Implementing fundamental computational primitives using a matrix multiplication accelerator (MMA)
A method for performing a fundamental computational primitive in a device is provided, where the device includes a processor and a matrix multiplication accelerator (MMA). The method includes configuring a streaming engine in the device to stream data for the fundamental computational primitive from memory, configuring the MMA to format the data, and executing the fundamental computational primitive by the device.
Latest TEXAS INSTRUMENTS INCORPORATED Patents:
This application claims benefit of U.S. Provisional Patent Application Ser. No. 62/465,620, filed Mar. 1, 2017, which is incorporated herein by reference in its entirety.
BACKGROUNDApplications such as speech recognition, intelligent industrial control, object detection and recognition, and vision are increasingly being migrated to embedded devices. Hardware acceleration may be needed in such devices to support the computational needs of the algorithms used in such applications.
SUMMARYExamples of the present disclosure relate to implementing fundamental computational primitives using a matrix multiplication accelerator. In one aspect, a method for performing a fundamental computational primitive in a device is provided, where the device includes a processor and a matrix multiplication accelerator (MMA). The method includes configuring a streaming engine in the device to stream data for the fundamental computational primitive from memory, configuring the MMA to format the data, and executing the fundamental computational primitive by the device.
In one aspect, a device is provided that includes a memory, a processor coupled to the memory, and a matrix multiplication accelerator (MMA) coupled to the processor, the MMA including a multiplier buffer and a first multiplicand buffer, wherein the device is operable to configure a streaming engine in the device to stream data for a fundamental computational primitive from the memory, configure the MMA to format the data and execute the fundamental computational primitive.
Specific examples of the disclosure will now be described in detail with reference to the accompanying figures. Like elements in the various figures are denoted by like reference numerals for consistency.
Examples of the disclosure provide for implementing fundamental computational primitives used by applications such as speech recognition, intelligent industrial control, object detection and recognition, and vision using a matrix multiplication accelerator (MMA). The fundamental computational primitives include, for example, twodimensional (2D) convolution as used in convolutional neural networks (CNNs), small and large matrix matrix multiplication, matrix matrix point wise multiplication, matrix matrix addition, vector matrix multiplication, vector summation, affine transformation, fast Fourier transform, discrete cosign transform, convolution, correlation, matrix assignment, matrix permutation, and matrix transposition.
The MMA 104 includes sufficient memory to store two 32×32 multiplicand buffers 144 of 16bit elements for storing two B matrices and two 32×32 result buffers 134 of 16bit elements for storing two C matrices. The multiplicand buffers 144 may be referred to as B matrix buffers herein and the result buffers 134 may be referred to as C matrix buffers herein. The MMA 104 further includes memory to store a 1×32 multiplier buffer 138 of 16bit elements for storing a row of the multiplier matrix A. The multiplier buffer 138 may be referred to as the A matrix buffer herein. As is explained in more detail herein, the B matrix buffers 144 are used as ping pong buffers in some operations such that data is loaded into one of the buffers in background as data in the other buffer is used for operation execution. Similarly, the C matrix buffers 134 are used as foreground and background buffers such that, e.g., the results of operation execution are stored in one buffer while the contents of another buffer are output from the MMA 104.
On each cycle, the MMA 104 performs a single instruction, i.e., a Load, Store, and Execute instruction, referred to as the LSE instruction herein. As the name of this instruction implies, the MMA 104 can perform a load operation, a store operation, and an execute operation in a single cycle. In general, in a cycle, a vector of data is loaded into the A matrix buffer 138 and a matrix multiplication operation is performed between a B matrix stored in a selected B matrix buffer 144 and the data vector in the A matrix buffer 138. That is, the matrix product of the data vector in the A matrix buffer 138 with each column of the B matrix in the selected B matrix buffer 144 is computed. The result of the matrix multiplication operation is a row of data elements that is stored in a row of a C matrix in a selected C matrix buffer 134. Depending on the content of the fields of the LSE instruction, a cycle can also include loading a row of data in the B matrix buffer not being used for the matrix multiplication, i.e., the background B matrix buffer, storing a row of data from a C matrix buffer into external memory, and/or performing a specified operation on the results of the matrix product operations before storing the results in the selected C matrix buffer 134.
The load operation portion of the LSE instruction includes fields identifying the location in the buffer 124 of the data to be loaded into the A matrix buffer 138, the location in the buffer 128 of the data to be loaded in a B matrix buffer 144, the B matrix buffer 144 that is the target of the load operation, and the row in target B matrix buffer to be loaded. The load operation portion also includes a field for indicating whether a load operation is to be performed.
The store operation portion of the LSE instruction includes fields identifying the location in the buffer 118 where the data in a C matrix buffer 134 is to be stored, the C matrix buffer 134 holding the data to be stored, and the row in the target C matrix buffer 134 containing the data to be stored. The store operation portion also includes a field for indicating whether a store operation is to be performed.
The execute operation portion of the LSE instruction includes fields identifying the target C matrix buffer 134 and the row in the target C matrix buffer 134 that is to receive the result of the execute operation, and the operation to be performed with the results of the matrix multiplication before storing in the target C matrix buffer 134. The operations that can be specified include =, +=, −=, or none. The = operation causes the results to be directly stored in the specified row with no alteration. The += operation causes elements in the results to be added to corresponding elements in the specified row, with the results of the additions replacing the contents of the specified row. The −= operation causes elements in the results to be subtracted from corresponding elements in the specified row, with the results of the subtractions replacing the contents of the specified row. The none operation, as the name implies, indicated that no operation is to be performed. The none operation is used, for example, during the initial load of data into a B matrix buffer 144 prior to performing the matrix multiplication or when moving the final results stored in a C matrix buffer 134 out of the MMA 104.
The MMA 104 further includes configurable format components 130, 136, 140 for formatting, respectively, the data output by the MMA 104 and the data input to the MMA 104. The format A component 136 and the format B component 140 are configurable to format the respective input data according to a specified type, e.g., 16bit float, 16bit fixed signed, 16bit fixed unsigned, 8bit fixed signed, and 8bit fixed unsigned, and the Q point, i.e., the number of fractional bits, for fixed point inputs. The format C component 130 is configurable to format the output data according to a specified type, e.g., 16bit float, 16bit fixed signed, 16bit fixed unsigned, 8bit fixed signed, and 8bit fixed unsigned, and the Q point, i.e., the number of fractional bits, for fixed point outputs. The format A component 136 is further configurable to define a lookup table (LUT) that allows the A data in L2 106 to be stored in 4bit precision to save memory and expanded to 16bit precision in the A matrix buffer 138 using a mapping of 4 bits to 16 bits that doesn't need to be uniform. This is potentially useful for all computational primitives and is particularly useful for CNN style 2D convolution.
