createDeviceEngine

Note: This function is not supported by Community Edition. You can get a trial of Shark from DolphinDB official website.

Syntax

createDeviceEngine(name, metrics, dummyTable, outputTable, [keyColumn],[keepOrder])

Arguments

name is a string of the engine name. It is the only identifier of a reactive state engine on a data/compute node. It can have letter, number and "_" and must start with a letter.

metrics is metacode specifying the formulas for calculation. It can include one or more expressions, built-in functions, or user-defined functions. User-defined functions must have a single return value and must not contain embedded for loops or any loops that exceed 100 iterations. It can also include a constant or a vector of constants (in which case, the output column must be of the array vector type). For more information about metacode, refer to Metaprogramming.

dummyTable is a table object whose schema must be the same as the subscribed stream table. Whether dummyTable contains data does not matter.

outputTable is the output table for the results. It can be an in-memory table or a DFS table. Create an empty table and specify the column names and types before calling the function.

keyColumn (optional) is a STRING scalar/vector indicating the grouping column(s).

keepOrder (optional) specifies whether to preserve the insertion order of the records in the output table. The default value is false, meaning data is sorted by keyColumn.

Details

Create a device engine which conducts the calculations defined in the metrics with GPU acceleration.

Note: The device engine does not retain any state information from one batch of data to the next. In other words, when two batches are processed, the second batch is processed independently of the first batch.

The following functions can be accelerated using GPUs:

  • Basic unary operation: not, neg, cast, log, log2, log10 ,log1p ,abs, sign, sqrt, sin, sinh, asin, asinh, cos, cosh, acos, acosh, tan, tanh, atan, atanh, reciprocal, cbrt, exp, exp2, expm1

  • Basic binary operation: add, sub, mul, div, ratio, pow, lt, gt, le, ge, eq, ne, and, or, or__, max, min

  • Binary operation on integers: mod, bitAnd, bitOr, bitXor, lshift, rshift

  • Ternary operation: iif

  • Unary moving functions: mavg, msum, mcount, mprod, mvar, mvarp, mstd, mstdp, mskew, mkurtosis, mmin, mmax, mimin, mimax, sma, wma, mfirst, mlast, mrank, mmaxPositiveStreak, mmed, mpercentile, mmad (useMedian is currently not supported)

  • TALib-series Unary moving functions: sma, ema, wma, dema, tema, trima, t3, wilder, gema, linearTimeTrend, ma, talib (only the m- /moving functions and mTopN-/moving TopN functions are accepted)

  • Binary moving functions: mcorr, mbeta, mcovar, mwsum, mwavg

  • Other moving functions: linearTimeTrend, mslr

  • Unary cumulative functions: cumsum, cumprod, cummin, cummax, cumvar, cumvarp, cumstd, cumstdp, cumnunique, cumfirstNot, cumlastNot, cumavg, cumcount, cumPositiveStreak

  • Binary cumulative functions: cumcorr, cumcovar, cumbeta, cumwsum, cumwavg

  • Order-sensitive functions: deltas, ratios, ffill, move, prev, next, percentChange, iterate, prevState, ewmMean, ewmVar, ewmStd, ewmCov, ewmCorr

    Note: For ewmVar, ewmStd, ewmCov, and ewmCorr, the adjust parameter must be set to false and bias must be true.
  • Moving TopN functions: msumTopN, mavgTopN, mstdpTopN, mstdTopN, mvarTopN, mvarpTopN, mwsumTopN, mcorrTopN, mcovarTopN, mbetaTopN,mskewTopN,mkurtosisTopN

  • Row-based functions: rowMin, rowMax, rowAnd, rowOr, rowXor, rowProd, rowSum, rowSum2, rowSize, rowCount, rowAvg, rowVar, rowVarp, rowStd, rowStdp

  • Time-based moving functions: tmsum, tmsum2, tmavg, tmprod, tmcount, tmvar, tmvarp, tmstd, tmstdp, tmcovar, tmcorr, tmwavg, tmwsum, tmbeta, tmfirst, tmlast, tmmin, tmmax, tmskew, tmkurtosis, tmove

  • Other functions: TrueRange, topRange, lowRange, stateMavg

Note: Starting from version 3.00.1, when the absolute value of a calculation result is less than DBL_EPSILON*10000 (approximately 2.22*10^-12), all moving functions and cumulative window functions will retain full precision instead of rounding the result.

Details

// create a device engine
dummyTb = table(1:0, `sym`id`value, [SYMBOL,INT,DOUBLE])
share table(100:0, `sym`id`flag`value`factor, [SYMBOL,INT,SYMBOL,DOUBLE,DOUBLE]) as result
de = createDeviceEngine(name="myDe", metrics=[<id>,<"flag"+"_A">,<value>,<mavg(value,5)>], dummyTable=dummyTb, outputTable=result, keyColumn="sym")


// simulate data
data1 = table(take("A", 100) as sym, 1..100 as id, double(10+1..100) as value)
data2 = table(take("B", 100) as sym, 1..100 as id, double(20+1..100) as value)
data3 = table(take("C", 100) as sym, 1..100 as id, double(30+1..100) as value)
data = data1.unionAll(data2).unionAll(data3).sortBy!(`id)

// write data
de.append!(data)
select top 10 * from result
sym id flag value factor
A 1 flag_A 11
A 2 flag_A 12
A 3 flag_A 13
A 4 flag_A 14
A 5 flag_A 15 13
A 6 flag_A 16 14
A 7 flag_A 17 15
A 8 flag_A 18 16
A 9 flag_A 19 17
A 10 flag_A 20 18