Plan 25.1c: vk_mac.ll's @lvk_ac (posix_memalign under a 16-byte header of scope/offset/size, atomic
counters) passed at every render3d create/destroy (49 sites; the caps probe keeps its own null pair);
lvk_ac_bytes/_peak/_allocs/_scope_bytes for the fence, Vk.alloc_bytes. vk_win.ll: null and 0.
MoltenVK 1.4.2 counted 0 live bytes through them in steady. Fence findings: gvk_pipeline freed its 17
create infos (and reuses one bufs list); actor_init fills ac_spare with 512 records (actor_fresh,
m4_new, v3_new at first placement in play). Compiled, not run (the user's call).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A co-op guest loading its models segfaulted in copyBufferToBuffer under vkQueueSubmit (0x68, AGX
LegacyBlitContext): a buffer primed into the frame's command buffer was released later in that
frame, and gvk_buf_release destroys a buffer no draw has marked, so the copy read a freed one. The
copies now run in their own command buffer before the frame's begins, checked against the slot's
handle as they are recorded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
lp_os_platform and lp_os_arch malloc'd 8 KB per call (the uname buffer) and kept none of it:
string_temps now asks both every round, 327 MB over 20,000 before, 0 after. Reseeded. render3d:
MoltenVK made a mapped buffer's MTLBuffer at its first bind (fn_gvk_draw +4 blocks in the boat
window); gvk_buf_reserve queues it and the next frame's command buffer copies 4 bytes out of it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gvk_tex_storage freed none of ici, out, req and vci (8 blocks a two-texture model); tex_load_ex kept
dds_path_of's string. Found with malloc_history over 400 load/release rounds (712 bytes a round, all
of it these). steady: the model's bound is 4 KB over 200 (was 1.6 MB), and the frame is the least of
three settled windows - a valley self-test beside it read 87 KB once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A freed texture or buffer id goes on a spare list the next one takes; tex_note_size keeps sizes by id.
model_release frees prims, meshes, material names, the skin and leaves gltf_cached. actor_release
takes an actor off the stage and actor_new reuses its record (ECS's Things come and go all day).
steady.ludic: an actor round at 0 bytes over 2000.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Metal aborted a net.sh guest ("MTLPixelFormatBC7_RGBAUnorm is not color renderable"): the crash
report's main thread was in vkQueueSubmit from gvk_once_end, MoltenVK encoding a vkCmdBlitImage
through a render pipeline. water.ludic bound its reflection by sampler name and then asked the
BOUND texture for mips - on Vulkan the bound texture was not the reflection but the last one
bound, since 23.3 a BC7 kit texture. gpu_bind_sampler now makes its texture the bound one, as
OpenGL did (and water binds it explicitly); gvk_tex_mips and gvk_mips_now skip compressed images;
gvk_pass_begin leaves out a compressed colour attachment and says so.
tests/net.sh passes (main e515bd87 built against it; host and guest agree on 196 animals, the
people, the board and the lost hiker).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The device's textureCompressionBC is asked for and remembered (gvk_has_bc). tex_load_ex prefers a
DX10 .dds with the full chain beside the .png (not for an edge-padded cut-out atlas):
texture_dds.ludic reads BC7 / BC5 / BC4, gpu_tex_compressed makes the image with every level and
no colour-attachment use (a compressed image is only sampled and copied into), and
gvk_tex_upload_blocks copies each level's blocks from one staging buffer. A colour map is BC7
sampled as sRGB, a data map BC7 read as it is. examples/rendering/bc.ludic holds it, in the suite;
ludic.lab's plate carries its .dds (its three shots render at 56-60 dB against the .png's).
