Compiler Support to AI on FPGAs

Sparsity and blocked algorithms for AIE2 FPGAs

This is a consulting-research project for the implementation of the Fast Matrix Multiplication using Winograd 's algorithms for high performance computing systems: symmetric and heterogeneous systems. Recently I have been working with GPUs , TR, and in the future with FPGAs.

Compiler support for the implementation of convolution networks into FPGAs.

Inference as a Computation DAG, memory computation, instruction scheduling, graph optimization, weight reorganization, and code generation. Support for the third generation of FPGA kernels.

CNN Front ends: Caffe, TensorFlow, MXNET, and Keras.

Large scale matching algorithms for the unbiased computation of lift (7 patents)

Deterministic geographical distributed graph algorithms for the computation of user profiles (features)

Lift as time series.

Currently all projects are unified using Spark-Meso clusters and writing applications in Python and MLib, and we are moving away from Postgree (dedicated hardware) and R (still research and fast development are done in R)

time series forecast and geo location targeting for brand advertising