feat(vk): NVIDIA Reflex: each queue submit tagged with the game frame its GPU work belongs to #2499

github-sync · 2026-11-11T21:09:30Z

pmwork-personal commented

2026-11-11 21:09:30 +00:00

(Migrated from github.com)

Follows @nullifyshot's lead on #2118 (comment). Under DLSS frame generation, Reflex's own "to the GPU's end" read 243–442 ms there (RTX 5080, 4x), next to ~21 ms from simulation start to present end. Hookshot chained no VkLatencySubmissionPresentIdNV to its queue submits, so the driver had to guess which frame each submit's GPU work belongs to. Without frame generation it guesses right. With it, the display thread's copies and DLSS-G's work sit between the game's frames and get counted as the next game frame's rendering.

What changes

reflex_submit_id(next, frame) (vk_reflex.cpp) chains a VkLatencySubmissionPresentIdNV to a VkSubmitInfo2 (vk.xml: structextends="VkSubmitInfo,VkSubmitInfo2"). It is used at the four submit sites in vk_replay.cpp:

submit_recording's direct submit and the parallel segment jobs' submit: the frame being made.
A batch (queue_flush_locked): the frame of the batch's first command buffer, kept in pend.reflex when the batch starts.
present_image's copy: the frame it shows, by the same rule as the present id: the set's game frame (reflex_present_frame), no tag for a made frame, and the frame being made without frame generation.

There is no tag when Reflex is off, when no frame is marked, or when that present id was presented already (present ids only go up). Each tag is taken right before its vkQueueSubmit2, after any flush, and that flush tags its own batch. RX_VK_REFLEX_SUBMIT_IDS=0 leaves the tags off, for an A/B comparison.

In game: RTX 5090, 5120x2160 swapchain, DLAA and DLSS-G 3x at 4416x1864 (600 of 600 presents made by DLSS-G), V-Sync Auto, Reflex On + Boost, RX_VK_REFLEX_LOG=1, Kokiri Forest. Same build both runs; run 2 only adds RX_VK_REFLEX_SUBMIT_IDS=0.

	Simulation start → present end	To the GPU's end	Game fps	Reports
Tags on	40–51 ms	46–56 ms	~44–47	13
Tags off (`RX_VK_REFLEX_SUBMIT_IDS=0`)	44–59 ms	425–1060 ms	~42–44	5

With the tags, "to the GPU's end" lands a few ms after present end, the same pattern as without frame generation (12.1 vs 10.1 ms in nullifyshot's run), so the driver now paces on real GPU timings. The tags-off run also felt worse to play. Its simulation-to-present numbers are a bit higher too, but it was a short run (5 reports), so I wouldn't read much into that gap yet.

What this doesn't measure: input to screen. @nullifyshot, if you have time, your HOS_INPUT_LIVE method on this head would show what it's actually worth. The ~⅔ of a game frame that interpolation holds the real frame back stays as it was; Reflex can't remove that.

Tests: test_reflex_swapchain_lifetime.py copies reflex_submit_id and gains 8 checks (the frame being made, a kept frame, frame 0, no frame marked, one already presented, Reflex off, retired before the lock). draw_stream_visibility_test.cpp, fsr_image_retirement_test.cpp and parallel_init_order_test.cpp get stand-ins. tests/run_all.py (Windows, llvm-mingw clang++): 391 passed, 31 skipped, 1 failed: test_vram_budget_row, which reads only vk_menu.cpp, vk_menu.h and vk_memory.cpp (not touched here) and fails the same way on main.

