44ffe0aae681.patch
-38
View File
@@ -1,38 +0,0 @@
From 237d8799be3afe9a1e7ca9156a5d44ffe0aae681 Mon Sep 17 00:00:00 2001
From: Natalie Vock <natalie.vock@gmx.de>
Date: Fri, 28 Feb 2025 14:21:57 +0100
Subject: [PATCH] radv/rt: Limit monolithic pipelines to 50 stages
Beyond that, monolithic pipelines just bloat to incredible sizes,
destroying compile times for questionable, if any, runtime perf benefit.
Indiana Jones: The Great Circle has more than 100 stages and takes
several minutes to compile its RT pipeline on Deck when using monolithic
compilation, and yet separate shaders still end up faster (probably
because instruction cache coherency in traversal is better).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33818>
---
src/amd/vulkan/radv_pipeline_rt.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/src/amd/vulkan/radv_pipeline_rt.c b/src/amd/vulkan/radv_pipeline_rt.c
index 0658840966c30..bcda4aae5528d 100644
--- a/src/amd/vulkan/radv_pipeline_rt.c
+++ b/src/amd/vulkan/radv_pipeline_rt.c
@@ -602,7 +602,11 @@ radv_rt_compile_shaders(struct radv_device *device, struct vk_pipeline_cache *ca
bool library = pipeline->base.base.create_flags & VK_PIPELINE_CREATE_2_LIBRARY_BIT_KHR;
- bool monolithic = !library;
+ /* Beyond 50 shader stages, inlining everything bloats the shader a ton, increasing compile times and
+ * potentially even reducing runtime performance because of instruction cache coherency issues in the
+ * traversal loop.
+ */
+ bool monolithic = !library && pipeline->stage_count < 50;
for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
if (rt_stages[i].shader || rt_stages[i].nir)
continue;
--
GitLab
anda/lib/mesa/34005.patch
-46
View File
@@ -1,46 +0,0 @@
From c6270978411609f52afb8c6a2219fcb94f9013bb Mon Sep 17 00:00:00 2001
From: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Date: Tue, 11 Mar 2025 15:29:37 +0100
Subject: [PATCH] radv/amdgpu: fix device deduplication
To correctly deduplicate device inside the winsys, it should use the
fd or amdgpu_device_handle. Using the allocated ac_drm_device as key
is obviously broken.
Not deduplicating devices breaks memory budget and a bunch of games
were broken.
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12686
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12775
Fixes: a565f2994fe ("amd: move all uses of libdrm_amdgpu to ac_linux_drm")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34005>
---
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
index be8df8708c813..8b57abeb0b1ca 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
@@ -234,7 +234,7 @@ radv_amdgpu_winsys_create(int fd, uint64_t debug_flags, uint64_t perftest_flags,
goto fail;
}
- struct hash_entry *entry = _mesa_hash_table_search(winsyses, dev);
+ struct hash_entry *entry = _mesa_hash_table_search(winsyses, (void *)ac_drm_device_get_cookie(dev));
if (entry) {
ws = (struct radv_amdgpu_winsys *)entry->data;
++ws->refcount;
@@ -325,7 +325,7 @@ radv_amdgpu_winsys_create(int fd, uint64_t debug_flags, uint64_t perftest_flags,
radv_amdgpu_bo_init_functions(ws);
radv_amdgpu_cs_init_functions(ws);
- _mesa_hash_table_insert(winsyses, dev, ws);
+ _mesa_hash_table_insert(winsyses, (void *)ac_drm_device_get_cookie(dev), ws);
simple_mtx_unlock(&winsys_creation_mutex);
return &ws->base;
--
GitLab
anda/lib/mesa/bazzite.patch
+581
View File
@@ -0,0 +1,581 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Antheas Kapenekakis <git@antheas.dev>
Date: Sat, 15 Mar 2025 16:38:53 +0100
Subject: [NA] Developer files, readme, etc
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Antheas Kapenekakis <git@antheas.dev>
Date: Sat, 15 Mar 2025 16:39:08 +0100
Subject: [BEGIN] SteamOS Changes
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Date: Fri, 14 Jan 2022 15:58:45 +0100
Subject: STEAMOS: radv: min image count override for FH5
Otherwise in combination with the vblank time reservation in
gamescope the game could get stuck in low power states.