The MMA 104 also includes a row offset component 142 that is configurable to specify an offset for each element of a row of data to be loaded in a B matrix buffer 144. The row offset component 142 stores thirtytwo fivebit offset values, one for each of the thirtytwo elements in a row. The row offset values specified in the row offset component 142 can be used to place elements in a row of data elements being loaded into different rows in the background B matrix buffer different from the row number specified in the load portion of the LSE instruction. The offset value corresponding to a data element is added to the row number of the B matrix buffer specified in the load portion of the LSE instruction to determine the row of the B matrix buffer in which the data element will be loaded. The column number of the data element is not affected.
More specifically, on a cycle of the MMA 104, a new row of data can be loaded into the background B matrix buffer, i.e., the B matrix buffer 144 that is not being used for execution. If the row offset values in the row offset component 142 for all elements in the row of data are zero, the data elements will be loaded in the row of the background B matrix buffer specified in the LSE instruction for the cycle. For example, when loading a new row of data in the first row of the background B matrix buffer, the first element will be loaded in row 1, column 1, the second element will be loaded in row 1, column 2, etc. However, if a row offset value in the row offset component 142 is nonzero, the row in which the corresponding data element is loaded is determined by the row specified in the LSE instruction and the row offset value. For example, assume the row offset values are 0, 1, 2, . . . 31. When loading a new row of data in which the first row of the background B matrix buffer is specified in the LSE instruction, the first element will be loaded in row 1, column 1, the second element will be loaded in row 2, column 2, the third element will be loaded in row 3, column 3, etc., thus forming a diagonal in the background B matrix buffer.
The MMA 104 further includes a configurable nonlinearity component 132 for applying a nonlinearity to the output of a C matrix buffer 134. The nonlinearity implemented is a rectifying linear unit (ReLU) and, if activated, is applied to the output of a C matrix buffer 134 on an elementwise basis as follows: if the input to the nonlinearity component 132 is negative, set the output of the nonlinearity component 132 to zero, and if the input to the nonlinearity component 132 is nonnegative, set the output of the nonlinearity component 132 to the input of the nonlinearity.
In the example device 100, the processor 102 is a digital signal processor (DSP) that includes a level one data (L1D) cache memory 112, a level 2 (L2) unified instruction and data cache memory 106, and two streaming engines (SE0 and SE1) 108, 110. Examples of such a processor are described in U.S. Pat. No. 9,606,803, issued Mar. 28, 2017, which is incorporated by reference herein. Further, examples of streaming engines are described in U.S. Pat. No. 9,606,803 and United States Patent Application Publication 2017/0308381, published Oct. 26, 2017, which is incorporated by reference herein.
The processor 102 is configured to operate as a source of input data for the MMA 104 and to receive output data from the MMA 104. More specifically, the processor 102 is configured to receive data vectors for the MMA 104 from the streaming engines 108, 110 in respective register files 120, 126, to apply formatting 122 to the data as needed for the fundamental computational primitive being executed by the device 100, and store the data vectors in respective buffers 124, 128 for consumption by the MMA 104. The source A buffer 124 stores the data to be loaded into the A matrix buffer 138 and the source B buffer 128 stores the data to be loaded into a B matrix buffer 144.
Examples of input formatting 122 include zero padding, even/odd vector generation, value copying, known matrix creation, and linked operations. Even/odd vector generation receives two vectors. If the even option is selected, all even number elements of the two vectors are used to create the output vector for input to the MMA 104. If the odd option is selected, all odd number elements of the two vectors are used to create the output vector. The even/odd formatting is useful, for example, for fast Fourier transforms (FFTs) and convolutions using a stride greater than one. Value copying formatting generates a vector for input to the MMA 104 in which a scalar value read from L2 106 is replicated to all elements of the vector. Value copying is useful, for example, for bias creation. Known matrix creation formatting creates a sequence of output vectors for input to the MMA 104 that together form a common known matrix pattern, e.g., an identity matrix. Zero padding formatting adds zeros to a vector prior to input to the MMA 104. Linked operations take output vectors of the MMA 104 and provide the vectors as input to the MMA 104 for the A matrix buffer 138 or a B matrix buffer 144. Linked operations are useful for example, for Z=W*X*Y style operations.
The processor 102 is also configured to receive data vectors from the MMA 104 in the destination C buffer 118, to apply formatting 116 to the data as needed for the fundamental computational primitive being executed by the device 100, and to store the data vectors in the register file 114. The data vectors are stored into external memory (not shown) via the level one data cache 112 and the level two unified cache 106. Examples of output formatting 116 include seam removal, stride removal, zero padding, and matrix transpose.
The streaming engines 108, 110 are configured to transfer streams of data elements from the level two cache 106 to respective register files 120, 126. A stream is defined to be a sequence of elements of the same type and size. The streaming engines 120, 126 are programmable to define a stream specific to a fundamental computational primitive by specifying the following stream attributes: address of the first element of the stream, size and type of the elements in the stream, formatting for the data in the stream, and the address sequence associated with the stream, i.e., the addressing order in which to access the elements to place them in the stream. When a stream is opened, a streaming engine 120, 126 calculates the address, fetches the defined data type from L2, performs any specified data type formatting, e.g., zero extension or sign extension, maps the data into vectors, and delivers the data vectors directly to a respective register file 120, 126.
The addressing sequence of the streaming engines 108, 110 permits multidimensional memory accesses. That is, each streaming engine 120, 126 executes an address sequence for elements of a stream in terms of a pointer walking through memory. Each streaming engine 120, 126 implements a multiplelevel parameterized nested loop that controls the path the pointer takes. In this nested loop, an iteration count for a loop level indicates the number of times the loop at that level repeats and a dimension for a loop level defines the distance between pointer positions in the loop level.
