Philip Rebohle
9721938e57
[d3d11] Introduce d3d11.invariantPosition option
...
Apparently it's better to make app profiles than risk tanking
performance for all applications. Invariance rules in D3D11 are
undocumented so it's not clear what the correct solution is.
2020-01-23 01:28:19 +01:00
Philip Rebohle
11f08c7dea
[dxbc] Declare vertex positions as invariant
...
Otherwise, games relying on different vertex shaders to produce
the same numerical results may suffer from Z-fighting issues.
Fixes #1364 .
2020-01-22 23:59:51 +01:00
Philip Rebohle
6690a31d8d
[dxbc] Don't enable SPV_KHR_shader_draw_parameters
...
Core in SPIR-V 1.3.
2020-01-17 17:46:59 +01:00
Philip Rebohle
31948cae8c
[spirv] Allow specifying the SPIR-V version explicitly
...
We're going to use some SPIR-V 1.4 features for D3D11 if supported,
but 1.4 is not supported by all implementations.
2019-12-18 17:36:46 +01:00
Philip Rebohle
27ea176295
[d3d11] Remove d3d11.strictDivision option
2019-11-30 20:39:32 +01:00
Philip Rebohle
3b030d9569
[dxbc] Implement workaround to replace NaN render target outputs by zero
2019-11-30 20:39:32 +01:00
Philip Rebohle
49e7df96ec
[dxbc] Emit a spec constant for uniform buffers
...
This partially reverts commit fd547b666e
.
For some reason, not doing so breaks Overwatch (because of course it does).
2019-11-28 00:55:38 +01:00
Philip Rebohle
e95bc3256f
[d3d11] Move handling of constantBufferRangeCheck option to D3D11Options
2019-10-30 10:57:36 +01:00
Philip Rebohle
c4e5323e0e
[dxbc] Remove old constant buffer range check
2019-10-28 17:52:50 +01:00
Philip Rebohle
fd547b666e
[dxbc] Don't emit a spec constant for uniform buffers
2019-10-28 15:44:02 +01:00
Philip Rebohle
3d213efe53
[dxbc] Use SSBOs for dynamically indexed constant buffers if needed
...
SSBOs are tightly bound-checked on all Nvidia GPUs, so this allows for
a more accurate workaround for games relying on OOB access behaviour.
2019-10-28 15:42:46 +01:00
Philip Rebohle
68760f5b20
[dxbc] Parse dynamically indexed flag for constant buffers
2019-10-28 14:15:44 +01:00
Philip Rebohle
6888a98c89
[dxbc] Always set geometry shader invocation count
...
Fixes validation errors when the DXBC shader does not specify
the invocation count itself.
2019-10-26 23:19:59 +02:00
Philip Rebohle
144f4badef
[dxbc] Always allocate at least one input array element
...
Turns out that shaders with no inputs can still have an input signature,
in which case we were generating a zero-length array which is illegal.
2019-10-26 23:13:59 +02:00
Philip Rebohle
2d6a3396dd
[dxbc] Declare length of shader input array late
...
Otherwise, we generate out-of-bounds array reads and writes
in broken ENB shaders.
2019-10-23 16:36:52 +02:00
Philip Rebohle
59d4556641
[dxbc] Extend scalar constants to vectors if required
...
Needed for ENB. Fixes #865 .
2019-10-23 15:26:22 +02:00
Philip Rebohle
ec8ca8a403
[dxbc] Assume vec4 type for i/o vars not declared in signature
...
Works around an ENB bug.
2019-10-23 15:26:22 +02:00
Philip Rebohle
3e12a116bb
[dxbc] Declare temp registers on demand
...
And ignore dcl_temps. Needed to successfully compile some broken
ENB shaders (this does work on Windows too). See #865 .
2019-10-23 13:55:04 +02:00
Philip Rebohle
8211225ad4
[dxbc] Swap OpSelect arguments for Ne/Dne, don't use OpLogicalNot
...
Saves one instruction.
2019-10-18 03:22:20 +02:00
Philip Rebohle
803ec3542d
[dxbc] Fix sample positions
...
From the D3D11.3 functional specification:
"The sample position is relative to the pixel's center"
Fixes wine test failures.
2019-10-11 18:41:10 +02:00
Philip Rebohle
59d4e8a36a
[dxbc] Fix NaN handling in not-equal comparison
...
Ne and Dne are unordered in DXBC. Avoid FUnordNotEqual to
avoid potential issues with drivers and debugging tools.
2019-09-30 00:19:47 +02:00
Philip Rebohle
b6cf518962
Revert "[dxbc] Always decorate SV_POSITION as invariant"
...
This reverts commit 5ee790a4a3
.
Breaks Shadow of the Tomb Raider on Nvidia for unknown reasons.
2019-09-24 23:01:36 +02:00
Philip Rebohle
5ee790a4a3
[dxbc] Always decorate SV_POSITION as invariant
...
Apparently fixes Z-Fighting in Saints Row IV on Nvidia.
2019-09-24 22:19:43 +02:00
Philip Rebohle
f38cfd592e
[dbxc] Actually fix callc implementation
...
Yeah maybe we should close our blocks properly. There seems to be
no way to generate a callc instruction with fxc so this is untested.
2019-09-22 22:36:42 +02:00
Philip Rebohle
94ce76da6b
[dxbc] Implement label, call and callc instructions
...