---
src/util/00-radv-defaults.conf | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/src/util/00-radv-defaults.conf b/src/util/00-radv-defaults.conf
index d2dbe4d5e11..1851504036a 100644
--- a/src/util/00-radv-defaults.conf
+++ b/src/util/00-radv-defaults.conf
@@ -220,5 +220,9 @@ Application bugs worked around in this file:
<application name="Total War: WARHAMMER III" application_name_match="TotalWarhammer3">
<option name="radv_disable_depth_storage" value="true"/>
</application>
+
+ <application name="Forza Horizon 5" application_name_match="ForzaHorizon5.exe">
+ <option name="vk_x11_override_min_image_count" value="4" />
+ </application>
</device>
</driconf>
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Date: Thu, 22 Feb 2024 22:32:45 +0100
Subject: STEAMOS: Dynamic swapchain override for gamescope limiter for DRI3
only
The original patch (from Bas) contained WSI VK support too but it's
been removed because the Gamescope WSI layer already handles that.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
---
.../frontends/dri/loader_dri3_helper.c | 42 ++++++++++++++++++-
.../frontends/dri/loader_dri3_helper.h | 1 +
2 files changed, 41 insertions(+), 2 deletions(-)
diff --git a/src/gallium/frontends/dri/loader_dri3_helper.c b/src/gallium/frontends/dri/loader_dri3_helper.c
index 9e4ca3f5707..7863623f8de 100644
--- a/src/gallium/frontends/dri/loader_dri3_helper.c
+++ b/src/gallium/frontends/dri/loader_dri3_helper.c
@@ -297,6 +297,30 @@ dri3_update_max_num_back(struct loader_dri3_drawable *draw)
}
}
+static unsigned
+gamescope_swapchain_override()
+{
+ const char *path = getenv("GAMESCOPE_LIMITER_FILE");
+ if (!path)
+ return 0;
+
+ static simple_mtx_t mtx = SIMPLE_MTX_INITIALIZER;
+ static int fd = -1;
+
+ simple_mtx_lock(&mtx);
+ if (fd < 0) {
+ fd = open(path, O_RDONLY);
+ }
+ simple_mtx_unlock(&mtx);
+
+ if (fd < 0)
+ return 0;
+
+ uint32_t override_value = 0;
+ pread(fd, &override_value, sizeof(override_value), 0);
+ return override_value;
+}
+
void
loader_dri3_set_swap_interval(struct loader_dri3_drawable *draw, int interval)
{
@@ -311,10 +335,12 @@ loader_dri3_set_swap_interval(struct loader_dri3_drawable *draw, int interval)
* PS. changing from value A to B and A < B won't cause swap out of order but
* may still gets wrong target_msc value at the beginning.
*/
- if (draw->swap_interval != interval)
+ if (draw->orig_swap_interval != interval)
loader_dri3_swapbuffer_barrier(draw);
- draw->swap_interval = interval;
+ draw->orig_swap_interval = interval;
+ if (gamescope_swapchain_override() != 1)
+ draw->swap_interval = interval;
}
static void
@@ -443,6 +469,12 @@ loader_dri3_drawable_init(xcb_connection_t *conn,
draw->swap_interval = dri_get_initial_swap_interval(draw->dri_screen_render_gpu);
+ draw->orig_swap_interval = draw->swap_interval;
+
+ unsigned gamescope_override = gamescope_swapchain_override();
+ if (gamescope_override == 1)
+ draw->swap_interval = 1;
+
dri3_update_max_num_back(draw);
/* Create a new drawable */
@@ -1087,6 +1119,12 @@ loader_dri3_swap_buffers_msc(struct loader_dri3_drawable *draw,
if (draw->type == LOADER_DRI3_DRAWABLE_WINDOW) {
dri3_fence_reset(draw->conn, back);
+ unsigned gamescope_override = gamescope_swapchain_override();
+ if (gamescope_override == 1)
+ draw->swap_interval = 1;
+ else
+ draw->swap_interval = draw->orig_swap_interval;
+
/* Compute when we want the frame shown by taking the last known
* successful MSC and adding in a swap interval for each outstanding swap
* request. target_msc=divisor=remainder=0 means "Use glXSwapBuffers()
diff --git a/src/gallium/frontends/dri/loader_dri3_helper.h b/src/gallium/frontends/dri/loader_dri3_helper.h
index 9061e9755e2..6cc64be298a 100644
--- a/src/gallium/frontends/dri/loader_dri3_helper.h
+++ b/src/gallium/frontends/dri/loader_dri3_helper.h
@@ -170,6 +170,7 @@ struct loader_dri3_drawable {
bool block_on_depleted_buffers;
bool queries_buffer_age;
int swap_interval;
+ int orig_swap_interval;
const struct loader_dri3_vtable *vtable;
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Date: Mon, 24 Feb 2025 17:48:21 +0100
Subject: radv: stop computing the UUID using the physical device cache key
Otherwise, the UUID changes for games that have shader-based drirc
workarounds and this breaks precompiled shaders on SteamDeck.
