International Business Machines Corporation

John D Keenleyside

Berthold Reinwald

Shirish Tatikonda

Matthias Boehm

Alexandre Valentinovich Evfimievski

Sunil Jeevananda Kamath

Keith W Campbell

Seyed Sina Meraji

Kajatheepan Kanagaratnam

Arash Ashari

Kevin Alexander Stoodley

Shereen M Ghobrial

According to one exemplary embodiment, a method for processing a query associated with a database is provided. The method may include receiving the query. The method may include estimating a number of groups. The method may include copying a plurality of data from the database to graphics processing unit (GPU) memory. The method may include creating a hash table in GPU memory. The method may include determining if a group associated with the database is present in the hash table. The method may include adding the group to the hash table based on determining that the group is not present in the hash table. The method may include aggregating a value associated with the group in the hash table based on determining that the group is present in the hash table. The method may include determining a plurality of results. The method may then include retrieving the plurality of results.

Fast query processing in columnar databases with GPUs

A method for optimization of machine learning (ML) workloads on a graphics processor unit (GPU). The method includes identifying a computation having a generic pattern commonly observed in ML processes. Hierarchical aggregation spanning a memory hierarchy of the GPU for processing is performed for the identified computation including maintaining partial output vector results in shared memory of the GPU. Hierarchical aggregation for vectors is performed including performing intra-block aggregation for multiple thread blocks of a partial output vector results on GPU global memory.

Pipelined approach to fused kernels for optimization of machine learning workloads on graphical processing units

A method for optimization of machine learning (ML) workloads on a graphics processor unit (GPU). The method includes identifying a computation having a generic pattern commonly observed in ML processes. An optimized fused GPU kernel is employed to exploit temporal locality for inherent data-flow dependencies in the identified computation. Hierarchical aggregation spanning a memory hierarchy of the GPU for processing for the identified computation is performed. GPU kernel launch parameters are estimated following an analytical model that maximizes thread occupancy and minimizes atomic writes to GPU global memory.

The invention enables a compiler (during its first compilation pass) to insert identifiers during an early optimization phase into machine-independent code. The identifiers are used for identifying specific instructions of the machine-independent code. The compiler may use a table to record the assignment of identifiers and their corresponding portion of machine-independent code. During subsequent optimization phases of the first compilation pass, portions of machine-independent code are transformed into portions of machine-dependent code, and the machine-dependent code inherits identification associated with the portions of machine-independent code that were used to generate the portions of machine-dependent code. After the first compilation pass, executable code is made to interact with test data to generate statistical execution data which is collected by previously inserted counters. The collected counter data is consolidated so that it relates to the instructions of the machine-independent code. During a second compilation pass of the compiler, the compiler reads the machine-independent code, the collected or captured counter data, an indication of the association between the identifiers and the instructions (which can be stored in a table). Subsequently, the compiler begins locating instructions of the machine-independent code relative to other instructions of the machine-independent code during an early or first optimization phase of the second compilation pass.

Using identifiers and counters for controlled optimization compilation

A system and method for aligning data in stack memory in a data processing system where the stack memory provides temporary storage for storing parameters for a function call. The method first determines if any of the parameters in the function being call are of a selected type. If a parameter is of a selected type, code is generated for aligning the parameter on a stricter boundary than the default boundary for the stack memory. Code is then generated to align the remaining parameters in the function call on the default boundary in the stack memory. The aligned parameter in the stack provides a reference point which is used by the called function to align locally scoped variables in the stack. By aligning a parameter of a selected type on stricter boundary in the stack, for example, a double precision floating point aligned on an 8 byte boundary, the execution performance of the compiled program code is improved.

Growing community of inventors

Pickering, Canada

John D Keenleyside