AMDGPUCoExecSchedStrategy.cpp File Reference

Coexecution-focused scheduling strategy for AMDGPU. More...

#include "AMDGPUCoExecSchedStrategy.h"
#include "AMDGPUIGroupLP.h"
#include "GCNHazardRecognizer.h"
#include "llvm/Support/Debug.h"

Go to the source code of this file.

Macros
#define	DEBUG_TYPE   "machine-scheduler"

Functions
static SUnit *	pickOnlyChoice (SchedBoundary &Zone)
static void	allowCoExec (llvm::AMDGPU::CoExecInfo &Info, llvm::AMDGPU::CoExecMaskT ExtraBits, unsigned StartIndex)
	Apply ExtraBits to every slot in Info starting with StartIndex Used by MFMA co-exec rules, because MFMA co-exec slots are incremental, i.e.

Variables

static cl::opt< CarriedLatency >

BlockCarriedLatency ("amdgpu-block-carried-latency", cl::Hidden, cl::init(CarriedLatency::Off), cl::desc("Estimate block-carried latency and include it in the effective " "candidate stall cost."), cl::values(clEnumValN(CarriedLatency::Off, "off", "Disabled - do not pad latency."), clEnumValN(CarriedLatency::Fence, "fence", "Only pad latency for memory fence (e.g. those surrounding " "barrier_signal/wait)."), clEnumValN(CarriedLatency::All, "all", "Pad latency for any SU with an incoming ds_load dependency.")))

Detailed Description

Coexecution-focused scheduling strategy for AMDGPU.

Definition in file AMDGPUCoExecSchedStrategy.cpp.

Macro Definition Documentation

DEBUG_TYPE

#define DEBUG_TYPE   "machine-scheduler"

Definition at line 22 of file AMDGPUCoExecSchedStrategy.cpp.

Function Documentation

allowCoExec()

void allowCoExec ( llvm::AMDGPU::CoExecInfo & Info, llvm::AMDGPU::CoExecMaskT ExtraBits, unsigned StartIndex )

static

Apply ExtraBits to every slot in Info starting with StartIndex Used by MFMA co-exec rules, because MFMA co-exec slots are incremental, i.e.

for every slot N it supports all instructions which were supported by the previous slot N-1 and may support something extra.

Definition at line 68 of file AMDGPUCoExecSchedStrategy.cpp.

Referenced by llvm::AMDGPU::getMFMACoExecInfo().

pickOnlyChoice()

SUnit * pickOnlyChoice ( SchedBoundary & Zone )

static

Definition at line 55 of file AMDGPUCoExecSchedStrategy.cpp.

References llvm::ReadyQueue::empty(), llvm::SchedBoundary::Pending, and llvm::SchedBoundary::pickOnlyChoice().

Referenced by llvm::AMDGPUCoExecSchedStrategy::pickNode(), llvm::GCNSchedStrategy::pickNode(), and llvm::GCNSchedStrategy::pickNodeBidirectional().

Variable Documentation

BlockCarriedLatency

cl::opt< CarriedLatency > BlockCarriedLatency("amdgpu-block-carried-latency", cl::Hidden, cl::init(CarriedLatency::Off), cl::desc("Estimate block-carried latency and include it in the effective " "candidate stall cost."), cl::values( clEnumValN(CarriedLatency::Off, "off", "Disabled - do not pad latency."), clEnumValN(CarriedLatency::Fence, "fence", "Only pad latency for memory fence (e.g. those surrounding " "barrier_signal/wait)."), clEnumValN( CarriedLatency::All, "all", "Pad latency for any SU with an incoming ds_load dependency."))) ( "amdgpu-block-carried-latency" , cl::Hidden , cl::init(CarriedLatency::Off) , cl::desc("Estimate block-carried latency and include it in the effective " "candidate stall cost.") , cl::values( clEnumValN(CarriedLatency::Off, "off", "Disabled - do not pad latency."), clEnumValN(CarriedLatency::Fence, "fence", "Only pad latency for memory fence (e.g. those surrounding " "barrier_signal/wait)."), clEnumValN( CarriedLatency::All, "all", "Pad latency for any SU with an incoming ds_load dependency."))  )

static

Referenced by llvm::CandidateHeuristics::getCarriedLatency().