Instead, use this pdev cache key to compute the logical device hash
which is common to all pipelines.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
---
src/amd/vulkan/radv_device.c | 6 +++++-
src/amd/vulkan/radv_physical_device.c | 1 -
2 files changed, 5 insertions(+), 2 deletions(-)
diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index 2de839e5d6d..da732ae503e 100644
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -858,6 +858,7 @@ radv_device_init_cache_key(struct radv_device *device)
const struct radv_physical_device *pdev = radv_device_physical(device);
const struct radv_instance *instance = radv_physical_device_instance(pdev);
struct radv_device_cache_key *key = &device->cache_key;
+ struct mesa_blake3 ctx;
key->keep_shader_info = device->keep_shader_info;
key->trap_excp_flags = device->trap_handler_shader && instance->trap_excp_flags;
@@ -879,7 +880,10 @@ radv_device_init_cache_key(struct radv_device *device)
key->primitives_generated_query = true;
}
- _mesa_blake3_compute(key, sizeof(*key), device->cache_hash);
+ _mesa_blake3_init(&ctx);
+ _mesa_blake3_update(&ctx, &pdev->cache_key, sizeof(pdev->cache_key));
+ _mesa_blake3_update(&ctx, &device->cache_key, sizeof(device->cache_key));
+ _mesa_blake3_final(&ctx, device->cache_hash);
}
static void
diff --git a/src/amd/vulkan/radv_physical_device.c b/src/amd/vulkan/radv_physical_device.c
index f24203fcccc..b1a742d48ef 100644
--- a/src/amd/vulkan/radv_physical_device.c
+++ b/src/amd/vulkan/radv_physical_device.c
@@ -264,7 +264,6 @@ radv_device_get_cache_uuid(struct radv_physical_device *pdev, void *uuid)
return -1;
#endif
- _mesa_sha1_update(&ctx, &pdev->cache_key, sizeof(pdev->cache_key));
_mesa_sha1_final(&ctx, sha1);
memcpy(uuid, sha1, VK_UUID_SIZE);
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Antheas Kapenekakis <git@antheas.dev>
Date: Sat, 15 Mar 2025 16:39:25 +0100
Subject: [BEGIN] SteamOS Backports
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Natalie Vock <natalie.vock@gmx.de>
Date: Fri, 28 Feb 2025 14:21:57 +0100
Subject: radv/rt: Limit monolithic pipelines to 50 stages
Beyond that, monolithic pipelines just bloat to incredible sizes,
destroying compile times for questionable, if any, runtime perf benefit.
Indiana Jones: The Great Circle has more than 100 stages and takes
several minutes to compile its RT pipeline on Deck when using monolithic
compilation, and yet separate shaders still end up faster (probably
because instruction cache coherency in traversal is better).
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33818>
---
src/amd/vulkan/radv_pipeline_rt.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/src/amd/vulkan/radv_pipeline_rt.c b/src/amd/vulkan/radv_pipeline_rt.c
index 5a23dc99cc4..1421688d580 100644
--- a/src/amd/vulkan/radv_pipeline_rt.c
+++ b/src/amd/vulkan/radv_pipeline_rt.c
@@ -600,7 +600,11 @@ radv_rt_compile_shaders(struct radv_device *device, struct vk_pipeline_cache *ca
bool library = pipeline->base.base.create_flags & VK_PIPELINE_CREATE_2_LIBRARY_BIT_KHR;
- bool monolithic = !library;
+ /* Beyond 50 shader stages, inlining everything bloats the shader a ton, increasing compile times and
+ * potentially even reducing runtime performance because of instruction cache coherency issues in the
+ * traversal loop.
