* renderer_vulkan: Respect provoking vertex setting
* renderer_vulkan: Handle rasterization discard
* renderer_vulkan: Implement logic ops
* renderer_vulkan: Properly implement depth clamp and clip
* renderer_vulkan: Handle line width
* Fix build
* vk_pipeline_cache: Don't check depth clamp without a depth buffer
* liverpool: Fix line control offset
* vk_pipeline_cache: Don't run search if depth clamp is disabled
* vk_pipeline_cache: Allow using viewport range when it's more restrictive then depth clamp
* liverpool: Disable depth clip when near and far planes have different setting
* vk_graphics_pipeline: Move warning to pipeline
* vk_pipeline_cache: Revert viewport check and remove log
* vk_graphics_pipeline: Enable depth clamp when depth clip is disabled and extension is not supported
Without the depth clip extension depth clipping is controlled by depth clamping
* shader_recompiler: Replace buffer pulling with attribute divisor for instance step rates
* flatten_extended_userdata: Remove special step rate buffer handling
* Review comments
* spirv_emit_context: Name all instance rate attribs properly
* spirv: Merge ReadConstBuffer again
template function only has 1 user now
* attribute: Add missing attributes
* translate: Reimplement step rate instance id
* Resolve validation warnings
* shader_recompiler: Separate vertex inputs from LS stage, cleanup tess
* buffer_cache: Bring back upload batching and temporary buffer
Because that PR fused the write and read protections under a single function call, it was a requirement to move the actual memory copy part inside the lambda to perform it before the read protection kicks in. However on certain large data transfers it had potential for data corruption. If, for example, an upload had two copies, a 400MB and a 300MB one, the first one would fit in the staging buffer, very likely with an induced stall. However the second one wouldn't have space to fit alongside the other data, but it's also small enough for the buffer to fit it, so the staging buffer would cause a flush and wait to copy it, overwriting the previous transfer.
To address this the upload function has been reworked to allow for batching like previously but with the new locking behavior. Also the condition to use temporary buffers has been expanded to also include cases when staging buffer will stall, which should increase performance a little in some cases.
* buffer_cache: Move buffer barriers and copy outside of lock range
* texture_cache: Async download of GPU modified linear images
* liverpool: Back to less submits
* texture_cache: Don't download depth images
* config: Add option for linear image readback
* vector_alu: Improve handling of mbcnt append/consume patterns
The existing implementation was written to handle a single pattern of mbcnt before the DS_APPEND instruction
v_mbcnt_hi_u32_b32 vX, exec_hi, 0
v_mbcnt_lo_u32_b32 vX, exec_lo, vX
ds_append vY offset:4 gds
v_add_i32 vX, vcc, vY, vX
In this case however the DS_APPEND is before the mbcnt pattern
ds_append vX gds
v_mbcnt_hi_u32_b32 vY, exec_hi, vX
v_mbcnt_lo_u32_b32 vZ, exec_lo, vY
The mbcnt instructions are always in pairs of hi/lo and in general are quite flexible. But they assume the subgroup size is 64 so they are not recompiled literally. Together with DS_APPEND they are used to derive a unique per thread index in a buffer (different from using thread_id as order could be random). DS_APPEND instruction works on per subgroup level, by adding number of active threads of subgroup to the GDS counter, essentially giving a multiple-of-64 base index to all threads. Then each thread executes the mbcnt pair which returns the number of active threads with id less than the itself and adds it with the base.
The recompiler translates DS_APPEND into an atomic increment of a storage buffer counter, which already gives the desired unique index, so this pattern is a no-op. On main it was set to zero as per the first pattern to avoid altering the DS_APPEND result. The new handling passes through the initial value of the pattern instead, which has the same effect but works on either case.
* vk_rasterizer: Always sync DMA buffers
* shader_recompiler: Simplify dma types
Only U32 is needed for S_LOAD_DWORD
* shader_recompiler: Perform address shift on IR level
Buffer instructions now expect address in the data unit they work on. Doing the shift on IR level will allow us to optimize some operations away on common case
* shader_recompiler: Optimize common buffer access pattern
* emit_spirv: Use 32-bit integer ops for fault buffer
Not many GPUs have 8-bit bitwise or operations so that would probably require some overhead to emulate from the driver
* resource_tracking_pass: Fix texel buffer shift
* Import memory
* 64K pages and fix memory mapping
* Queue coverage
* Buffer syncing, faulted readback adn BDA in Buffer
* Base DMA implementation
* Preparations for implementing SPV DMA access
* Base impl (pending 16K pages and getbuffersize)
* 16K pages and stack overflow fix
* clang-format
* clang-format but for real this time
* Try to fix macOS build
* Correct decltype
* Add testing log
* Fix stride and patch phi node blocks
* No need to check if it is a deleted buffer
* Clang format once more
* Offset in bytes
* Removed host buffers (may do it in another PR)
Also some random barrier fixes
* Add IR dumping from my read-const branch
* clang-format
* Correct size insteed of end
* Fix incorrect assert
* Possible fix for NieR deadlock
* Copy to avoid deadlock
* Use 2 mutexes insteed of copy
* Attempt to range sync error
* Revert "Attempt to range sync error"
This reverts commit dd287b48682b50f215680bb0956e39c2809bf3fe.
* Fix size truncated when syncing range
And memory barrier
* Some fixes (and async testing (doesn't work))
* Use compute to parse fault buffer
* Process faults on submit
* Only sync in the first time we see a readconst
Thsi is partialy wrong. We need to save the state into the submission context itself, not the rasterizer since we can yield and process another sumission (if im not understanding wrong).
* Use spec const and 32 bit atomic
* 32 bit counter
* Fix store_index
* Better sync (WIP, breaks PR now)
* Fixes for better sync
* Better sync
* Remove memory coveragte logic
* Point sirit to upstream
* Less waiting and barriers
* Correctly checkout moltenvk
* Bring back applying pending operations in wait
* Sync the whole buffer insteed of only the range
* Implement recursive shared/scoped locks
* Iterators
* Faster syncing with ranges
* Some alignment fixes
* fixed clang format
* Fix clang-format again
* Port page_manager from readbacks-poc
* clang-format
* Defer memory protect
* Remove RENDERER_TRACE
* Experiment: only sync on first readconst
* Added profiling (will be removed)
* Don't sync entire buffers
* Added logging for testing
* Updated temporary workaround to use 4k pages
* clang.-format
* Cleanup part 1
* Make ReadConst a SPIR-V function
---------
Co-authored-by: georgemoralis <giorgosmrls@gmail.com>
* shader_recompiler: Add lowering pass for when 64-bit float is unsupported.
* shader_recompiler: Fix PackDouble2x32/UnpackDouble2x32 type.
* shader_recompiler: Remove extra bit cast implementations.