The innermost loop, i.e., loop 0, consumes physically contiguous elements from memory and has an implicit dimension of one and the pointer moves from element to element in consecutive, increasing order in this loop level. In each level outside the inner loop, a loop moves the pointer to a new location based on the size of the dimension specified for the loop. This form of addressing allows programs to specify regular paths through memory in a small number of parameters
Table 1 shows example pseudo code for such a nested loop with six levels. In this pseudo code, ICNTx is the iteration count for level x, DIMx is the dimension for level x, and ELEM_BYTES is the size of each element in bytes. In other examples, the nested loop may have more or fewer levels.
The processor 102 also executes control software for each fundamental computational primitive defined for the device 100. The control software causes configuration of the streaming engines 108, 110 and the MMA 104 as needed to execute the fundamental computational primitive, controls the execution of the MMA, and causes the application of any input formatting 122 and/or output formatting 116 needed for the fundamental computational primitive.
The control software also configures 202 the MMA 104 as needed to perform the fundamental computational primitive using matrix multiplication. That is, the control software configures the format components 130, 136, 140, the row offset component 142, and the nonlinearity component 132 as needed for the fundamental computational primitive. Examples of configuring the MMA 104 for different computational primitives are described herein.
Once the configuration is complete, the control software starts 204 the configured streaming engines 108, 110 and executes 206 the configured fundamental computational primitive. In general, to execute the fundamental computational primitive, the control software causes the MMA 104 to execute a sequence of LSE instructions to load data elements into the A matrix buffer 138 and a background B matrix buffer 144, to execute the matrix multiplication between the A matrix buffer 138 and the foreground B matrix buffer 144, to store the result of the matrix multiplication in a selected C matrix buffer 134, and to move data from a background C matrix buffer 134 to the buffer 118. Note that any formatting and offsets configured in the MMA 104 are applied before data elements are loaded in the buffers 138, 144 and when results are moved from a C matrix buffer 134 to the buffer 118. As part of execution of the fundamental computational primitive, the control software may also cause input formatting 122 and output formatting 116 specific to the fundamental computational primitive to be performed on the processor 102.
To perform this primitive, the T H matrices are stored in the L2 cache 106 such that there are T*K contiguous elements containing the first row of each of the T H matrices followed by Z_{k }zeros, T*K contiguous elements containing the second row of each of the T H matrices followed by Z_{k }zeros, . . . , T*K contiguous elements of the Mth row of each of the H matrices followed by Z_{k }zeros, where Z_{K}=32−T*K. In addition, the T X matrices are stored the L2 cache 106 such that there are T*N contiguous elements containing the first row of each of the T X matrices followed by Z_{N }zeros, T*N contiguous elements containing the second row of each of the T X matrices followed by Z_{N }0s, . . . , T*N contiguous elements of the Mth row of each of the T X matrices followed by Z_{N }0s, where Z_{N}=32−T*N.
The streaming engine 110 is configured to read the elements of the T X matrices from the L2 cache 106 and provide vectors for loading in a B matrix buffer 144 of the MMA 104 that contain elements of successive rows of the T X matrices. The streaming engine 108 is configured to read the elements of the T H matrices from the L2 cache 106 and provide vectors for loading in the A matrix buffer 138 that contain elements of successive rows of the T H matrices.
The row offset component 140 of the MMA 104 is configured to cause the elements of the rows in each vector from the streaming engine 110 to be loaded at an offset t*K in a B matrix buffer 144. Thus, the elements of a row from X(0) are loaded with an offset of 0, the elements of the row from X(1) are loaded with an offset of K, the elements of the row from X(2) are loaded with an offset of 2K, etc.
To perform the multiplication, appropriately configured LSE instructions are executed on the MMA 104 to load a B matrix buffer 144 with an initial batch of X matrices. Once a B matrix buffer is loaded, further LSE instructions are executed to load the rows of the corresponding H matrices in the A matrix buffer 138, perform the multiplication, and store the results in a C matrix buffer 134. Further, if multiple batches are to be processed, the LSE instructions will also load another batch of X matrices in a background B matrix buffer 144 and move the results of a previous batch from a C matrix buffer 134 out of the MMA 104. Thus, to perform the batch small matrix matrix multiplication, T*K elements are loaded into the A matrix buffer 138 for M cycles, T*N elements are loaded into a B matrix buffer 144 (in background except for the initial batch) for K cycles, and T*N elements are moved out of a C matrix buffer for M cycles.
In other examples, rather than storing the Z_{K }and Z_{N }zeros in the L2 cache 106, the streaming engines 108, 110 or the input formatting 122 are configured to perform zero padding to add the required number of zeros to each vector prior to storing the vectors in the source A buffer 124 or the source B buffer 128.
Table 2 is example pseudo code illustrating performance of this primitive by the MMA 104. The pseudo code assumes that the streaming engine 108 is configured to read elements of the multiplier matrix H from the L2 cache 106 and to provide vectors to the A matrix buffer 138 such that each row of H tiles is loaded N times, i.e., H(0,0:(K−1)) is loaded N times, H(1,0:(K−1)) is loaded N times, H(2,0:(K−1)) is loaded N times, etc. That is, all rows of the H matrix are stored in the L2 cache 106 consecutively. The streaming engine 108 is configured to load the following sequence N times: H(0,0), H(0,1), . . . , H(0,K−1); then the following sequence N times: H(1,0), H(1,1), . . . , H(1,K−1); . . . ; then the following sequence N times: H(M−1,0), H(M−1,1), . . . , H(M−1, K−1).
The pseudo code also assumes that the streaming engine 110 is configured read element of X tiles form the L2 cache 106 and to provide each X tile M times to be loaded in a B matrix buffer 144, i.e., a sequence of loading [X(0:(K−1),0), . . . , X(0:(K−1), N−1)] is repeated M times. That is, all rows of the X matrix are stored in the L2 cache 106 consecutively. The streaming engine 110 is configured to load the following sequence N times: X(0,0), X(1,0), . . . , X(K−1,0), X(0,1), X(1,1), . . . , X(K−1,1), . . . , X(0,N−1), X(1,N−1), X(K−1, N−1).
In this pseudo code, Bback refers to the current background buffer of the B matrix buffers 144 and Bfore refers to the current foreground buffer used for execution.