+ */
+ bool monolithic = !library && pipeline->stage_count < 50;
for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
if (rt_stages[i].shader || rt_stages[i].nir)
continue;
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Antheas Kapenekakis <git@antheas.dev>
Date: Sat, 15 Mar 2025 16:39:33 +0100
Subject: [BEGIN] Our Mesa backports
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Rhys Perry <pendingchaos02@gmail.com>
Date: Tue, 25 Feb 2025 18:07:30 +0000
Subject: aco: insert dependency waits in certain situations
This seems to fix some artifacts, but we're not sure why, so it might not
be a correct or optimal solution.
fossil-db (navi31):
Totals from 28424 (35.81% of 79377) affected shaders:
Instrs: 30112910 -> 30348977 (+0.78%); split: -0.00%, +0.78%
CodeSize: 159542980 -> 160485336 (+0.59%); split: -0.00%, +0.59%
Latency: 221438396 -> 221500856 (+0.03%); split: -0.00%, +0.03%
InvThroughput: 38154231 -> 38159984 (+0.02%); split: -0.00%, +0.02%
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Backport-to: 25.0
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33853>
---
src/amd/compiler/aco_insert_NOPs.cpp | 101 +++++++++++++++++++++++----
1 file changed, 87 insertions(+), 14 deletions(-)
diff --git a/src/amd/compiler/aco_insert_NOPs.cpp b/src/amd/compiler/aco_insert_NOPs.cpp
index de062be2c74..1005f82812c 100644
--- a/src/amd/compiler/aco_insert_NOPs.cpp
+++ b/src/amd/compiler/aco_insert_NOPs.cpp
@@ -259,6 +259,9 @@ struct NOP_ctx_gfx11 {
std::bitset<128> sgpr_read_by_valu_as_lanemask;
std::bitset<128> sgpr_read_by_valu_as_lanemask_then_wr_by_salu;
+ std::bitset<128> sgpr_read_by_valu_as_lanemask2;
+ std::bitset<128> sgpr_read_by_valu_as_lanemask_then_wr_by_valu;
+
/* WMMAHazards */
std::bitset<256> vgpr_written_by_wmma;
@@ -278,8 +281,11 @@ struct NOP_ctx_gfx11 {
valu_since_wr_by_trans.join_min(other.valu_since_wr_by_trans);
trans_since_wr_by_trans.join_min(other.trans_since_wr_by_trans);
sgpr_read_by_valu_as_lanemask |= other.sgpr_read_by_valu_as_lanemask;
+ sgpr_read_by_valu_as_lanemask2 |= other.sgpr_read_by_valu_as_lanemask2;
sgpr_read_by_valu_as_lanemask_then_wr_by_salu |=
other.sgpr_read_by_valu_as_lanemask_then_wr_by_salu;
+ sgpr_read_by_valu_as_lanemask_then_wr_by_valu |=
+ other.sgpr_read_by_valu_as_lanemask_then_wr_by_valu;
vgpr_written_by_wmma |= other.vgpr_written_by_wmma;
sgpr_read_by_valu |= other.sgpr_read_by_valu;
sgpr_read_by_valu_then_wr_by_valu |= other.sgpr_read_by_valu_then_wr_by_valu;
@@ -297,8 +303,11 @@ struct NOP_ctx_gfx11 {
valu_since_wr_by_trans == other.valu_since_wr_by_trans &&
trans_since_wr_by_trans == other.trans_since_wr_by_trans &&
sgpr_read_by_valu_as_lanemask == other.sgpr_read_by_valu_as_lanemask &&
+ sgpr_read_by_valu_as_lanemask2 == other.sgpr_read_by_valu_as_lanemask2 &&
sgpr_read_by_valu_as_lanemask_then_wr_by_salu ==
other.sgpr_read_by_valu_as_lanemask_then_wr_by_salu &&
+ sgpr_read_by_valu_as_lanemask_then_wr_by_valu ==
+ other.sgpr_read_by_valu_as_lanemask_then_wr_by_valu &&
vgpr_written_by_wmma == other.vgpr_written_by_wmma &&
sgpr_read_by_valu == other.sgpr_read_by_valu &&
sgpr_read_by_valu_then_wr_by_salu == other.sgpr_read_by_valu_then_wr_by_salu;
@@ -1377,6 +1386,30 @@ handle_valu_partial_forwarding_hazard(State& state, aco_ptr<Instruction>& instr)
return global_state.hazard_found;
}
+static bool
+instr_reads_lanemask(Instruction* instr, Operand* op)
+{
+ if (!instr->isVALU())
+ return false;
+ if (instr->isVOPD()) {
+ *op = Operand(vcc, s1);
+ return instr->opcode == aco_opcode::v_dual_cndmask_b32 ||
+ instr->vopd().opy == aco_opcode::v_dual_cndmask_b32;
+ }
+ switch (instr->opcode) {
+ case aco_opcode::v_addc_co_u32:
+ case aco_opcode::v_subb_co_u32:
+ case aco_opcode::v_subbrev_co_u32:
+ case aco_opcode::v_cndmask_b16:
+ case aco_opcode::v_cndmask_b32:
+ case aco_opcode::v_div_fmas_f32:
+ case aco_opcode::v_div_fmas_f64:
+ *op = instr->operands.back();
+ return !instr->operands.back().isConstant();
+ default: return false;
+ }
+}
+
void
handle_instruction_gfx11(State& state, NOP_ctx_gfx11& ctx, aco_ptr<Instruction>& instr,
std::vector<aco_ptr<Instruction>>& new_instructions)
@@ -1473,14 +1506,47 @@ handle_instruction_gfx11(State& state, NOP_ctx_gfx11& ctx, aco_ptr<Instruction>&
sa_sdst = 0;
}
+ /* VALU reading a SGPR as a lane mask and later written as a lane mask shouldn't be read again
+ * as a lane mask without a wait.
