Shader optimizations to reduce VGPR usage and increase occupancy#45023
Conversation
|
|
||
| light_data.attenuation_energy[0] = Math::make_half_float(storage->light_get_param(base, RS::LIGHT_PARAM_ATTENUATION)); | ||
| light_data.attenuation_energy[1] = Math::make_half_float(sign * storage->light_get_param(base, RS::LIGHT_PARAM_ENERGY) * Math_PI); | ||
| light_data.attenuation = storage->light_get_param(base, RS::LIGHT_PARAM_ATTENUATION); |
There was a problem hiding this comment.
This just cleaned up how parameters are sent to the reflection probe struct
| uint8_t color_specular[4]; //rgb color, a specular (8 bit unorm) | ||
| uint16_t cone_attenuation_angle[2]; // attenuation and angle, (16bit float) | ||
| uint8_t shadow_color_enabled[4]; //shadow rgb color, a>0.5 enabled (8bit unorm) | ||
|
|
| } | ||
| #endif //LOW_END_MODE | ||
| // multiply by albedo | ||
| diffuse_light *= albedo; // ambient must be multiplied by albedo at the end |
| #ifndef LOW_END_MODE | ||
| if (scene_data.ssao_enabled) { | ||
| float ssao = texture(sampler2D(ao_buffer, material_samplers[SAMPLER_LINEAR_CLAMP]), screen_uv).r; | ||
| ao = min(ao, ssao); |
There was a problem hiding this comment.
We need to delete SSAO_ao_affect from environment etc.
There was a problem hiding this comment.
For reference, I opened an issue to track this: #54611
1131b2c to
8b6c33c
Compare
|
Im getting 2 types of artifacts with shaders. One is random blocks of shadow that flicker in and out (likely entire clusters missing a light). And the second is that the derivative shadow optimization doesnt scale well with distance from a shadow. So when you get far enough away from a shadow the edges take on a very aliased look
|
8b6c33c to
dc2ec58
Compare
|
@akien-mga Should hopefully be ok to merge but the static checks are not passing and I already have clang-format-11. |
|
@reduz I would recommend avoiding nested ternary operators, it should fix the clang-format issue at the same time. |
dc2ec58 to
9683c70
Compare
|
I haven't reviewed the code yet, but running this build there are some artifacts created on triangle edges. It must be due to the shadow derivative path.
edit: I can confirm that commenting out the shadow derivative path fixes this issue |
9683c70 to
e71492a
Compare
|
@clayjohn disabled it for now |
e71492a to
1f56612
Compare
|
I went over the files, it wasn't an in-depth review but I didn't spot anything wrong. Other than that, Github lists some files as edited, but they don't contain any changes, not sure what's going on. Also, static checks have failed, I think they should be restarted. |
|
I restarted CI jobs for static checks and Android which failed for a seemingly unrelated reason, likely a CI oops. |
Ah no it's not a CI issue, @reduz did remove executable permission on those files, so they can't run on CI: |
1f56612 to
a5fbe24
Compare
Clustering is now GPU based, uses an implementation based on the Activision algorithm.
a5fbe24 to
099dee3
Compare
|
Thanks! |
Reduced from 116 to 80 VGPRs in normal usage, still more work is needed.
Some not often used features are gone for the sake of this and some are disabled, will re-enable later.
-Shadow color per light is gone
-Oren Nayar is gone (implement manually if you really want it, makes little difference)
-SSAO color is gone
-SSAO ao affect is gone because this confused users, min(ao,ssao) now used, should work in most cases.
Also
-softshadows (based on angular distance) are likely going to be removed from directional light, too expensive dont look great, and dont make much of a difference
-projectors are disabled.