🤖 Generated with Claude Code

Follows @nullifyshot's lead on #2118 ([comment](https://git.magicbean.dev/MagicBeanDev/Hookshot/pulls/2118)). Under DLSS frame generation, Reflex's own "to the GPU's end" read 243–442 ms there (RTX 5080, 4x), next to ~21 ms from simulation start to present end. Hookshot chained no `VkLatencySubmissionPresentIdNV` to its queue submits, so the driver had to guess which frame each submit's GPU work belongs to. Without frame generation it guesses right. With it, the display thread's copies and DLSS-G's work sit between the game's frames and get counted as the next game frame's rendering. **What changes** `reflex_submit_id(next, frame)` (vk_reflex.cpp) chains a `VkLatencySubmissionPresentIdNV` to a `VkSubmitInfo2` (vk.xml: `structextends="VkSubmitInfo,VkSubmitInfo2"`). It is used at the four submit sites in vk_replay.cpp: - **`submit_recording`'s direct submit and the parallel segment jobs' submit:** the frame being made. - **A batch (`queue_flush_locked`):** the frame of the batch's first command buffer, kept in `pend.reflex` when the batch starts. - **`present_image`'s copy:** the frame it shows, by the same rule as the present id: the set's game frame (`reflex_present_frame`), no tag for a made frame, and the frame being made without frame generation. There is no tag when Reflex is off, when no frame is marked, or when that present id was presented already (present ids only go up). Each tag is taken right before its `vkQueueSubmit2`, after any flush, and that flush tags its own batch. `RX_VK_REFLEX_SUBMIT_IDS=0` leaves the tags off, for an A/B comparison. **In game:** RTX 5090, 5120x2160 swapchain, DLAA and DLSS-G 3x at 4416x1864 (600 of 600 presents made by DLSS-G), V-Sync Auto, Reflex On + Boost, `RX_VK_REFLEX_LOG=1`, Kokiri Forest. Same build both runs; run 2 only adds `RX_VK_REFLEX_SUBMIT_IDS=0`. | | Simulation start → present end | To the GPU's end | Game fps | Reports | |---|---|---|---|---| | Tags on | 40–51 ms | **46–56 ms** | ~44–47 | 13 | | Tags off (`RX_VK_REFLEX_SUBMIT_IDS=0`) | 44–59 ms | **425–1060 ms** | ~42–44 | 5 | With the tags, "to the GPU's end" lands a few ms after present end, the same pattern as without frame generation (12.1 vs 10.1 ms in nullifyshot's run), so the driver now paces on real GPU timings. The tags-off run also felt worse to play. Its simulation-to-present numbers are a bit higher too, but it was a short run (5 reports), so I wouldn't read much into that gap yet. What this doesn't measure: input to screen. @nullifyshot, if you have time, your `HOS_INPUT_LIVE` method on this head would show what it's actually worth. The ~⅔ of a game frame that interpolation holds the real frame back stays as it was; Reflex can't remove that. **Tests:** `test_reflex_swapchain_lifetime.py` copies `reflex_submit_id` and gains 8 checks (the frame being made, a kept frame, frame 0, no frame marked, one already presented, Reflex off, retired before the lock). `draw_stream_visibility_test.cpp`, `fsr_image_retirement_test.cpp` and `parallel_init_order_test.cpp` get stand-ins. `tests/run_all.py` (Windows, llvm-mingw clang++): 391 passed, 31 skipped, 1 failed: `test_vram_budget_row`, which reads only `vk_menu.cpp`, `vk_menu.h` and `vk_memory.cpp` (not touched here) and fails the same way on main. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

Joey-Magnus (Migrated from github.com) requested changes

2026-11-11 21:46:32 +00:00

Joey-Magnus (Migrated from github.com) left a comment

The tagging looks right. Each tag is taken under dev.present_mu right before its vkQueueSubmit2, after any flush (which tags its own batch), so nothing can overwrite the thread_local struct in between. Taking dev.present_mu before r.mu is the order reflex_submit_ends() already uses. Made frames and ids already presented get no tag, and the A/B numbers are convincing. On other GPUs the cost is one uncontended lock per submit, which is fine.

One rule issue: gpu/vk/vk_reflex.cpp:245 reads RX_VK_REFLEX_SUBMIT_IDS with atoi(...) == 0. Since #2042 the convention is env_value_off/env_on from compat/hookshot_env.h. With atoi, =true or =on turns the tags off. const char* e = getenv(...); off = e && env_value_off(e); would fix it. Please also add the switch to the Reflex row in docs/development.md:170.