+ *
+ * TODO: this fixes #12623 and #11480, but needs further investigation as to why.
+ */
+ Operand lanemask_op;
+ if (instr_reads_lanemask(instr.get(), &lanemask_op)) {
+ unsigned reg = lanemask_op.physReg().reg();
+ if (ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu[reg] ||
+ (state.program->wave_size == 64 &&
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu[reg + 1])) {
+ bool is_vcc = reg == vcc || reg == vcc_hi;
+ bld.sopp(aco_opcode::s_waitcnt_depctr, is_vcc ? 0xfffd : 0xf1ff);
+ if (is_vcc)
+ wait.va_vcc = 0;
+ else
+ wait.va_sdst = 0;
+ }
+ }
+
if (va_vdst == 0) {
ctx.valu_since_wr_by_trans.reset();
ctx.trans_since_wr_by_trans.reset();
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu.reset();
}
if (sa_sdst == 0)
ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_salu.reset();
+ if (wait.va_sdst == 0) {
+ std::bitset<128> old = ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu;
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu.reset();
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu[vcc] = old[vcc];
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu[vcc_hi] = old[vcc_hi];
+ }
+
+ if (wait.va_vcc == 0) {
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu[vcc] = false;
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu[vcc_hi] = false;
+ }
+
if (state.program->wave_size == 64 && instr->isSALU() &&
check_written_regs(instr, ctx.sgpr_read_by_valu_as_lanemask)) {
unsigned reg = instr->definitions[0].physReg().reg();
@@ -1511,21 +1577,28 @@ handle_instruction_gfx11(State& state, NOP_ctx_gfx11& ctx, aco_ptr<Instruction>&
if (!op.isConstant() && op.physReg().reg() < 126)
ctx.sgpr_read_by_valu_as_lanemask.reset();
}
- switch (instr->opcode) {
- case aco_opcode::v_addc_co_u32:
- case aco_opcode::v_subb_co_u32:
- case aco_opcode::v_subbrev_co_u32:
- case aco_opcode::v_cndmask_b16:
- case aco_opcode::v_cndmask_b32:
- case aco_opcode::v_div_fmas_f32:
- case aco_opcode::v_div_fmas_f64:
- if (instr->operands.back().physReg() != exec) {
- ctx.sgpr_read_by_valu_as_lanemask.set(instr->operands.back().physReg().reg());
- ctx.sgpr_read_by_valu_as_lanemask.set(instr->operands.back().physReg().reg() + 1);
- }
- break;
- default: break;
+ }
+
+ if (instr_reads_lanemask(instr.get(), &lanemask_op)) {
+ unsigned reg = lanemask_op.physReg().reg();
+ if (state.program->wave_size == 64 && reg != exec) {
+ ctx.sgpr_read_by_valu_as_lanemask.set(reg);
+ ctx.sgpr_read_by_valu_as_lanemask.set(reg + 1);
}
+ ctx.sgpr_read_by_valu_as_lanemask2.set(reg);
+ if (state.program->wave_size == 64)
+ ctx.sgpr_read_by_valu_as_lanemask2.set(reg + 1);
+ }
+
+ if (instr->opcode != aco_opcode::v_readlane_b32_e64 &&
+ instr->opcode != aco_opcode::v_readfirstlane_b32 &&
+ !instr->definitions.empty() &&
+ instr->definitions.back().getTemp().type() == RegType::sgpr) {
+ unsigned reg = instr->definitions.back().physReg().reg();
+ if (ctx.sgpr_read_by_valu_as_lanemask2[reg])
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu[reg] = true;
+ if (state.program->wave_size == 64 && ctx.sgpr_read_by_valu_as_lanemask2[reg + 1])
+ ctx.sgpr_read_by_valu_as_lanemask_then_wr_by_valu[reg + 1] = true;
}
}
} else {
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: "Ivan A. Melnikov" <iv@altlinux.org>
Date: Fri, 7 Mar 2025 19:29:31 +0400
Subject: gallium/radeon: Make sure radeonsi PCI IDs are also included
When importing libdrm_radeon code [1][2] it was somehow missed
that what libdrm has in one r600_pci_ids.h, Mesa has split
into r600_pci_ids.h and radeonsi_pci_ids.h. So, devices
with ids from radeonsi_pci_ids.h were not considered valid for
radeon_surface_manager_new.