To perform this primitive, the streaming engine 110 is configured read elements of the B matrix from the L2 cache 106 and to provide each row of the B matrix in turn for loading in a B matrix buffer 144 of the MMA 104. That is, the first vector from the streaming engine 110 will contain the first row, row 0, of the B matrix, the second vector from the streaming engine will contain the second row of the B matrix, etc. The streaming engine 108 is configured to read elements of the A matrix from the L2 cache 106 and to provide each row of the A matrix in turn for loading in the A matrix buffer 138. That is, the first vector from the streaming engine 108 will contain the first row, row 0, of the A matrix, the second vector from the streaming engine will contain the second row of the A matrix, etc.
The row offset component 140 of the MMA 104 is configured to cause the elements of a row of the B matrix to be loaded diagonally in a B matrix buffer 144. That is, the offsets for the row elements are set to sequential values ranging from 0 to 31, such that the first element of a row is loaded in row 0, column 0, the second element is loaded in row 1, column 1, the third element is loaded in row 2, column 2, etc.
To perform the point wise multiplication, appropriately configured LSE instructions are executed on the MMA 104 to load a B matrix buffer 144 with the initial row of the B matrix. Once a B matrix buffer is loaded, further LSE instructions are executed to load the corresponding row of the A matrix in the A matrix buffer 138, perform the matrix multiplication, and store the results in the corresponding row of a C matrix buffer 134. Further, the LSE instructions will also load the next row of the B matrix in the background B matrix buffer 144. This process of loading a row of the B matrix on the diagonal in the background B matrix, executing a matrix multiply on the foreground B matrix buffer, and storing the results is repeated until all rows of the B matrix have been processed. LSE instructions to move the contents of the C matrix buffer 134 out of the MMA 104 are then executed.
To perform this primitive, the streaming engine 108 is configured to read elements of A from the L2 cache 106 and to provide each row of A in turn to be loaded into the A matrix buffer 138. The input formatting 122 is configured to generate vectors of the identity matrix I to be loaded in to a B matrix buffer 144. Appropriately configured LSE instructions are executed in the MMA 104 to load each row of A in the A matrix buffer 138, perform the matrix multiplication between the row of A loaded in the A matrix buffer 138 and the identity matrix in a B matrix buffer 144, and store the results in corresponding locations of a C matrix buffer 134. The =operation is specified in the LSE instructions for storing the results in the C matrix buffer. Thus, each element of A is stored unchanged in a corresponding location in the C matrix buffer.
The streaming engine 108 is then configured to read elements of B from the L2 cache 106 and to provide each row of B in turn to be loaded into the A matrix buffer 138. Appropriately configured LSE instructions are executed in the MMA 104 to load each row of B in the A matrix buffer 138, perform the matrix multiplication between the row of B loaded in the A matrix buffer 138 and the identity matrix in a B matrix buffer 144, and store the results in corresponding locations of the C matrix buffer 134 storing the result of A*I. The +=operation is specified in the LSE instructions for storing the results in the C matrix buffer, thus causing the value of each data element of B to be added to the value of a corresponding element of A stored in the corresponding location of the C matrix buffer. LSE instructions to move the contents of the C matrix buffer 134 out of the MMA are then executed.
To perform this primitive, the streaming engine 108 is configured to read elements of T corresponding x vectors from the L2 cache 106 and to provide vectors for loading in the A matrix buffer 138. That is, the vectors from the streaming engine 108 contain the elements of x(0), . . . x(T−1). The loading of the vectors via the streaming engine 108 is similar to that described in reference to
The row offset component 140 of the MMA 104 is configured to cause the elements of the rows in each vector from the streaming engine 110 to be loaded at an offset t*K in a B matrix buffer 144. Thus, the elements of replicated row 0 of the H matrix are loaded with an offset of 0, the elements of replicated row 1 of the H matrix are loaded with an offset of K, the elements of replicated row 2 of the H matrix are loaded with an offset of 2K, etc.
To perform the multiplication, appropriately configured LSE instructions are executed on the MMA 104 to load a B matrix buffer 144 with an initial batch of replicated H matrices. Once a B matrix buffer is loaded, further LSE instructions are executed to load the vector of corresponding x vectors in the A matrix buffer 138, perform the multiplication, and store the results in a C matrix buffer 134. Further, if multiple batches are to be processed, appropriately configured LSE instructions will load another vector in the A matrix buffer 138, perform the multiplication, and store the previously computed vector. Note that the B matrix buffer 144 does not need to be reloaded for each new vector loaded in the A matrix buffer 138.
To perform this primitive, the streaming engine 108 is configured to read element of x vectors from the L2 cache 106 and provide vectors for loading in the A matrix buffer 138 that contain T×vectors. That is, the vectors from the streaming engine 108 contains the vectors x(0), . . . x(T−1). The loading of the vectors via the streaming engine 108 is similar to that described in reference to
To perform the summations, appropriately configured LSE instructions are executed on the MMA 104 to load a B matrix buffer 144 on the diagonal with multiple copies of a 1×K vector of all ones. Once a B matrix buffer is loaded, further LSE instructions are executed to load the x(0) . . . x(T−1) vectors in the A matrix buffer 138, perform the multiplication, and store the results in a row of a C matrix buffer 134. Further, if multiple batches are to be processed, appropriately configured LSE instructions will load another vector in the A matrix buffer 138, perform the multiplication, and store the previously computed vector. Note that the B matrix buffer 144 does not need to be reloaded for each new vector loaded in the A matrix buffer 138.
To perform this primitive, the streaming engine 110 is configured to read elements of the H matrix and v from the L2 cache 106 and provide vectors for loading in a B matrix buffer 144 of the MMA 104. To replicate the H matrix concatenated with v, K−1 rows of H are stored contiguously in the L2 cache 106 followed by one row of v and the input formatting 122 is configured to replicate this T times and add Zn zeros. Alternatively, T copies of each row of H are stored contiguously in the L2 cache 106 each followed by Zn zeros, then T copies of v are stored contiguously in the L2 cache 106 followed by Zn zeros.
The streaming engine 108 is configured to read elements of T corresponding x vectors from the L2 cache 106 and provide vectors for loading in the A matrix buffer 138. In one example, T copies of x(t) are stored contiguously in the L2 cache 106, and the input formatting 122 is configured to add a one between copies and Zk zeros at the end. Alternatively, T copies of x(t) are stored contiguously in the L2 cache 106 with a one between copies and Zk zeros at the end.