After that it's ready from my side. A batch that spans two game frames is credited to the first frame's id. That seems an acceptable approximation, but the comment should say so.

The tagging looks right. Each tag is taken under `dev.present_mu` right before its `vkQueueSubmit2`, after any flush (which tags its own batch), so nothing can overwrite the thread_local struct in between. Taking `dev.present_mu` before `r.mu` is the order `reflex_submit_ends()` already uses. Made frames and ids already presented get no tag, and the A/B numbers are convincing. On other GPUs the cost is one uncontended lock per submit, which is fine. One rule issue: `gpu/vk/vk_reflex.cpp:245` reads `RX_VK_REFLEX_SUBMIT_IDS` with `atoi(...) == 0`. Since #2042 the convention is `env_value_off`/`env_on` from `compat/hookshot_env.h`. With `atoi`, `=true` or `=on` turns the tags off. `const char* e = getenv(...); off = e && env_value_off(e);` would fix it. Please also add the switch to the Reflex row in `docs/development.md:170`. After that it's ready from my side. A batch that spans two game frames is credited to the first frame's id. That seems an acceptable approximation, but the comment should say so.

nullifyshot commented

2026-11-11 21:57:05 +00:00

(Migrated from github.com)

Ran the input-to-screen test on ca483d88 with the same PC, method and spot as on #2118: RTX 5080, 360 Hz G-SYNC, DLAA, DLSS-G 4x, Reflex On + Boost, RX_VK_REFLEX_LOG=1. The game presses its own buttons through HOS_INPUT_LIVE: 20 camera flicks, 20 sword swings and 20 steps per run. Each is timed from the press being due to the first desktop frame (Desktop Duplication) that shows the reaction. All three runs are the same build. Run 2 only adds RX_VK_REFLEX_SUBMIT_IDS=0, and run 3 sets RX_VK_DLSS_FG=0 and RX_VK_FRAMEGEN=0. All 180 presses were measured, and DLSS-G made 600 of 600 presents in every window of runs 1 and 2.

Action (median of 20)	DLSS-G 4x, tags on	DLSS-G 4x, tags off	Frame generation off
Camera flick	38.8 ms	39.1 ms	28.5 ms
Sword (Y)	69.3 ms	75.0 ms	49.1 ms
Walk	73.0 ms	79.9 ms	50.1 ms

Reflex's own timing over the runs:

	Simulation start → present end	To the GPU's end	Game fps
DLSS-G, tags on	18.3–19.8 ms	38–40 ms	~68
DLSS-G, tags off	21.3–21.6 ms	147–969 ms	~70
Frame generation off	11.4–11.7 ms	24.3–24.9 ms	~87

The tags fix the GPU-end timing. It's steady at about 20 ms past present end, the same kind of gap as with frame generation off (about 13 ms). Without the tags it jumps around by hundreds of ms, as before.
The tags look worth about 6–7 ms on the sword and walk, and nothing on the camera flick. The 10–90 % ranges of the two runs overlap, so I'd call it a likely gain, not a proven one. Simulation start → present end also drops about 2–3 ms with the tags, which points the same way.
The cost of DLSS-G vs off stays large: +10 ms on the camera and +20–23 ms on the sword and walk with the tags. As you said, the held-back real frame is still there.

One thing that isn't this PR: compared with my #2118 run on 3a39e3f, this build's game frame rate is lower in roughly the same spot (about 68 vs 77 fps with DLSS-G, and 87 vs 97 with it off). The sword and walk are also slower in every mode. Main moved about 100 commits between the two, so that's somewhere in main, not the tags.

