Systolic-CNN: an OpenCL-defined scalable runtime-flexible programmable accelerator architecture for accelerating convolutional neural network inference in cloud/edge computing
View Patent ↗An OpenCL-defined scalable runtime-flexible programmable accelerator architecture for accelerating convolutional neural network (CNN) inference in cloud/edge computing is provided, referred to herein as Systolic-CNN. Existing OpenCL-defined programmable accelerators (e.g., field-programmable gate array (FPGA)-based accelerators) for CNN inference are insufficient due to limited flexibility for supporting multiple CNN models at runtime and poor scalability resulting in underutilized accelerator resources and limited computational parallelism. Systolic-CNN adopts a highly pipelined and paralleled one-dimensional (1-D) systolic array architecture, which efficiently explores both spatial and temporal parallelism for accelerating CNN inference on programmable accelerators (e.g., FPGAs). Systolic-CNN is highly scalable and parameterized, and can be easily adapted by users to efficiently utilize the coarse-grained computation resources for a given programmable accelerator. In addition, Systolic-CNN is runtime-flexible and can be time-shared to accelerate a variety of CNN models at runtime without the need to recompile the programmable accelerator kernel hardware or reprogram the programmable accelerator.
1 . A method for accelerating a convolutional neural network (CNN) process on a programmable accelerator, the method comprising:
establishing on the programmable accelerator a convolution layer and additional layers which are runtime-flexible for a plurality of CNN models without recompiling the programmable accelerator using optimized values of at least three architectural parameters selected such that the programmable accelerator utilizes up to 100% of available computing resources and reduces runtime relative to using non-optimized values of the at least three architectural parameters, wherein each of the at least three architectural parameters has a different impact on a demand for the available computing resources;
receiving a first request to perform a first CNN inference process; and
at runtime, accelerating the first CNN inference process using the convolution layer and the additional layers with spatial and temporal parallel execution,
wherein the at least three architectural parameters comprise: pe_num, reuse_fac, and vec_fac, wherein:
the pe_num parameter defines a number of processing elements (PEs) in a one-dimensional (1-D) systolic array of PEs that perform temporally paralleled convolution in a deep pipeline and a parallelism of output feature maps (OFMs);
the reuse_fac parameter defines a parallelism of inner product (IP) units inside each of the PEs and a number of times an identical input feature map (IFM) is reused by each of the PEs for convolution computation within an identical OFM; and
the vec_fac parameter defines a single instruction multiple data (SIMD) width of a partial computation between a weight vector and IFM vector across different channels inside each IP unit in each of the PEs.
2 . The method of claim 1 , further comprising:
receiving a second request to perform a second CNN inference process; and
accelerating the second CNN inference process using the convolution layer and the additional layers without recompiling the programmable accelerator.
3 . The method of claim 2 , wherein the first CNN inference process uses at least one of the additional layers which the second CNN inference process does not use.
4 . The method of claim 2 , wherein the second CNN inference process performs a type of inference which is distinct from the first CNN inference process.
5 . The method of claim 1 , wherein accelerating the first CNN inference process comprises executing data-independent loops on the convolution layer in parallel spatially.
6 . The method of claim 5 , wherein executing the data-independent loops of the convolutional layer in parallel spatially comprises executing the data-independent loops over different processing elements (PEs) of the programmable accelerator.
7 . The method of claim 5 , wherein accelerating the first CNN inference process comprises executing data-dependent loops of the convolutional layer in parallel temporally.
8 . The method of claim 5 , wherein the additional layers comprise two or more of a batch normalization (BNORM) layer, a local response normalization (LRN) layer, a max pooling layer, an average pooling layer, an element-wise sum (ELTWISE) layer, and a rectified linear unit (ReLU) layer.
9 . The method of claim 1 , wherein the programmable accelerator comprises a field-programmable gate array (FPGA).