The row offset component 140 of the MMA 104 is configured to cause elements in each vector from the streaming engine 110 to be loaded at appropriate offsets in a B matrix buffer 144. For example, assuming that T copies of each row of H are stored contiguously in the L2 cache 106 each followed by Zn zeros, then T copies of v are stored contiguously in the L2 cache 106 followed by Zn zeros, the row offset component 140 is configured such that elements of replicated row 0 of the [H;v] matrix are loaded with an offset of 0, the elements of replicated row 1 of the [H;v] matrix are loaded with an offset of K, the elements of replicated row 2 of the [H;v] matrix are loaded with an offset of 2K, etc.
To perform the multiplication, appropriately configured LSE instructions are executed on the MMA 104 to load a B matrix buffer 144 with an initial batch of replicated [H;v] matrices. Once a B matrix buffer is loaded, further LSE instructions are executed to load the vector of corresponding x vectors in the A matrix buffer 138, perform the multiplication, and store the results in a C matrix buffer 134. Further, if multiple batches are to be processed, appropriately configured LSE instructions will load another vector in the A matrix buffer 138, perform the multiplication, and store the previously computed vector. Note that the B matrix buffer 144 does not need to be reloaded for each new vector loaded in the A matrix buffer 138.
To compute the output feature maps 1004, the filters or kernels can be read in contiguous 32×32 blocks, applied to 32×32 blocks of the tile filtering matrix X 1006 in the MMA 104, and the results of applying the filters to the input feature maps 1000 can be computed separately by the MMA 104 for 32×32 blocks of the tile filtering matrix 1006 and combined to generate the output feature maps 1004. The rows of the filters or kernels 1002 can be stored in the L2 cache 106 consecutively and the filter matrix H 1008 can be read from the L2 cache 106 in contiguous 32×32 blocks. As illustrated in
Y(:,n)=H*Xfilter(:,n),n=0, . . . ,N−1
or in block row order (shown in this example) as per
Y(m,:)=H(m,:)*Xfilter,m=0, . . . ,M−1
where Xfilter is the tile filtering matrix.
To perform this primitive when large feature map tiles are used, the streaming engine 108 is configured to read filter vectors of the filter matrix H 1008 from the L2 cache 106 in block row order for loading in the A matrix buffer 138 of the MMA 104. The streaming engine 110 is configured to read elements of row of feature maps form the L2 cache 106 and to provide each row in turn of 32×32 tiles in the feature maps for loading in a B matrix buffer 144 of the MMA 104. Table 3 is an example of pseudo code for configuring the streaming engine 108 and Table 4 is an example of pseudo code for configuring the streaming engine 110. In this pseudo code, Bf is the number of bytes per feature map data element, S is the stride, VECLEN is the vector length, and the tile size is Lr×Lc×Ni where Lc=32*Bc+Fc−1 and Bc is a positive integer indicating the number of valid input feature map tile columns. The value of Bc can be chosen, for example, as the largest possible integer such that 32*Bc+Fc−1 the number of columns in the feature map. Smaller values of Bc can also be used.
In the pseudo code of Table 3, the values of ICNTx and DIMx are chosen such that the filter matrix H is generated in block row order as previously specified. The rows of the H matrix are either stored with a zero pad in the L2 cache 106 or the streaming engine 108 or the input formatting 122 is configured to add the zero padding. In the pseudo code of Table 3, the values of ICNTx and DIMs are chosen such that the filtering matrix X is generated in block column order as previously specified. The for loop indicates that the whole pattern repeats ceil(No/32) times.
The pattern in the pseudo code of Table 3 and Table 4 creates a large matrix matrix multiplication problem, i.e., from the perspective of the MMA 104, the processing performed is a large matrix matrix multiplication problem that is performed as previously described in reference to
To perform this primitive when small feature map tiles are used, the streaming engine 108 is configured to read filter vectors of the filter matrix H 1008 in block row order from the L2 cache 106 for loading in the A matrix buffer 138 of the MMA 104. The streaming engine 110 is configured to read elements of feature maps from the L2 cache 106 and to provide each row in turn of 32×32 tiles in the feature maps for loading in a B matrix buffer 144 of the MMA 104. Table 5 is an example of pseudo code for configuring the streaming engine 108 and Table 6 is an example of pseudo code for configuring the streaming engine 110. In this pseudo code, Bf is the number of bytes per feature map data element, S is the stride, VECLEN is the vector length, the tile size is Lr×Lc×Ni where Tc (32+Fc−1)/Bc and Bc is a positive integer indicating the number of valid input feature map tile columns, and M=max(Bc,S). The value of Bc is the maximum integer that satisfies Bc≤(32+Fc−1)/Lc.
In the pseudo code of Table 5, the values of ICNTx and DIMx are chosen such that the filter matrix H is generated in block row order as previously specified. The rows of the H matrix are either stored with a zero pad in the L2 cache 106 or the streaming engine 108 or the input formatting 122 is configured to add the zero padding. In the pseudo code of Table 6, the values of ICNTx and DIMs are chosen such that the filtering matrix X is generated in block column order as previously specified. The for loop indicates that the whole pattern repeats ceil(No/32) times.
The pattern in the pseudo code of Table 5 and Table 6 creates a large matrix matrix multiplication problem, i.e., from the perspective of the MMA 104, the processing performed is a large matrix matrix multiplication problem that is performed as previously described in reference to
When large feature tiles are used, if the stride S is greater than one, the configuration of the streaming engine 108 handles the row stride. The column stride is handled by configuring the output formatting 116 to write out one of every S columns generated by the MMA 104. When small feature tiles are used, seams need to be removed from the output of the MMA 104. If no stride S is specified, then the output formatting 116 is configured to remove Bc−1 seams of size Fc−1 from the output and to remove the last 32−Bc*Lc+Fc−1 samples from the output and ignore the end samples if (Lr−Fr+1)/Bc is not an integer. If a stride S is specified, the output formatting 116 is configured to perform seam removal on the output and to remove (S−1)/S rows and columns from the output. In some examples, when zero padding is needed for the next convolution layer, the output formatting 116 is configured to add the zero padding to the output of the MMA 104.