Fixes #1200 .
2019-09-22 22:28:58 +02:00
Philip Rebohle
17a6c4168e
[dxbc] Implement stencil ref export from fragment shaders
...
This is an optional feature in D3D11.3, but easy enough to support.
2019-09-16 12:42:15 +02:00
Philip Rebohle
26afaba410
[dxbc] Implement MSAD4 instruction
...
Apparently we're required to support this for Direct3D 11.2+.
There are currently no known games that require this instruction.
2019-09-16 12:42:14 +02:00
Philip Rebohle
6ab074c95b
[dxbc] Only use atomic append/consume optimization in compute shaders
...
- For fragment shaders, this isn't safe since ballots include helper invocations
- For vertex shaders, if drivers don't support subgroup operations in those
stages, we don't want it to affect the performance of compute shadres.
2019-07-26 14:07:39 +02:00
Philip Rebohle
9c5102e257
[dxvk] Don't duplicate geometry shader system value outputs
...
Fixes #1121 . The basic idea here is that all built-ins that can be
written by the GS will be consumed as built-ins by the FS anyway,
so we do not need to keep the o# variable around.
2019-07-24 18:18:20 +02:00
Philip Rebohle
8644d75722
[dxbc] Use OpDemoteToHelperInvocationEXT for discards if supported
2019-07-11 19:33:17 +02:00
Liam Middlebrook
2e83aae472
[dxbc] Print shader signatures at debug loglevel
2019-07-11 12:40:44 +02:00
Liam Middlebrook
ff4fa29304
[dxbc] Add string output function for DxbcRegMask
2019-07-11 12:40:44 +02:00
Liam Middlebrook
2164a44887
[dxbc] Add ostream operator for DxbcScalarType
2019-07-11 12:40:44 +02:00
Philip Rebohle
06d4e06d8a
[dxbc] Enable early discard on RADV/ACO
...
Still can't enable on LLVM due to GPU hangs. The way we detect it
may have to change in the future, but for now, this should do.
2019-07-08 16:30:56 +02:00
Philip Rebohle
58dcf77c08
[dxbc] Disable atomic counter optimization on old RADV versions
...
Breaks TressFX in Tomb Raider 2013 for some reason, whereas this works
fine on 19.1 with both LLVM 8.0 and 9.0-git.
2019-06-15 12:58:31 +02:00
Philip Rebohle
d5ba36443e
[dxbc] Fix useRawSsbo option
...
Add some break to fix some breakage.
2019-06-11 20:23:47 +02:00
Philip Rebohle
d755b18ccb
[dxbc] Repurpose useRawSsbo option
...
Basically, abuse the minSsboAlignment property to convince the
compiler to always/never emit raw SSBOs for raw and structured
buffers, and use the actual device limit by default.
2019-06-11 20:20:40 +02:00
Philip Rebohle
0f7c23c022
[dxbc] Always use raw storage buffers if alignment requirements are met
2019-06-11 20:20:40 +02:00
Philip Rebohle
163e6aaa50
[dxbc] Store alignment for raw and structured buffers
2019-06-11 20:20:40 +02:00
Philip Rebohle
7d9a75c82c
[dxbc] Use subgroup operations for atomic append/consume operations
...
Reduces the number of atomic operations performed per subgroup to 1.
2019-05-15 19:32:27 +02:00
Philip Rebohle
d94d89c3ef
[dxbc] Add option to use subgroup ops for atomic counter operations
...
This can greatly reduce the number of atomic operations when using
append/consume buffers.
2019-05-15 18:49:02 +02:00
Philip Rebohle
04e6479690
[dxbc] Remove old spec constant code
2019-05-01 01:57:34 +02:00
Philip Rebohle
dc3cfc9fa0
[dxbc] Use new spec constant API for rasterizer sample count
2019-05-01 01:57:34 +02:00
Philip Rebohle
7111af423d
[dxbc] Add new emitNewSpecConstant method
...
Convenience method to declare new specialization constants.
2019-05-01 01:57:34 +02:00
Philip Rebohle
dc52212873
[dxbc] Disable Constant Buffer Range Check on AMD
...
The hardware already behaves as intended, no need to waste GPÜ cycles.
2019-04-30 16:07:24 +02:00
Philip Rebohle
3b1e753bb5
[dxvk] Re-implement early discard with quad granularity
...
May perform better on some hardware in situations where we cannot
discard a full subgroup. Closes #753 .
2019-04-30 14:46:03 +02:00
Danylo Piliaiev
261d31cac6
[dxbc] Fix xfb passthrough for system values
...
vReg should be always allocated for system values which is
necessary for emitXfbOutputSetup and is already done for
hull shader passthrough.
Signed-off-by: Danylo Piliaiev <danylo.piliaiev@globallogic.com>
2019-04-30 12:19:21 +02:00
Philip Rebohle
06c144f075
[dxbc] Store sample positions as vec2 array
...
We can append the zeroes in shader code instead. May
improve generated code on drivers that use scratch
memory or temporary uniform buffers for large arrays.
2019-04-29 14:04:42 +02:00
Philip Rebohle
b14ad7b30c
[dxbc] Remove some old TODOs
...
This is already implemented properly.
2019-04-29 11:48:09 +02:00
Philip Rebohle
04152055b7
[dxbc] Correctly report a sample count of 0 for unbound images
2019-04-29 11:48:09 +02:00