Ran the input-to-screen test on ca483d88 with the same PC, method and spot as on #2118: RTX 5080, 360 Hz G-SYNC, DLAA, DLSS-G 4x, Reflex On + Boost, `RX_VK_REFLEX_LOG=1`. The game presses its own buttons through `HOS_INPUT_LIVE`: 20 camera flicks, 20 sword swings and 20 steps per run. Each is timed from the press being due to the first desktop frame (Desktop Duplication) that shows the reaction. All three runs are the same build. Run 2 only adds `RX_VK_REFLEX_SUBMIT_IDS=0`, and run 3 sets `RX_VK_DLSS_FG=0` and `RX_VK_FRAMEGEN=0`. All 180 presses were measured, and DLSS-G made 600 of 600 presents in every window of runs 1 and 2. | Action (median of 20) | DLSS-G 4x, tags on | DLSS-G 4x, tags off | Frame generation off | |---|---|---|---| | Camera flick | 38.8 ms | 39.1 ms | 28.5 ms | | Sword (Y) | **69.3 ms** | 75.0 ms | 49.1 ms | | Walk | **73.0 ms** | 79.9 ms | 50.1 ms | Reflex's own timing over the runs: | | Simulation start → present end | To the GPU's end | Game fps | |---|---|---|---| | DLSS-G, tags on | 18.3–19.8 ms | **38–40 ms** | ~68 | | DLSS-G, tags off | 21.3–21.6 ms | **147–969 ms** | ~70 | | Frame generation off | 11.4–11.7 ms | 24.3–24.9 ms | ~87 | - **The tags fix the GPU-end timing.** It's steady at about 20 ms past present end, the same kind of gap as with frame generation off (about 13 ms). Without the tags it jumps around by hundreds of ms, as before. - **The tags look worth about 6–7 ms on the sword and walk**, and nothing on the camera flick. The 10–90 % ranges of the two runs overlap, so I'd call it a likely gain, not a proven one. Simulation start → present end also drops about 2–3 ms with the tags, which points the same way. - **The cost of DLSS-G vs off stays large:** +10 ms on the camera and +20–23 ms on the sword and walk with the tags. As you said, the held-back real frame is still there. One thing that isn't this PR: compared with my #2118 run on 3a39e3f, this build's game frame rate is lower in roughly the same spot (about 68 vs 77 fps with DLSS-G, and 87 vs 97 with it off). The sword and walk are also slower in every mode. Main moved about 100 commits between the two, so that's somewhere in main, not the tags.

hckrman101 (Migrated from github.com) reviewed

2026-11-11 22:20:54 +00:00

hckrman101 (Migrated from github.com) left a comment

Nice, and the numbers back it up: your 5090 run and @nullifyshot's input-to-screen run both show the GPU-end timing going from hundreds of ms to a few ms past present end, with a likely 6–7 ms gain on sword/walk.

I checked the four sites. Each tag is taken under dev.present_mu right before its vkQueueSubmit2, so the thread_local struct can't be overwritten in between. present_image flushes the batch (tagged with its own frame) before tagging its copy. The lock order (present_mu then r.mu) matches the existing callers. Made frames, unmarked frames and ids already presented get no tag. VkLatencySubmissionPresentIdNV is allowed on VkSubmitInfo2, and on() covers the extension being enabled. The test stand-ins match the new signature. CI is green.

I agree with @Joey-Magnus's review and won't repeat it. It's still open at this head: RX_VK_REFLEX_SUBMIT_IDS should use env_value_off from gpu/vk/compat/hookshot_env.h rather than atoi(...) == 0, the switch needs adding to the Reflex row in docs/development.md, and the comment should say a batch is credited to its first frame. Once those are in, this is good to go from my side.

One small addition to that last point, inline.