To perform this primitive, the streaming engine 110 is configured to provide each row of the B matrix in turn from the L2 cache 106 for loading in a B matrix buffer 144 of the MMA 104. That is, the first vector from the streaming engine 110 will contain the first row, row 0, of the B matrix, the second vector from the streaming engine 110 will contain the second row of the B matrix, etc. The streaming engine 108 is configured to provide each row of the P matrix in turn from the L2 cache 106 for loading in the A matrix buffer 138. That is, the first vector from the streaming engine 108 will contain the first row, row 0, of the P matrix, the second vector from the streaming engine will contain the second row of the A matrix, etc.
To perform the row permutation, appropriately configured LSE instructions are executed on the MMA 104 to load a B matrix buffer 144 with the B matrix. Once a B matrix buffer is loaded, further LSE instructions are executed to load rows of the P matrix in the A matrix buffer 138, perform the matrix multiplications, and store the results in corresponding rows of a C matrix buffer 134. LSE instructions to move the contents of the C matrix buffer 134 out of the MMA 104 are executed once all matrix multiplications have been completed.
To perform this primitive, the streaming engine 110 is configured to provide each row of the P matrix in turn from the L2 cache 106 for loading in a B matrix buffer 144 of the MMA 104. That is, the first vector from the streaming engine 110 will contain the first row, row 0, of the P matrix, the second vector from the streaming engine will contain the second row of the P matrix, etc. The streaming engine 108 is configured to provide the vector c from the L2 cache 106 for loading in the A matrix buffer 138.
To perform the column permutation, appropriately configured LSE instructions are executed on the MMA 104 to load a B matrix buffer 144 with the P matrix. Once a B matrix buffer is loaded, further LSE instructions are executed to load the c vector in the A matrix buffer 138, perform the matrix multiplications, and store the results in a corresponding row of a C matrix buffer 134. LSE instructions to move the contents of the C matrix buffer 134 out of the MMA 104 are executed once all matrix multiplications have been completed.
The device 100 can also be configured to perform batch small onedimensional (1D) complex to complex fast Fourier transforms (FFTs) with interleaved real and imaginary parts and with separate real and imaginary parts and large 1D complex to complex FFTs with interleaved real and imaginary parts and with separate real and imaginary parts. In the following FFT descriptions, the subscript “co” denotes complex, the subscript “re” denotes real, and the subscript “im” denotes imaginary.
A small complex to complex FFT of length N 16 with interleaved real and imaginary parts can be computed as a brute force discrete Fourier transform (DFT) as per
y_{co}^{N}=x_{co}^{N}*F_{co}^{N }
where
As used herein, a brute force DFT is using a DFT matrix to multiply an input vector to create an output vector that is the DFT of the input.
For N=16, vector matrix multiplication can be used to compute y_{co}^{16}=x_{co}^{16}*F_{co}^{16 }as x_{co}^{16 }and y_{co}^{16 }are 1×32 vectors with interleaved real and imaginary values and F_{co}^{16 }is a 32×32 matrix. That is, a vector x_{co}^{16 }can be loaded into the A matrix buffer 138 of the MMA 104 and the matrix F_{co}^{16 }can be loaded in a B matrix buffer 144, the matrix product computed, and the results stored in a C matrix buffer 134. For smaller values of N, small vector matrix multiplication as previously described herein in reference to
Batches of small complex to complex FFTs with separate real and imaginary parts can be computed as brute force discrete Fourier transforms (DFT) as per
Y_{re}^{M,N}=X_{re}^{M,N}*F_{re}^{N}−X_{im}^{M,N}*F_{im}^{N }
Y_{im}^{M,N}=X_{im}^{M,N}*F_{re}^{N}+X_{re}^{M,N}*F_{im}^{N }
where
where
x_{re}^{m,N}=[x_{re}^{m,N}(0) . . . x_{re}^{m,N}(N−1)]
x_{im}^{m,N}=[x_{im}^{m,N}(0) . . . x_{im}^{m,N}(N−1)]
y_{re}^{N}=[y_{re}^{N}(0) . . . y_{re}^{N}(N−1)].
Batches of M 32point complex to complex FFTs with separate real and imaginary parts can be computed in the device 100 using matrix matrix multiplication by splitting the computation of Y_{re}^{M,32 }and Y_{im}^{M,32 }into two computations as per
Y_{re}^{M,32}=X_{re}^{M,32}*F_{re}^{32 }
Y_{re}^{M,32}−=X_{im}^{M,32}*F_{im}^{32 }
Y_{im}^{M,32}=X_{im}^{M,32}*F_{re}^{32 }
Y_{im}^{M,32}+=X_{re}^{M,32}*F_{im}^{32}.
To perform these computations, the streaming engines 108, 110 and the MMA 104 are configured for matrix matrix multiplication. Further, the input formatting 122 is configured to split the inputs into the real and imaginary parts.
Smaller complex to complex FFTs with separate real and imaginary parts can be computed in the device 100 in batches of batches using the above formulation and batch small matrix matrix multiplication as previously described herein where N=K=FFT size 32.
A large complex to complex 1D FFT can be computed as follows: 1) load the data X in contiguous order in rows of a matrix; 2) FFT the columns of the matrix; 3) point wise multiply the matrix by the twiddle factors e^{j*(2*π/N)*row*col }for IFFT or e^{−j*(2*π/N)*row*col }for FFT; 4) FFT the rows of the matrix; 5) transpose the matrix; and 6) store the rows of the resulting matrix in contiguous order.
This processing can be implemented in the device 100 as follows given a data matrix X. To FFT the columns of X, load the matrix X in a B matrix buffer 144 and load the DFT matrix F in the A matrix buffer 138. Then use matrix matrix point wise multiplication as described herein to perform the multiplication with the twiddle factors. To FFT the rows of X, load the DFT matrix F in a B matrix buffer 144 and the matrix X in the A matrix buffer 138. Note that loading the data matrix X to the A matrix buffer or a B matrix buffer to allow the FFT of columns or rows without transpose. The final transpose of the matrix X can be performed in the MMA 104 or the output formatting 116.
Table 7 is an example illustrating computation of a complex to complex 1D FFT of length 1024 with separate real and imaginary parts. Note that T_{re}^{32}=F_{re}^{32 }and T_{im}^{32}=F_{im}^{32}. Also, all load operations explicitly mentioned are into a B matrix buffer 144, and with the exception of the initialization load, are performed in parallel with the matrix multiplications. The input formatting 122 is configured to split the inputs into the real and imaginary parts.