ludic-dev test 307/307, no Vulkan SDK in the environment.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- gvk_sampler built a key string on every call it was reached (a texture read two ways in turn
misses its per-texture cache each time); samplers are found by their seven numbers instead
- gvk_tex_mips_into allocated a VkImageBlit per level, every frame (the exposure measure's
chain); it comes from the scratch ring, as does gvk_tex_grow_mips's copy
- gvk_program_new made a new program - modules, pipelines - on every call, and every actor asks
for one; a variant is now made once and shared (a program is immutable), and one that cannot
be made stays 0 for every later asker
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- model_release(model) frees each primitive's mesh, and each texture once no other model uses it:
the glTF texture cache counts the primitives using each texture (a path loaded again, and a LOD
chain borrowing its LOD0's material, each take a reference); the last one frees the texture and
forgets it along with any remembered material naming it, so a later load is a fresh one
- a scatter layer's instance and sort arrays start at 256 and double as layer_add or a stream
fills them, to the layer's cap, instead of the whole cap up front (0.4 GB in Maroon Lake)
- gvk_tex_read's copy struct comes from the scratch ring
- examples/rendering/release.ludic holds it (RELEASE OK on Vulkan and OpenGL), in the suite
smooth renders pixel-identical before and after (max |d| 0). ludic-dev test 306/306.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- every texture upload malloc'd a CPU copy of its pixels and never freed it (236 MB over the
valley's boot, and again for every model loaded later): the pixels are converted straight
into the mapped staging buffer
- the per-level and per-layer views gvk_view_of made outlived their texture, and on MoltenVK a
view keeps its Metal texture alive: a released texture takes its views with it, and the cache
is keyed by numbers instead of a string built on every call
- the Vulkan structs filled for a draw, pass, barrier, descriptor set, buffer, allocation or
upload (about sixty call sites) come from a reused 1 MB scratch ring (gvk_tmp)
- the descriptor-set cache and the retired buffers are emptied in place, not replaced; the grass
cull's dispatch arguments are made once
- macOS drains an autorelease pool each frame (Vk.frame_pool, lvk_frame_pool in vk_mac.ll)
- R3D_VK_PROF reports Vulkan objects made and destroyed by kind, and every cache's length
- examples/rendering/smooth presents through render3d, so it runs on Vulkan too
The full valley on headless Vulkan loads to 2.18 GB and holds (it passed 8 GB while loading
before). ludic-dev test 305/305, selfhost-test 33/33.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- a CAMetalLayer on the view (cocoa.ll win_metal_layer), VK_EXT_metal_surface, QuartzCore linked
with Vk.*; the drawable measured after the layer sets the backing scale
- vk_mac.ll opens MoltenVK directly after any loader: a bundle ships only libMoltenVK.dylib
- a covered window is not presented to (win_visible); one frame in flight on macOS
(R3D_VK_INFLIGHT), with images, buffers and descriptor pools held until it is done
- win_held / win_mouse / win_pad / win_touch / win_text / win_present pass slice elements (arg_buf):
every windowed program died on its first input poll
- variants.list and SPIR-V for the three NEAR_FADE foliage programs
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ludic migrate state packages <every example program> packages/ludic.lab/example/plate.ludic
1804 vars into 126 states, 64 into lets; 23498 edits in 460 files
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A numbers float file adapts decimal literals to a fixed operand or slot, and refuses to
promote a computed int to a float implicitly: there it is almost always float bits. Explicit
float(x) is always allowed.
render3d's numbers are float, converted by tools/migrate/floatbits.py - a whole-program
inference of which ints carried IEEE bits (union-find over flows, calls, returns, buffers,
nested buffers and lexical scopes) and a rewriter to operators, Math.* and float literals,
with float_bits / float_from_bits left only where bits really cross (runtime scratch
buffers, mixed buffers). Seed regenerated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Input.text, App.monitor_count / window_to_monitor / window_fixed,
gl_sleep_us, and the grass and collide changes, uncommitted on main and
depended on by the game; carried here so the language work starts from
what the game actually uses. main's working tree is untouched.
HDR output: an HDR10 swapchain (A2B10G10R10, ST 2084 over BT.2020) when the
setting asks and the display offers it, with HDR metadata. The tonemap's HDR10
variant keeps the SDR picture up to a 200-nit paper white and rolls highlights
on to 1000 nits; the overlay's converts the interface to the same white. The
screen and LDR images go 10-bit with it; screenshots refuse while it is on.
OpenGL and the Vulkan SDR frame are unchanged. The instance asks for
VK_EXT_swapchain_colorspace. HDR metadata only where the loader has
vkSetHdrMetadataEXT: Streamline's interposer does not, and calling the thunk
crashed the game the moment the swapchain came up HDR10. PC 4K monitor: HDR10,
validation 0. R3D_HDR overrides the setting.
