20170113 7 th seminar Chap 1 Chap 2

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Literature surveys : Active noise control based on GPU 김 영 석 (Kim, Youngseok) System Dynamics and Applied Control Lab. Center for Noise and Vibration Control KAIST 2017. 01. 13 SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Brief comparison of CPU and GPU § Short comparison of CPU and GPU – CPU (Central Processing Unit) Sequence operation • Few cores optimized for sequential serial processing – GPU (Graphics Processing Unit) Parallel operation GPGPU (General Purpose GPU) • Massively parallel architecture consisting of thousands of smaller, more efficient cores • Understanding the efficiency of GPU algorithms for matrix-matrix multiplication, Stanford Univ. – FPU (Floating Point Unit) • Part of CPU and GPU Implemented as distinct chips • Specialized circuits for doing floating point operations (addition, subtraction, exponentiation, etc) Addition of three floating point numbers CPU computation Addition of million floating point numbers GPU computation SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Brief comparison of CPU and GPU SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Process of GPU computation § Conditions for real-time GPU implementation – Adaptive algorithm in block data processing Jorge Lorente Giner et al, Real-time adaptive algorithms using a Graphics Processing Units, 2012, Waves vol. 4 pp. 59 - 68 SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Process of GPU computation § Jorge Lorente Giner et al, Real-time adaptive algorithms using a Graphics Processing Units, 2012, Waves vol. 4 pp. 59 - 68 SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: GPU algorithm implementation § CUDA – A parallel computing platform and programming model invented by NVIDIA – Programmable with c, c++, Fortran, Python SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Literature surveys § Multi-channel active noise control – A high computational capacity is required • A number of channels, defining one channel for each pair of error sensor – secondary source • A number of filtering operations GPUs seem suitable for multi-channel ANC applications where the processing of each channel could be done in parallel – Literature surveys about cases of utilizing GPU for active noise control • GPU implementation of multi-channel adaptive algorithms for local active noise control, 2014 SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Literature surveys § Literature survey about cases of using GPU in active noise control – GPU implementation of multi-channel adaptive algorithms for local active noise control, 2014 • Processing each channel in parallel • Frequency-domain block-based filtered-x LMS GPU is better exploited when working with blocks of samples instead of sample-by-sample Most of the common audio cards work with block data buffers GPU ü Output signal generation ü Error signal calculation ü Update of the adaptive filters SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Literature surveys § Literature survey about cases of using GPU in active noise control – GPU implementation of multi-channel adaptive algorithms for local active noise control, 2014 • Processing each channel in parallel • Frequency-domain block-based filtered-x LMS GPU is better exploited when working with blocks of samples instead of sample-by-sample Most of the common audio cards work with block data buffers I : Reference signal J : Secondary sources K : Error sensors B : Size of the sample blocks L : Length of adaptive filters M : Length of estimated secondary path filter SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Idea of utilizing GPUs for active noise control § GPU d. B Sub-band filter bank Frequency domain Speaker saturation or coherence consideration Frequency [Hz] SDAC Lab. No. Vi. C KAIST

20170113 7 th seminar Chap 1: Chap 2: Chap 3: Chap 4: Future work § Studying programming language C § Studying GPU programming – GPU Programming and Its Applications [EE 817, Spring] § Derive an analytical solution of optimal solution considering a speaker saturation or coherence SDAC Lab. No. Vi. C KAIST