def openMirrorKernel(self, suffix):
kernel = copy.deepcopy(self.currentKernel)
kernel.kernelName = self.currentKernel.kernelName + suffix
self.kernels.append(kernel)
self.mirrorKernel = kernel
self.kernelCalls.append(KernelCall.generated(kernel))
return kernel
def createAgg(name, minEntries, payload, codegen):
mem = HashTableMemory(minEntries, codegen)
mem.ht = mem.addColumn(qlib.Type.AGG_HT + "<" + payload.typeName + ">",
name)
call = KernelCall.library(qlib.Krnl.INIT_AGG_HT,
[mem.ht.getGPU(), mem.numEntries],
payload.typeName)
codegen.gpumem.cudaMallocHT.add(call.get())
mem.addToKernel(codegen.currentKernel)
return mem
def createMulti(name, minHtEntries, payload, numPayloads, codegen):
mem = HashTableMemory(minHtEntries, codegen)
mem.ht = mem.addColumn(qlib.Type.MULTI_HT, name)
mem.payload = mem.addSizedColumn(payload.typeName, name + "_payload",
numPayloads)
call = KernelCall.library(qlib.Krnl.INIT_MULTI_HT,
[mem.ht.getGPU(), mem.numEntries])
codegen.gpumem.cudaMallocHT.add(call.get())
mem.addToKernel(codegen.currentKernel)
return mem
def htInsertMultiMatch(self, ctxt):
# execute only when current thread has active elements
with IfClause(ctxt.vars.activeVar, ctxt.codegen):
# compute a (possibly) non-unique hash over all join attributes
hashVar = Variable.val(CType.UINT64,
"hash" + str(self.algExpr.opId),
ctxt.codegen, intConst(0))
with IfClause(ctxt.vars.activeVar, ctxt.codegen):
Hash.attributes(self.algExpr.buildKeyAttributes, hashVar, ctxt)
htRangeOffset = Variable.val(CType.INT,
"offs" + str(self.algExpr.opId))
ctxt.codegen.gpumem.local(htRangeOffset, intConst(0))
scanCall = KernelCall.library("scanMultiHT", [
self.htmem.ht.getGPU(), self.htmem.numEntries,
htRangeOffset.getGPU()
])
ctxt.codegen.kernelCalls.append(scanCall)
ctxt.codegen.openMirrorKernel("_ins")
emit(
call(qlib.Fct.HASH_COUNT_MULTI,
[self.htmem.ht, self.htmem.numEntries, hashVar]),
ctxt.codegen.currentKernel)
ctxt.codegen.mirrorKernel.addVar(htRangeOffset)
payl = self.payload.materialize("payl", ctxt.codegen.mirrorKernel,
ctxt)
emit(
call(qlib.Fct.HASH_INSERT_MULTI, [
self.htmem.ht, self.htmem.payload, htRangeOffset,
self.htmem.numEntries, hashVar,
addressof(payl)
]), ctxt.codegen.mirrorKernel)
def openKernel(self, kernel):
self.kernels.append(kernel)
self.currentKernel = kernel
self.kernelCalls.append(KernelCall.generated(kernel))
return kernel
def initVar(self, var, init):
call = KernelCall.library(
"initArray",
[var.getGPU(), str(init), var.numElements], var.dataType)
self.cudaMallocHT.add(call.get())