DLSS:
- the vertical jitter offset flips with Streamline's image (rows from the top):
unflipped, Quality resolved the ground into concentric rings;
- preset K in every mode: the default M put Performance at 18 ms a frame at 4K
on an RTX 3070 Ti (33 fps against 41 with DLSS off; with K, 60);
- the camera is jittered only while this frame holds a token and the last
evaluate worked.
R3D_DLSS_PRESET, R3D_CAM_LOG (the camera and DLSS state a frame) and
R3D_NOGRAIN for measuring.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
streamline.ludic: slInit before the Vulkan instance, feature support per
adapter, a frame token per frame, Reflex sleep and PCL latency markers, and
DLSS super resolution on the lit HDR frame (Halton jitter, depth + zero
motion vectors with camera motion from clipToPrevClip, matrices carrying
the Vulkan path's y flip and depth remap). Bloom, tonemap and sharpen read
the upscaled size. The device asks for privateData and present_id, which
Streamline's hooks need. R3D_DLSS / R3D_REFLEX / R3D_SL_LOG for tests.
gpu_feature_implemented: DLSS and Reflex.
ludic bundle (Windows): app native "<dir>" copies native libraries beside
the executable.
Verified on an RTX 3070 Ti: DLSS Quality evaluates 1280x720 -> 1920x1080.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
gpu_rb_storage makes the image with the samples asked for; a pass takes its attachments' count and
every pipeline in it matches; a blit from a multisampled source resolves on the end of an empty
dynamic-rendering pass (colour averaged, depth from sample zero - vkCmdResolveImage cannot do depth).
gpu_msaa_max reads the device's colour-and-depth sample limits, and post_set_msaa clamps to it, so
the Anti-aliasing setting is live on Vulkan. The pipeline cache's pass key no longer packs samples
into three bits.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- A block sub-allocator: 64 MB blocks per memory type, images and buffers kept apart, first fit with
alignment, freed ranges merged and empty blocks given back; anything over 16 MB still gets its own
allocation. The self-tests' second world holds 94 allocations instead of 8735 (the driver's limit
refused a shadow map before). Camp bench unchanged, 105.3 fps.
- shadow_set_res keeps the size that worked when the card has no memory for the new one, and records
it in shadow_refused, instead of ending with no shadow map; gpu_tex_ok says whether a texture has an
image behind it.
- Image barriers skip an image that was never made (a failed allocation used to crash there), and
R3D_VK_ERRLOG=<file> appends every Vulkan failure line by line, so a crash no longer takes the
message with it.
- R3D_VK_PROF reports draws asked for and not made, so a layer missing from a frame is never silent.
Validation on (VK_INSTANCE_LAYERS): the game's self-tests 61 OK, 0 errors, on the RTX 3070 Ti.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- A draw's set carries its textures and lives in a pool of its own, keyed by each texture's handle,
generation and sampler; the uniform blocks are dynamic uniform buffers into the frame's ring,
bound with the draw's offsets, so a draw that changes only uniforms allocates and writes no set.
The sampler for an unbound slot is made once instead of looked up by string every draw.
Camp bench, RTX 3070 Ti, 400 frames, twice: 102.6 fps (53.3 before; OpenGL 114.3), sets 0.9 ms
against 8.5. MoltenVK (M4 Pro): sets 0.2 ms.
- Integer vertex attributes read by float inputs use USCALED formats (a_joints was UINT against a
vec4), and every shader input a mesh does not feed reads a shared zero buffer. NVIDIA drew
anyway; MoltenVK refused every skinned pipeline and the whole actor layer was missing from the
Mac's Vulkan frame. A draw skipped for want of a pipeline now says so, once per program.
- A failed image allocation is reported instead of bound as a null allocation.
Validation proven on (VK_INSTANCE_LAYERS): 0 errors over the game's self-tests (61 OK) and a frame.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- r3d_open opens the window with nothing baked; r3d_load_count / r3d_load_step run the rest (sky
and light, terrain, shadow and screen targets, cover and actors, grass) so a game can present a
loading frame between them. r3d_init is the same calls in a row.
- sky_set_quality(width): the prefiltered sky light at 256 / 512 / 1024, baked again.
- post_set_msaa(samples): multisampling on OpenGL, remade in place (post_msaa_live is false on
Vulkan); post_free frees the multisampled framebuffer too.
