orklev2-ci

name	orklev2-ci
description	Answer questions about orkid's ComputeInterface (CI), compute shader dispatch, indirect dispatch, SSBO binding for compute, memory barriers, image/sampler binding, dispatch phases, and PyTorch tensor integration. Use when the user asks about compute shaders, GPU compute, or dispatch.
user-invocable	false

Orkid ComputeInterface (CI) Reference

When answering questions about compute shaders or GPU compute in orkid, consult the files below. All under ork.lev2/.

Key Files

Component	File
CI Interface	inc/ork/lev2/gfx/ci.h (lines 8–44)
FxComputeShader	inc/ork/lev2/gfx/shadman.h
Vulkan CI	src/gfx/vulkan/headers/vulkan_ctx.h (lines 584–619)
Python Bindings	pyext/src/pyext_gfx.cpp

CI Interface (ci.h:8–44)

ImageBindAccess Enum

• EIBA_READ_ONLY = 0
• EIBA_WRITE_ONLY = 1
• EIBA_READ_WRITE = 2

Dispatch Phase Management

• beginDispatchPhase() — suspends active render pass if needed
• endDispatchPhase() — resumes render pass

Compute Dispatch

• dispatchCompute(shader, numgroups_x, numgroups_y, numgroups_z) — direct dispatch
• dispatchComputeIndirect(shader, indirect_ptr) — indirect dispatch from buffer

Resource Binding

• bindStorageBuffer(shader, binding_index, buffer) — bind SSBO to compute shader
• bindImage(shader, binding_index, texture, access) — bind image for load/store
• bindSampler(shader, binding_index, texture) — bind texture sampler

Memory Barriers

• storageBarrier() — ensures prior SSBO writes are visible to subsequent dispatches

PyTorch Integration (ENABLE_PYTORCH builds)

• copyTensorIntoStorageBuffer(ssbo, tensor, dest_offset) — tensor → SSBO
• storageBufferFromTensor(tensor) → FxShaderStorageBuffer* — wrap tensor as SSBO

Typical Usage

auto* CI = ctx->CI();
auto* FXI = ctx->FXI();

// Get compute shader from loaded FxShader
auto* cs = FXI->computeShader(shader, "cs_blur");

CI->beginDispatchPhase();

// Bind resources
CI->bindStorageBuffer(cs, 0, input_ssbo);
CI->bindStorageBuffer(cs, 1, output_ssbo);
CI->bindImage(cs, 2, texture, ComputeInterface::EIBA_WRITE_ONLY);

// Dispatch workgroups
CI->dispatchCompute(cs, groups_x, groups_y, groups_z);

// Barrier between passes
CI->storageBarrier();

// Second dispatch can read first dispatch's output
CI->dispatchCompute(cs2, groups_x, groups_y, groups_z);

CI->endDispatchPhase();

Vulkan Implementation Notes

The Vulkan CI (VkComputeInterface) handles:

• Suspending/resuming active render passes for compute dispatch
• Pipeline binding for compute shaders (separate from graphics pipeline)
• Descriptor set binding for compute resources
• vkCmdDispatch / vkCmdDispatchIndirect calls
• VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT barriers

How to Answer

1. For CI methods: check ci.h:8–44
2. For FxComputeShader: check shadman.h
3. For SSBO creation/mapping: consult orklev2-fxi skill (FXI manages buffer creation)
4. For Vulkan details: check vulkan_ctx.h:584–619