Initialization

 Load X_{re}^{32,32 }
1. FFT columns
Y_{re}^{32,32}=F_{re}^{32}*X_{re}^{32,32 }and load X_{im}^{32,32 }
Y_{re}^{32,32}−=F_{im}^{32}*X_{im}^{32,32 }
Y_{im}^{32,32}=F_{re}^{32}*X_{im}^{32,32 }and load X_{re}^{32,32 }
Y_{im}^{32,32}+=F_{im}^{32}*X_{re}^{32,32 }and load Y_{re}^{32,32 }
2. Point wise multiply by twiddle factors
X_{re}^{32,32}=T_{re}^{32}·*Y_{re}^{32,32 }and load Y_{im}^{32,32 }
X_{re}^{32,32}−=T_{im}^{32}·*Y_{im}^{32,32 }
X_{im}^{32,32}=T_{re}^{32}·*Y_{im}^{32,32 }and load Y_{re}^{32,32 }
X_{im}^{32,32}+=T_{im}^{32}·*Y_{re}^{32,32 }and load F_{re}^{32 }
3. FFT rows
Y_{re}^{32,32}=X_{re}^{32,32}*F_{re}^{32 }and load F_{im}^{32 }
Y_{im}^{32,32}−=X_{im}^{32,32}*F_{im}^{32 }
Y_{im}^{32,32}=X_{re}^{32,32}*F_{im}^{32 }and load F_{re}^{32 }
Y_{im}^{32,32}+=X_{im}^{32,32}*F_{re}^{32 }
4. Transpose
X_{re}^{32,32}=(Y_{re}^{32,32})^{T }
X_{im}^{32,32}=(Y_{im}^{32,32})^{T }
The device 100 can also be configured to perform onedimensional (1D) discrete cosign transforms (DCTs) as per y=dct(x). The computation of batch small onedimensional (1D) discrete cosign transforms (DCTs) and large 1D DCTs is similar to the previously described small and large 1D complex to complex FFTs except that the DCT matrix is real and the data is real. The terms small and large in reference to DCT sizes are used identically as to the terms small and large FFT sizes. Small DCTs can be implemented via matrix vector multiplication. Large DCTs can be implemented via preprocessing (step 1), FFT (step 2) and post processing (step 3) as per:
x′(n)=x(2*n) and x′(N−1−n)=x(2*n−1) for n=0,1, . . . ,N/2−1 (step 1)
Compute y′=FFT(x′) (step 2)
y(n)=real(exp(−j*n*pi/(2*N))*y(n)) (step 3)
Any FFT method can be used to compute step 2 in the above sequence, including the previously described small FFT and large FFT.
The device 100 can also be configured to perform small and large onedimensional (1D) convolution/filtering as per y=hx. 1D convolution/filtering may be computed as per
y(n)=Σh(τ)*x(n−τ),τ=0, . . . ,L−1;n=L=1,L, . . .
where L is the filter length and τ is the filter coefficient index as per the definition of convolution. A small convolution is one in which L≤32 and a large convolution is one in which L>32. For small 1D convolution, assume the filter length L is in [1, 32] and the output is valid for n≥L−1. Small 1D convolution may be computed in the device 100 as per
C(t,:)=A(t,:)*B,t=0,1, . . .
where
Note that in B the last L−1 columns of all rows are zeroes and in C the last L−1 entries are zeroes and the variable t is used to index the number of times that 32−L outputs are generated.
To perform the 1D convolution, the streaming engine 110 is configured to provide elements of the filter h from the L2 cache 106 for initially loading a B matrix buffer 144 with 33L copies of the filter. Either multiple copies of h are stored in the L2 cache 106 or a single copy is stored in the L2 cache 106 and replicated by configuring the input formatting 122. The streaming engine 108 is configured to read 33−L new values of x(n) from the L2 cache 106 on each cycle and generate A(t,:). The MMA 104 is configured for vector matrix multiplication, and, on each cycle, the MMA 104 computes 33−L values of y(n) in C(t,:).
The strategy for computing large 1D convolutions/filters in the device 100 is based on the following observations regarding decomposition of a large filter into two smaller filter fragments:
where τ=0, . . . , L−1; n=L−1, L, . . . and τ_{1}=0, . . . , L_{1}−1; τ_{2}=0, . . . , L_{2}−1.
The filter h is broken into smaller fragments using a recursive application of the above decomposition until the sizes of the filter fragments are as optimal as possible, i.e., the filter sizes are chosen as a tradeoff between maximizing the throughput of the MMA 104 and minimizing the total number of filters. The graph of
The device 100 can also be configured to perform 1D correlation as per y=h*x. Correlation is the same as convolution with the filter coefficients in time reversed order. Thus, the 1D correlation can be implemented using small or large 1D convolution as previously described in which the input formatting 122 is configured to perform the time reversal on either h or x.
The device 100 can also be configured to implement matrix assignment as per C=A or C=B. That is, assignment can be performed in the MMA 104 as C=A*I or C=I*B where I is the identity matrix.
The device 100 can also be configured to implement matrix transpose BT. The transpose of a matrix is a new matrix in which the columns are the rows of the original matrix. To perform the transpose operation, the MMA 104 includes an operation mode, i.e., a copy operation, that specifies that elements in the A matrix buffer 138 are to be copied to a designated column of a B matrix buffer 144.
To perform the transpose of a matrix, the streaming engine 108 is configured to read the elements of rows of the matrix from the L2 cache 106 and provide each row of the matrix in turn for loading in a B matrix buffer 144. That is, the first vector from the streaming engine 108 will contain the first row, row 0, of the matrix, the second vector from the streaming engine will contain the second row of the matrix, etc. Appropriately configured LSE instructions that load the rows of the matrix from the buffer 124 and store the elements in successive columns of a B matrix buffer 144 are then executed. Once all rows of the matrix have been transposed, appropriately configured LSE instructions are executed to move the contents of the B matrix buffer to a C matrix buffer 134 and out of the MMA 104.
Other ExamplesWhile the disclosure has been described with respect to a limited number of examples, other examples can be devised which do not depart from the scope of the disclosure as disclosed herein.