- STREAM_BUDGET_US is a variable; r3d_fog_scale multiplies the fog the day sets.
- Vulkan buffers carry UNIFORM_BUFFER usage: the frame's ring is bound as uniform buffers and was
created without it (VUID-VkWriteDescriptorSet-descriptorType-00330, found on MoltenVK).
OpenGL frames at the five viewpoints unchanged; the game's 59 self-tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- Device: multiDrawIndirect, drawIndirectFirstInstance and drawIndirectCount where present.
- Buffers carry storage and indirect usage; a GPU-owned buffer is never swapped under a draw.
- Compute programs from shaders/compute.list (binding 0 parameters, 1.. storage buffers),
built by `ludic-dev shaders`; gpu_compute / gpu_dispatch / gpu_draw_mesh_indirect in gpu.ludic.
- R3D_VK_PROBE=1: a dispatch read back (OK on the RTX 3070 Ti).
- scatter_cull.comp: a tree layer's frustum test and LOD split on the GPU, with the lit, prepass,
impostor and shadow-LOD draws reading its records. Behind R3D_GPU_CULL=1 and off by default:
at the camp it is slower (43.0 fps against 53.3), because the frame's cost is per-draw
descriptor sets and it adds empty-level draws. Validation-clean; OpenGL frames unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
R3D_VK_PROF at the camp view (about 2950 draws a frame) showed the frame spent in bookkeeping:
19.5 ms a frame finding pipelines by string key and 16 ms building descriptor sets.
- Pipelines: a mesh carries an interned vertex-layout id (dropped only when an attribute's shape
changes, not when an instance buffer is swapped), render state packs into an int and the pass
formats into another; the program's last hit is tried first. The string path only builds.
- Samplers: each texture keeps the sampler for its parameters until they change.
- Descriptor sets: a program's last set is reused within the frame while its blocks and resolved
textures are unchanged; a uniform write that repeats the value it already holds changes nothing.
Headless on the RTX 3070 Ti at 1920x1080: 21.7 -> 53.3 fps (pipelines 0.2 ms, sets 8.5 ms, inside
draws 10 ms a frame; OpenGL 114 fps). The camp frame is unchanged and validation-clean. OpenGL frames
byte-identical at the five viewpoints; 59 self-tests pass; VKRES OK.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- A buffer re-uploaded after a draw this frame read it gets fresh storage, as OpenGL orphans
one; storage moved off or freed while the frame still reads it is destroyed after the frame's
submit. Draws mark the buffers they bind. No flush for buffer work.
- Mipmaps asked for mid-frame (the exposure measure, every frame) are recorded into the open frame
after its pass; growing a chain the first time keeps its one-shot path.
- R3D_VK_PROF prints, every 120 frames, draws and flushes per frame and the milliseconds spent
finding pipelines, filling descriptor sets and inside draws.
The gain was small - the camp view headless at 1920x1080 on the RTX 3070 Ti went from 21.3 to
21.7 fps (OpenGL: 114 fps) - so these flushes were not what holds the frame; the profile is how
the rest is found. The frame is unchanged and validation-clean; OpenGL frames byte-identical at the
five viewpoints with 59 self-tests passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three things the first Vulkan frames on the RTX 3070 Ti showed against OpenGL on the same PC:
- Exposure: the adaptation pass reads the HDR scene's smallest mip, and a render target made
without pixels had one level, so exposure came from a single texel. A target asked for mipmaps
now grows a full chain (level 0 kept), and passes draw through level-0 views keyed by the
image's generation.
- Foliage: the depth prepass and the lit pass (depth EQUAL) are different variants. Vulkan vertex
stages now declare an invariant gl_Position so both land on the same depth.
- Alpha to coverage is enabled only on a multisampled pass. OpenGL ignores it without MSAA; Vulkan
with one sample dropped every fragment under half alpha.
vk_resources also checks a big-endian 16-bit RGB upload (a normal map). VKRES OK; ludic-dev test
140 passed; the PC's Vulkan frame is validation-clean; OpenGL frames byte-identical at the five
viewpoints with 59 self-tests passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
gpu.ludic's calls branch to the Vulkan backend when it was chosen and came up; every OpenGL
statement is unchanged, only guarded. R3D_GFX=vk selects it in a headless run (the window's
swapchain is the next milestone) and falls back to OpenGL, with the reason, when the device or
the SPIR-V manifest is missing.
