Portable RT: Fixes for shader-binding-table assembly (sbt data) #1993
Conversation
|
CI gfxreconstruct build queued with queue ID 356754. |
fabian-lunarg
force-pushed
the
fabian_work_on_sbt_data
branch
from
7074c8a to
e136f23
Compare
|
CI gfxreconstruct build queued with queue ID 356756. |
|
CI gfxreconstruct build # 5935 running. |
|
CI gfxreconstruct build # 5935 passed. |
| { | ||
| auto* previous_pipeline = object_table_->GetVkPipelineInfo( | ||
| command_buffer_info->bound_pipelines.at(VK_PIPELINE_BIND_POINT_COMPUTE)); | ||
| GFXRECON_ASSERT(previous_pipeline); |
There was a problem hiding this comment.
previous_pipeline MUST be valid?
There was a problem hiding this comment.
yes. at this point we already checked the bound_pipelines map and know a compute-pipeline was previously bound (rare'ish case anyway, since the command-buffer is used to record a raytracing pipeline). so in fact, we only assert that it can also be mapped.
There was a problem hiding this comment.
double-checked the assumption, I think it's correct. the only caveat I could think of if a compute-pipeline would have been bound and then unbound again. but that's not allowed ("pipeline must be a valid VkPipeline handle")
There was a problem hiding this comment.
@antonio-lunarg is your concern resolved?
| { | ||
| auto* previous_pipeline = object_table_->GetVkPipelineInfo( | ||
| command_buffer_info->bound_pipelines.at(VK_PIPELINE_BIND_POINT_COMPUTE)); | ||
| GFXRECON_ASSERT(previous_pipeline); |
There was a problem hiding this comment.
@antonio-lunarg is your concern resolved?
|
CI gfxreconstruct build queued with queue ID 358099. |
|
CI gfxreconstruct build # 5951 running. |
|
CI gfxreconstruct build queued with queue ID 358164. |
|
CI gfxreconstruct build # 5952 running. |
|
CI gfxreconstruct build # 5952 passed. |
|
CI gfxreconstruct build queued with queue ID 359206. |
|
CI gfxreconstruct build # 5960 running. |
|
CI gfxreconstruct build # 5960 passed. |
fabian-lunarg
force-pushed
the
fabian_work_on_sbt_data
branch
from
f47de41 to
a68aeea
Compare
|
CI gfxreconstruct build queued with queue ID 359348. |
|
CI gfxreconstruct build # 5963 running. |
|
CI gfxreconstruct build # 5963 passed. |
…rG#1993) * Account for all bind-points when tracking bound pipelines * Adjust internal compute-shader: pass-through data that could not be mapped * Keep track of opaque/assigned acceleration-structure addresses * Rebind previously bound compute-pipelines, if any * Adjust internal compute-shader: pass-through data that could not be mapped (SBT) * rework I/O address calculation, account for data-payloads and pass-through * Drop GL_GOOGLE_include_directive from internal compute-shader, not used * Use accelerationStructureCaptureReplay only when allocator supports opaque addresses
GL_GOOGLE_include_directivefrom internal compute-shader, unused and caused issues on radvaccelerationStructureCaptureReplayonly when allocator supports opaque addressesthis fixes portable replay (using
-m rebind) of applications using shader-record buffers, sourcing data from shader-binding-tables.in glsl e.g.
layout(shaderRecordEXT, std430) buffer ShaderRecordBuffer{...}in SPIRV:
StorageClassShaderRecordBufferKHRin presence of that feature, when re-assembling a new sbt, it's not sufficient to remap handles.
also the extra data needs to be copied/passed-through.
this is fixed in this PR.
the simplest example around is: https://github.com/SaschaWillems/Vulkan/tree/master/examples/raytracingsbtdata
also this PR fixes issues encountered with portable-Rt on radv