Nice, and the numbers back it up: your 5090 run and @nullifyshot's input-to-screen run both show the GPU-end timing going from hundreds of ms to a few ms past present end, with a likely 6–7 ms gain on sword/walk. I checked the four sites. Each tag is taken under `dev.present_mu` right before its `vkQueueSubmit2`, so the thread_local struct can't be overwritten in between. `present_image` flushes the batch (tagged with its own frame) before tagging its copy. The lock order (`present_mu` then `r.mu`) matches the existing callers. Made frames, unmarked frames and ids already presented get no tag. `VkLatencySubmissionPresentIdNV` is allowed on `VkSubmitInfo2`, and `on()` covers the extension being enabled. The test stand-ins match the new signature. CI is green. I agree with @Joey-Magnus's review and won't repeat it. It's still open at this head: `RX_VK_REFLEX_SUBMIT_IDS` should use `env_value_off` from `gpu/vk/compat/hookshot_env.h` rather than `atoi(...) == 0`, the switch needs adding to the Reflex row in `docs/development.md`, and the comment should say a batch is credited to its first frame. Once those are in, this is good to go from my side. One small addition to that last point, inline.

gpu/vk/vk_replay.cpp

					
				@ -419,3 +423,4 @@

				    si.pNext = reflex_submit_id(nullptr);   // (NVIDIA Reflex: the frame being made)

				    const VkResult res = vkQueueSubmit2(dev.queue, 1, &si, r.fence);

				    r.submitted = res == VK_SUCCESS;

				    static int failures;

Nit, on top of Joey's note: the batch's frame is read only when it starts. So a batch started before the frame's sleep returned (reflex_frame() = 0) stays untagged, even when the rest of its command buffers belong to the new frame. That's harmless, since batching is off by default outside RADV, so on NVIDIA this only matters with RX_VK_BATCH=1. It's worth a word in the same comment, though.

Nit, on top of Joey's note: the batch's frame is read only when it starts. So a batch started before the frame's sleep returned (`reflex_frame()` = 0) stays untagged, even when the rest of its command buffers belong to the new frame. That's harmless, since batching is off by default outside RADV, so on NVIDIA this only matters with `RX_VK_BATCH=1`. It's worth a word in the same comment, though.

hckrman101 commented

2026-11-11 22:55:54 +00:00

(Migrated from github.com)

This now conflicts with main in gpu/vk/vk_reflex.cpp. Could you merge main into the branch (or rebase) when you get a chance? I'll take another look once it's up to date.

This now conflicts with main in `gpu/vk/vk_reflex.cpp`. Could you merge main into the branch (or rebase) when you get a chance? I'll take another look once it's up to date.

pmwork-personal commented

2026-11-11 23:07:02 +00:00

(Migrated from github.com)

Thanks both, and thanks @nullifyshot for the input-to-screen run.

Updated: main merged first (8678b655). #2325's reflex_present_id_features now sits next to reflex_submit_id. With Reflex on, its presenter thread isn't used, so the present's tag stays on the game's thread, where reflex_present_frame is set. Then the review notes (fa75d051):

RX_VK_REFLEX_SUBMIT_IDS is read with env_value_off from hookshot_env.h. test_reflex_swapchain_lifetime.py now runs its fixture with =on (tags on; the old atoi reading fails it at check 21) and with =0 (untagged).
The Reflex row in docs/development.md lists the switch.
The header of vk_reflex.cpp and the batch start in vk_replay.cpp say how a batch is credited: to the frame of its first command buffer, read when the batch starts. So one that spans two game frames counts as the first's, and one begun before the frame's sleep returned goes untagged. Batching is on by default only with RADV, which has no Reflex.

tests/run_all.py (Windows, llvm-mingw): 432 passed, 35 skipped, 3 failed, none of them in this PR: test_env_on (main's #2470 reads HOS_FPS_LOG as set-means-on in host/oot_high_fps.c:15), test_vram_budget_row (fails the same on main here) and test_portable_builder_startup (its dotnet publish needs packages this PC doesn't have).

On the frame-rate drop between the two builds (about 68 vs 77 fps with DLSS-G, 87 vs 97 without): agreed, that's somewhere in main, not here. It's next on my list for the frame-rate work.