- Render state is cached as before and turned into pipelines at the draw; u_* and sampler binds
go into the variant's uniform blocks; meshes, buffers and textures are gvk_* objects;
framebuffer binds are dynamic-rendering passes; same-size blits are image copies; the screen,
the photograph read-back, the present and the screenshot go through the frame.
- Work that submits on its own (uploads, read-backs, new or freed images and buffers) flushes the
frame first, so it runs in OpenGL's order. A read may take fewer channels than the image has
(the height field's R from its RGBA32F bake). Pipeline keys name vertex bindings by order, not
buffer handle, so re-pointed instance buffers keep their pipeline.
The valley renders at frame 90 validation-clean on the RTX 3070 Ti and on MoltenVK. OpenGL frames
byte-identical at the five viewpoints; 59 self-tests pass with no GL error.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The drawing half of the Vulkan backend, compiled into render3d and not yet reached from it:
- gvk_program: a program handle as its manifest variant - two SPIR-V modules, a descriptor set
layout (the vertex block at 0, the fragment block at 1, the samplers at their manifest bindings)
and a pipeline layout. Nothing is compiled at run time.
- gvk_pipeline: one pipeline per program, recorded vertex layout, render state and pass formats,
built the first time that combination draws. Attributes the shader does not read are left out;
the front face is clockwise, since neither API flips y between clip space and its target rows.
- gvk_uniform / gvk_u_set / gvk_bind_texture: loose uniforms written into each stage's block at
the manifest's offsets and array strides; gvk_draw_set copies the blocks into a per-frame ring
at the device's alignment and fills a descriptor set from a per-frame pool, with a white 1x1
texture for a sampler nothing was bound to.
- The frame: one command buffer; a framebuffer bind ends the pass and the next begins at its first
clear or draw (a clear that comes first is the load op); attachments move to attachment layouts
for the pass and back to SHADER_READ_ONLY after it, a cascade drawn through a view of its layer.
gvk_present submits and waits; gvk_screenshot reads the screen image back bottom row first.
r3d.ludic now imports gpu_manifest.ludic too. Textures remember their size.
OpenGL frames byte-identical at the five viewpoints; 59 self-tests pass; VKRES OK and VKDEVICE OK.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
gpu_vk_res.ludic gains what gpu.ludic's texture and buffer handles stand for on Vulkan:
- gvk_tex_storage / _upload / _mips / _read / _release: an image and view per handle, a full mip
chain when pixels come with it (the renderer asks for mipmaps after the upload) and one level
for a target, every level in SHADER_READ_ONLY between uses. Uploads are converted to what the
image stores: a missing alpha filled opaque (three-channel formats are stored with four), 16-bit
PNG samples byte-swapped when the unpack state says so, 32-bit float HDR halved into half
floats. Mips are blitted down level by level; a read-back brings level 0 home.
- gvk_sampler: one VkSampler per filter / wrap / compare / anisotropy combination, made when first
asked for, with GL's defaults where the renderer set nothing.
- gvk_buf_*: vertex, index and instance buffers behind one handle, kept when an upload fits.
Host-visible while the backend comes up.
gpu_vk.ludic switches on anisotropic sampling where the device has it and reads its limit.
examples/rendering/vk_resources.ludic checks it all: VKRES OK, validation-clean, every allocation
freed, on the RTX 3070 Ti (16x anisotropy) and on MoltenVK. One run on the Mac crashed while a
headless game run was using the GPU and did not come back in two reruns. OpenGL frames
byte-identical at the five viewpoints; 59 self-tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The Vulkan backend is two files. gpu_vk.ludic is the device, memory and one-shot commands,
pure Vulkan, and still runs on its own (vk_device.ludic). gpu_vk_res.ludic is the resources in
the renderer's vocabulary - OpenGL's names for formats, filters and blend factors, which only
exist where Gl.* is named - starting with the format table. r3d.ludic imports both after
gpu.ludic; nothing calls them yet.
OpenGL frames byte-identical at the five viewpoints; 59 self-tests pass; VKDEVICE OK.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