This commit changes that, thus fixing radeonsi for these
devices.
[1] commit 1299f5c50a490fadeb60b61677596f13399ee136
[2] commit 3aa7497cc0bb52c8099fb07b27f9aee5e18e58ca
Fixes: 1299f5c50a490fadeb60b61677596f13399ee136
Signed-off-by: Ivan A. Melnikov <iv@altlinux.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33940>
---
src/gallium/winsys/radeon/drm/radeon_surface.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/src/gallium/winsys/radeon/drm/radeon_surface.c b/src/gallium/winsys/radeon/drm/radeon_surface.c
index 8a3302df684..3c469ad0c6e 100644
--- a/src/gallium/winsys/radeon/drm/radeon_surface.c
+++ b/src/gallium/winsys/radeon/drm/radeon_surface.c
@@ -132,6 +132,9 @@ static int radeon_get_family(struct radeon_surface_manager *surf_man)
switch (surf_man->device_id) {
#define CHIPSET(pci_id, name, fam) case pci_id: surf_man->family = CHIP_##fam; break;
#include "pci_ids/r600_pci_ids.h"
+#undef CHIPSET
+#define CHIPSET(pci_id, fam) case pci_id: surf_man->family = CHIP_##fam; break;
+#include "pci_ids/radeonsi_pci_ids.h"
#undef CHIPSET
default:
return -EINVAL;
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Date: Tue, 11 Mar 2025 15:29:37 +0100
Subject: radv/amdgpu: fix device deduplication
To correctly deduplicate device inside the winsys, it should use the
fd or amdgpu_device_handle. Using the allocated ac_drm_device as key
is obviously broken.
Not deduplicating devices breaks memory budget and a bunch of games
were broken.
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12686
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12775
Fixes: a565f2994fe ("amd: move all uses of libdrm_amdgpu to ac_linux_drm")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34005>
---
src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
index be8df8708c8..8b57abeb0b1 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_winsys.c
@@ -234,7 +234,7 @@ radv_amdgpu_winsys_create(int fd, uint64_t debug_flags, uint64_t perftest_flags,
goto fail;
}
- struct hash_entry *entry = _mesa_hash_table_search(winsyses, dev);
+ struct hash_entry *entry = _mesa_hash_table_search(winsyses, (void *)ac_drm_device_get_cookie(dev));
if (entry) {
ws = (struct radv_amdgpu_winsys *)entry->data;
++ws->refcount;
@@ -325,7 +325,7 @@ radv_amdgpu_winsys_create(int fd, uint64_t debug_flags, uint64_t perftest_flags,
radv_amdgpu_bo_init_functions(ws);
radv_amdgpu_cs_init_functions(ws);
- _mesa_hash_table_insert(winsyses, dev, ws);
+ _mesa_hash_table_insert(winsyses, (void *)ac_drm_device_get_cookie(dev), ws);
simple_mtx_unlock(&winsys_creation_mutex);
return &ws->base;
--
2.48.1
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Maarten Lankhorst <maarten.lankhorst@intel.com>
Date: Mon, 17 Feb 2025 14:55:29 -0800
Subject: anv: Mark images with format modifiers set as scanout.
We currently use the presence of struct WSI_IMAGE_CREATE_INFO_MESA.scanout to mark the BO as scanout,
but this only handles the linear case, and fails when drm format modifiers are used.
Also handle the case of exportable BO with tiling set to VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT.
This fixes the gamescope handling of using vulkan allocated images for scanout.