10 . A deep learning system, comprising:
a programmable accelerator; and
a memory storing instructions which, when executed, cause the programmable accelerator to:
establish processing resources on the programmable accelerator which are runtime-flexible for a plurality of convolutional neural network (CNN) models using optimized values of at least three architectural parameters selected such that the programmable accelerator utilizes up to 100% of available computing resources and reduces runtime relative to using non-optimized values of the at least three architectural parameters, wherein each of the at least three architectural parameters has a different impact on a demand for the available computing resources;
receive a request to perform a CNN inference process using one of the plurality of CNN models; and
perform the CNN inference process with the processing resources without recompiling the programmable accelerator,
wherein the at least three architectural parameters comprise: pe_num, reuse_fac, and vec_fac, wherein:
the pe_num parameter defines a number of processing elements (PEs) in a one-dimensional (1-D) systolic array of PEs that perform temporally paralleled convolution in a deep pipeline and a parallelism of output feature maps (OFMs);
the reuse_fac parameter defines a parallelism of inner product (IP) units inside each of the PEs and a number of times an identical input feature map (IFM) is reused by each of the PEs for convolution computation within an identical OFM; and
the vec_fac parameter defines a single instruction multiple data (SIMD) width of a partial computation between a weight vector and IFM vector across different channels inside each IP unit in each of the PEs.
11 . The deep learning system of claim 10 , wherein the processing resources comprise a convolution layer and additional layers.
12 . The deep learning system of claim 10 , wherein the programmable accelerator comprises a field-programmable gate array (FPGA).
13 . The deep learning system of claim 10 , wherein the deep learning system further comprises:
an off-chip memory coupled to the programmable accelerator; and
a host processor configured to provide the request to perform the CNN inference process to the programmable accelerator.
14 . The deep learning system of claim 13 , wherein the processing resources comprise a shift register-based input feature map (IFM) buffer for storing IFM data received from the off-chip memory.
15 . A convolutional neural network (CNN) accelerator architecture, comprising:
a one-dimensional (1-D) systolic array of processing elements (PEs) configured to execute a convolution layer of a CNN; and
an additional layer module configured to provide optional computations for the CNN;
wherein the CNN accelerator architecture is configured to accelerate a plurality of types of CNNs on a programmable accelerator at runtime without reconfiguring the programmable accelerator using optimized values of at least three architectural parameters selected such that the programmable accelerator utilizes up to 100% of available computing resources and reduces runtime relative to using non-optimized values of the at least three architectural parameters, wherein each of the at least three architectural parameters has a different impact on a demand for the available computing resources,
wherein the at least three architectural parameters comprise: pe_num, reuse_fac, and vec_fac, wherein:
the pe_num parameter defines a number of PEs in the 1-D systolic array of PEs that perform temporally paralleled convolution in a deep pipeline and a parallelism of output feature maps (OFMs);
the reuse_fac parameter defines a parallelism of inner product (IP) units inside each of the PEs and a number of times an identical input feature map (IFM) is reused by each of the PEs for convolution computation within an identical OFM; and
the vec_fac parameter defines a single instruction multiple data (SIMD) width of a partial computation between a weight vector and IFM vector across different channels inside each IP unit in each of the PEs.
16 . The CNN accelerator architecture of claim 15 , wherein the additional layer module comprises one or more of a batch normalization (BNORM) layer, a local response normalization (LRN) layer, a max pooling layer, an average pooling layer, an element-wise sum (ELTWISE) layer, and a rectified linear unit (ReLU) layer.
17 . The CNN accelerator architecture of claim 15 , wherein the additional layer module comprises multiple computation layers which are selected at runtime according to a type of CNN selected.
18 . The CNN accelerator architecture of claim 17 , wherein the multiple computation layers are cascaded such that a computation layer not selected provides data passthrough.
19 . The CNN accelerator architecture of claim 15 , further comprising a buffer configured to store an input feature map (IFM) for the CNN received from an external memory.
20 . The CNN accelerator architecture of claim 19 , wherein the buffer is a shift register-based buffer.