Thanks both, and thanks @nullifyshot for the input-to-screen run. Updated: main merged first (8678b655). #2325's `reflex_present_id_features` now sits next to `reflex_submit_id`. With Reflex on, its presenter thread isn't used, so the present's tag stays on the game's thread, where `reflex_present_frame` is set. Then the review notes (fa75d051): - `RX_VK_REFLEX_SUBMIT_IDS` is read with `env_value_off` from `hookshot_env.h`. `test_reflex_swapchain_lifetime.py` now runs its fixture with `=on` (tags on; the old `atoi` reading fails it at check 21) and with `=0` (untagged). - The Reflex row in `docs/development.md` lists the switch. - The header of `vk_reflex.cpp` and the batch start in `vk_replay.cpp` say how a batch is credited: to the frame of its first command buffer, read when the batch starts. So one that spans two game frames counts as the first's, and one begun before the frame's sleep returned goes untagged. Batching is on by default only with RADV, which has no Reflex. `tests/run_all.py` (Windows, llvm-mingw): 432 passed, 35 skipped, 3 failed, none of them in this PR: `test_env_on` (main's #2470 reads `HOS_FPS_LOG` as set-means-on in `host/oot_high_fps.c:15`), `test_vram_budget_row` (fails the same on main here) and `test_portable_builder_startup` (its `dotnet publish` needs packages this PC doesn't have). On the frame-rate drop between the two builds (about 68 vs 77 fps with DLSS-G, 87 vs 97 without): agreed, that's somewhere in main, not here. It's next on my list for the frame-rate work.

hckrman101 (Migrated from github.com) approved these changes

2026-11-11 23:12:14 +00:00

hckrman101 (Migrated from github.com) left a comment

Thanks, all three of @Joey-Magnus's asks are in, plus my nit:

RX_VK_REFLEX_SUBMIT_IDS is read with env_value_off from hookshot_env.h. The test now runs with =on (which the old atoi reading would have turned off) and =0.
The switch is in the Reflex row of docs/development.md.
The header of vk_reflex.cpp and the batch start in vk_replay.cpp both say a batch is credited to its first command buffer's frame, and that one started before the frame is marked stays untagged.

The merge is clean. The PR is 0 behind main. Against main, vk_reflex.cpp has only additions: the header comment, the include and reflex_submit_id, sitting next to #2325's reflex_present_id_features. None of main's lines were dropped. vk_replay.cpp differs from main only at the four tag sites and pend.reflex. With #2325's presenter thread, the copy's submit and its tag stay in present_image on the game's thread. A deferred present carries reflex_present_frame with it, so the tag still matches the frame that's shown.

CI was still pending at this head when I looked. The test_env_on failure you mention comes from main's #2470, not from this PR. Good to merge once CI is green.

Thanks, all three of @Joey-Magnus's asks are in, plus my nit: - `RX_VK_REFLEX_SUBMIT_IDS` is read with `env_value_off` from `hookshot_env.h`. The test now runs with `=on` (which the old `atoi` reading would have turned off) and `=0`. - The switch is in the Reflex row of `docs/development.md`. - The header of `vk_reflex.cpp` and the batch start in `vk_replay.cpp` both say a batch is credited to its first command buffer's frame, and that one started before the frame is marked stays untagged. The merge is clean. The PR is 0 behind main. Against main, `vk_reflex.cpp` has only additions: the header comment, the include and `reflex_submit_id`, sitting next to #2325's `reflex_present_id_features`. None of main's lines were dropped. `vk_replay.cpp` differs from main only at the four tag sites and `pend.reflex`. With #2325's presenter thread, the copy's submit and its tag stay in `present_image` on the game's thread. A deferred present carries `reflex_present_frame` with it, so the tag still matches the frame that's shown. CI was still pending at this head when I looked. The `test_env_on` failure you mention comes from main's #2470, not from this PR. Good to merge once CI is green.

hckrman101 (Migrated from github.com) reviewed

2026-11-11 23:12:21 +00:00

gpu/vk/vk_replay.cpp

					
				@ -419,3 +423,4 @@

				    si.pNext = reflex_submit_id(nullptr);   // (NVIDIA Reflex: the frame being made)

				    const VkResult res = vkQueueSubmit2(dev.queue, 1, &si, r.fence);

				    r.submitted = res == VK_SUCCESS;

				    static int failures;