Link: https://gitlab.freedesktop.org/mesa/mesa/-/issues/12633
Signed-off-by: Maarten Lankhorst <dev@lankhorst.se>
Signed-off-by: Matthew Schwartz <matthew.schwartz@linux.dev>
Normalspeak: fixes battlemage iGPUs in gamescope
---
src/intel/vulkan/anv_device.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
index 1884932bbc7..cbc1b4aad87 100644
--- a/src/intel/vulkan/anv_device.c
+++ b/src/intel/vulkan/anv_device.c
@@ -1533,6 +1533,9 @@ VkResult anv_AllocateMemory(
dedicated_info->image != VK_NULL_HANDLE) {
ANV_FROM_HANDLE(anv_image, image, dedicated_info->image);
+ if (image->vk.tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT)
+ alloc_flags |= ANV_BO_ALLOC_SCANOUT;
+
/* Apply implicit sync to be compatible with clients relying on
* implicit fencing. This matches the behavior in iris i915_batch
* submit. An example client is VA-API (iHD), so only dedicated
--
2.48.1
anda/lib/mesa/mesa.spec
+3 -21
View File
@@ -76,7 +76,7 @@ Summary: Mesa graphics libraries
# disabled by default, and has to be enabled manually. See `terra/release/terra-mesa.repo` for details.
Epoch: 1
Version: 25.0.1
Release: 5%?dist
Release: 6%?dist
License: MIT AND BSD-3-Clause AND SGI-B-2.0
URL: http://www.mesa3d.org
@@ -88,26 +88,8 @@ Source0: https://archive.mesa3d.org/%{srcname}-%{version}.tar.xz
# Fedora opts to ignore the optional part of clause 2 and treat that code as 2 clause BSD.
Source1: Mesa-MLAA-License-Clarification-Email.txt
#Patch10: https://src.fedoraproject.org/rpms/mesa/raw/e89544b7a4d811a64ca23b402add29524cc6f704/f/gnome-shell-glthread-disable.patch
#Patch11: https://src.fedoraproject.org/rpms/mesa/raw/e89544b7a4d811a64ca23b402add29524cc6f704/f/0001-llvmpipe-Init-eglQueryDmaBufModifiersEXT-num_modifie.patch
#Patch12: https://src.fedoraproject.org/rpms/mesa/raw/e89544b7a4d811a64ca23b402add29524cc6f704/f/0001-Revert-ac-radeonsi-remove-has_syncobj-has_fence_to_h.patch
# https://gitlab.com/evlaV/mesa/
Patch20: valve.patch
# Fix issues with Intel Battlemage under Valve's gamescope in DRM mode
# https://gitlab.freedesktop.org/mesa/mesa/-/issues/12633
Patch21: 12633.patch
# radv/amdgpu: fix device deduplication
# https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34005/diffs
Patch22: 34005.patch
Patch30: 237d8799be3afe9a1e7ca9156a5d44ffe0aae681.patch
Patch31: 13a3f9a972324a72dd507e09ac975b969e6c88e0.patch
# s390x: fix build
#Patch100: https://src.fedoraproject.org/rpms/mesa/raw/e89544b7a4d811a64ca23b402add29524cc6f704/f/fix-egl-on-s390x.patch
# https://github.com/bazzite-org/mesa
Patch20: bazzite.patch
BuildRequires: meson >= 1.3.0
BuildRequires: gcc
anda/lib/mesa/valve.patch
-143
View File
@@ -1,143 +0,0 @@
From 04afaf13b208f5c58c0b057f3dfc2dfa5c19a334 Mon Sep 17 00:00:00 2001
From: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Date: Fri, 14 Jan 2022 15:58:45 +0100
Subject: [PATCH 5/8] STEAMOS: radv: min image count override for FH5
Otherwise in combination with the vblank time reservation in
gamescope the game could get stuck in low power states.
