RdmaPlayoutProbe.hpp File Reference

Content-verified one-time probe of the GPU->peer (playout) P2P path. More...

Detailed Description

Content-verified one-time probe of the GPU->peer (playout) P2P path.

Pinning a CUDA buffer for a capture card only proves nvidia_p2p_get_pages accepted the range. It does not prove the card can actually read that memory over PCIe: with the IOMMU in full-translation mode, or with the card and the GPU behind different host bridges, the transfer is silently dropped while every SDK call still returns success. The card then plays out whatever its framestore already held — a constant frame — and the pump reports a full frame count with zero drops.

The capture direction is already content-verified by its vendor adapter (seed the pinned buffer, run a real card->GPU transfer, check the bytes changed). This is the playout counterpart, and it lives here rather than in an addon so every vendor gets the same assertion: seed the pinned buffer with a known non-degenerate pattern, ask the vendor to push it to the device, ask the vendor to read it back, and compare 1:1.

The GPU seeding is injected (RdmaPlayoutProbeIo) so the decision table is exercisable without a CUDA driver.

Go to the source code of this file.

Classes
struct	score::gfx::interop::RdmaPlayoutProbeIo

Namespaces
namespace	score
	Base toolkit upon which the software is built.
namespace	score::gfx
	Graphics rendering pipeline for ossia score.

Enumerations
enum class	score::gfx::interop::RdmaPlayoutProbeResult {   Delivered , Unverified , NoProbe , SeedFailed ,   TransferFailed , ReadbackFailed , ContentMismatch }

Functions
unsigned char	score::gfx::interop::rdmaPlayoutProbeByte (std::uint32_t index) noexcept
RdmaPlayoutProbeResult	score::gfx::interop::rdmaProbePlayoutPath (void *pinnedGpuPtr, std::uint32_t frameByteSize, const VendorDmaRegistrar &registrar, const RdmaPlayoutProbeIo &io) noexcept
	Run the probe. pinnedGpuPtr must already be pinned by the vendor.
bool	score::gfx::interop::rdmaPlayoutProbeEngages (RdmaPlayoutProbeResult r) noexcept
	True for the results on which an RDMA output rung may engage.
const char *	score::gfx::interop::rdmaPlayoutProbeMessage (RdmaPlayoutProbeResult r) noexcept

Variables
constexpr std::uint32_t	score::gfx::interop::rdmaPlayoutProbeMaxBytes = 64u * 1024u

Enumeration Type Documentation

RdmaPlayoutProbeResult

enum class score::gfx::interop::RdmaPlayoutProbeResult

strong

Enumerator
Delivered	Pushed and read back byte-identical: the peer really reads GPU memory.
Unverified	Pushed successfully, but the vendor supplied no readback hook, so delivery rests on a return code alone.
NoProbe	The vendor supplied no verifyTransfer at all — nothing was probed.
SeedFailed	The pattern could not be written into the pinned GPU buffer.
TransferFailed	verifyTransfer returned false.
ReadbackFailed	readbackTransfer returned false.
ContentMismatch	The peer returned bytes that are not the ones it was given.

Function Documentation

rdmaPlayoutProbeByte()

unsigned char score::gfx::interop::rdmaPlayoutProbeByte ( std::uint32_t  index )

noexcept

Deterministic, non-degenerate pattern. Never constant and never all-zero, so a dropped transfer (stale framestore, zero-filled scratch) cannot alias a correct result.

Variable Documentation

rdmaPlayoutProbeMaxBytes

constexpr std::uint32_t score::gfx::interop::rdmaPlayoutProbeMaxBytes = 64u * 1024u

inlineconstexpr

Bytes the probe seeds and compares: enough to cross several GPU pages, capped so the probe stays cheap at 8K rasters.