---
src/util/00-radv-defaults.conf | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/src/util/00-radv-defaults.conf b/src/util/00-radv-defaults.conf
index 1cbb2e087c9..43488ada6bc 100644
--- a/src/util/00-radv-defaults.conf
+++ b/src/util/00-radv-defaults.conf
@@ -207,6 +207,11 @@ Application bugs worked around in this file:
<application name="Rocket League" executable="RocketLeague">
<option name="radv_zero_vram" value="true" />
</application>
+
+ <application name="Forza Horizon 5" application_name_match="ForzaHorizon5.exe">
+ <option name="vk_x11_override_min_image_count" value="4" />
+ </application>
+
<application name="Crystal Project" executable="Crystal Project.bin.x86_64">
<option name="radv_zero_vram" value="true" />
</application>
2.42.0
From b1c0d3de07bf958317f386585ce541b1c336e929 Mon Sep 17 00:00:00 2001
From: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Date: Mon, 21 Feb 2022 18:43:54 +0100
Subject: [PATCH 6/8] STEAMOS: Dynamic swapchain override for gamescope limiter
---
src/gallium/frontends/dri/loader_dri3_helper.c | 42 +++++++++++++++++++++++++++++++--
src/gallium/frontends/dri/loader_dri3_helper.h | 1 +
src/loader/meson.build | 2 +-
4 files changed, 80 insertions(+), 3 deletions(-)
diff --git a/src/gallium/frontends/dri/loader_dri3_helper.c b/src/gallium/frontends/dri/loader_dri3_helper.c
index 2631a9e2fd5..dbf6db349c6 100644
--- a/src/gallium/frontends/dri/loader_dri3_helper.c
+++ b/src/gallium/frontends/dri/loader_dri3_helper.c
@@ -276,6 +276,30 @@ dri3_update_max_num_back(struct loader_dri3_drawable *draw)
}
}
+static unsigned
+gamescope_swapchain_override()
+{
+ const char *path = getenv("GAMESCOPE_LIMITER_FILE");
+ if (!path)
+ return 0;
+
+ static simple_mtx_t mtx = SIMPLE_MTX_INITIALIZER;
+ static int fd = -1;
+
+ simple_mtx_lock(&mtx);
+ if (fd < 0) {
+ fd = open(path, O_RDONLY);
+ }
+ simple_mtx_unlock(&mtx);
+
+ if (fd < 0)
+ return 0;
+
+ uint32_t override_value = 0;
+ pread(fd, &override_value, sizeof(override_value), 0);
+ return override_value;
+}
+
void
loader_dri3_set_swap_interval(struct loader_dri3_drawable *draw, int interval)
{
@@ -290,10 +314,12 @@ loader_dri3_set_swap_interval(struct loader_dri3_drawable *draw, int interval)
* PS. changing from value A to B and A < B won't cause swap out of order but
* may still gets wrong target_msc value at the beginning.
*/
- if (draw->swap_interval != interval)
+ if (draw->orig_swap_interval != interval)
loader_dri3_swapbuffer_barrier(draw);
- draw->swap_interval = interval;
+ draw->orig_swap_interval = interval;
+ if (gamescope_swapchain_override() != 1)
+ draw->swap_interval = interval;
}
static void
@@ -422,6 +448,12 @@ loader_dri3_drawable_init(xcb_connection_t *conn,
draw->swap_interval = dri_get_initial_swap_interval(draw->dri_screen_render_gpu);
+ draw->orig_swap_interval = draw->swap_interval;
+
+ unsigned gamescope_override = gamescope_swapchain_override();
+ if (gamescope_override == 1)
+ draw->swap_interval = 1;
+
dri3_update_max_num_back(draw);
/* Create a new drawable */
@@ -1066,6 +1098,12 @@ loader_dri3_swap_buffers_msc(struct loader_dri3_drawable *draw,
if (draw->type == LOADER_DRI3_DRAWABLE_WINDOW) {
dri3_fence_reset(draw->conn, back);
+ unsigned gamescope_override = gamescope_swapchain_override();
+ if (gamescope_override == 1)
+ draw->swap_interval = 1;
+ else
+ draw->swap_interval = draw->orig_swap_interval;
+
/* Compute when we want the frame shown by taking the last known
* successful MSC and adding in a swap interval for each outstanding swap
* request. target_msc=divisor=remainder=0 means "Use glXSwapBuffers()
diff --git a/src/gallium/frontends/dri/loader_dri3_helper.h b/src/gallium/frontends/dri/loader_dri3_helper.h
index cc2362dd599..fe73b3f329c 100644
--- a/src/gallium/frontends/dri/loader_dri3_helper.h
+++ b/src/gallium/frontends/dri/loader_dri3_helper.h
@@ -170,6 +170,7 @@ struct loader_dri3_drawable {
bool block_on_depleted_buffers;
bool queries_buffer_age;
int swap_interval;
+ int orig_swap_interval;
const struct loader_dri3_vtable *vtable;
diff --git a/src/gallium/frontends/dri/meson.build b/src/gallium/frontends/dri/meson.build
index a98c8c0..0d4f816 100644
--- a/src/gallium/frontends/dri/meson.build
+++ b/src/gallium/frontends/dri/meson.build
@@ -23,7 +23,7 @@ if with_platform_x11
deps_for_libdri += dep_xcb
if with_dri_platform == 'drm'
deps_for_libdri += [dep_xcb_dri3, dep_xcb_present, dep_xcb_sync,
- dep_xshmfence, dep_xcb_xfixes]
+ dep_xshmfence, dep_xcb_xfixes, dep_xcb_xrandr, idep_mesautil]
files_libdri += files('loader_dri3_helper.c')
endif
endif