Examples have been described herein in which the processor is a DSP and executes the control software that configures the device 100 to perform a fundamental computational primitive and controls the execution of the primitive. In other examples, a host processor coupled to the device 100 performs some or all of the control operations needed to execute the fundamental computational primitives. The processor may be any suitable type and combination of electronic data processors. For example, the processor can be one or more processors from Intel Corp. or Advanced Micro Devices, Inc., one or more Reduced Instruction Set Computers (RISC), one or more ApplicationSpecific Integrated Circuits (ASIC), one or more digital signal processors (DSP), or the like.
Examples have been described herein in which buffer dimensions in the MMA are 1×32 or 32×32. In other examples, the buffer dimensions in the MMA can be smaller or larger.
It is therefore contemplated that the appended claims will cover any such modifications of the examples as fall within the true scope of the disclosure.
Claims
1. A method comprising:
 configuring, by a processor, a streaming engine, based on a computational primitive, to stream data vectors, wherein the data vectors include data elements;
 formatting, by the processor, the data vectors based on the computational primitive;
 configuring, by the processor, a matrix multiplication accelerator (MMA) based on the computational primitive; and
 executing, by the MMA, the computational primitive to produce an MMA output.
2. The method of claim 1, wherein executing the computational primitive further comprises loading data into a first multiplicand matrix buffer concurrently with execution of matrix multiplication using data in a second multiplicand matrix buffer, the first and second multiplicand matrix buffers comprised in the MMA.
3. The method of claim 1, wherein configuring the MMA further comprises configuring the MMA to offset load data in a multiplicand matrix buffer comprised in the MMA.
4. The method of claim 1, wherein the data comprises a row of a matrix, and wherein executing the computational primitive further comprises copying the data to a column in a multiplicand matrix buffer comprised in the MMA.
5. The method of claim 1, further comprising formatting, by the processor, the MMA output to add zero padding to the MMA output.
6. The method of claim 1, further comprising formatting, by the processor, the MMA output to perform seam removal on the MMA output.
7. The method of claim 1, further comprising formatting, by the processor, the MMA output to perform column subsampling on the MMA output according to a specified stride.
8. The method of claim 1, wherein configuring the streaming engine further comprises configuring the streaming engine to tile data of a plurality of feature maps into a filtering matrix.
9. The method of claim 8, wherein a size of a tile of data is selected such that no seam is present in the output of the MMA.
10. The method of claim 8, wherein a size of a tile of data is selected such that a seam is always in a same location in the output of the MMA.
11. The method of claim 1, wherein the data vectors include interleaved real and imaginary values; and
 wherein the MMA is configured to perform matrix multiplication on the data vectors.
12. The method of claim 1, wherein the data vectors comprise a filter of a length greater than a width of a multiplicand matrix buffer comprised in the MMA, and the filter is decomposed into a plurality of smaller filters, wherein sizes of the smaller filters are chosen as a tradeoff between maximizing the throughput of the MMA and minimizing a total number of smaller filters.
13. A device comprising:
 a memory;
 a processor coupled to the memory; and
 a matrix multiplication accelerator (MMA) coupled to the processor, the MMA comprising a multiplier buffer and a first multiplicand buffer,
 wherein the processor is operable to: configure a streaming engine comprised in the device to stream data vectors based on a computational primitive, wherein the data vectors include data elements; format the data vectors based on the computational primitive; and configure the MMA based on the computational primitive; and
 wherein the MMA is operable to execute the computational primitive to produce an MMA output.
14. The device of claim 13, wherein the device is further operable to execute the computational primitive wherein data is loaded into the first multiplicand matrix buffer concurrently with execution of matrix multiplication using data in a second multiplicand matrix buffer comprised in the MMA.
15. The device of claim 13, wherein the device is further operable to configure the MMA to offset load data in the first multiplicand matrix buffer.
16. The device of claim 13, wherein the device is further operable to execute the computational primitive, wherein the data comprises a row of a matrix, and the data is copied to a column in the first multiplicand matrix buffer.
17. The device of claim 13, wherein the device is further operable to configure the processor to format the MMA output to add zero padding to the MMA output.
18. The device of claim 13, wherein the device is further operable to configure the processor to format the MMA output to perform seam removal on the MMA output.
19. The device of claim 13, wherein the device is further operable to configure the processor to format the MMA output to perform column sub sampling on the MMA output according to a specified stride.
20. The device of claim 13, wherein the device is further operable to configure the streaming engine to tile data of a plurality of feature maps into a filtering matrix.
21. The device of claim 20, wherein a size of a tile of data is selected such that no seam is present in output of the MMA.
22. The device of claim 20, wherein a size of a tile of data is selected such that a seam is always in a same location in output of the MMA.
23. The device of claim 13, wherein the data vectors include interleaved real and imaginary values; and
 wherein the MMA is configured to perform matrix multiplication on the data vectors.
24. The device of claim 13, wherein the data vectors comprise a filter of a length greater than a width of the first multiplicand matrix buffer, and the device is further operable to decompose the filter into a plurality of smaller filters, wherein sizes of the smaller filters are chosen as a tradeoff between maximizing the throughput of the MMA and minimizing a total number of the smaller filters.
4897717  January 30, 1990  Hamilton et al. 
5099447  March 24, 1992  Myszewski 
5745793  April 28, 1998  Atsatt et al. 
5870568  February 9, 1999  Culley et al. 
5982375  November 9, 1999  Nelson et al. 
8301749  October 30, 2012  Finklestein et al. 
9606803  March 28, 2017  Anderson et al. 
9645974  May 9, 2017  Patil et al. 
10114613  October 30, 2018  Bekas 
20020191692  December 19, 2002  Fallon et al. 
20040136316  July 15, 2004  Kwak et al. 
20090292758  November 26, 2009  Brokenshire et al. 
20090300091  December 3, 2009  Brokenshire et al. 
20110153707  June 23, 2011  Ginzburg et al. 
20120011348  January 12, 2012  Eichenberger et al. 
20120072917  March 22, 2012  Boldyrev et al. 
20120140814  June 7, 2012  Sole Rojals et al. 
20120251013  October 4, 2012  Porikli 
20120268298  October 25, 2012  Oh 
20130262548  October 3, 2013  Ge et al. 
20140167987  June 19, 2014  Pell et al. 
20140365548  December 11, 2014  Mortensen 
20160373158  December 22, 2016  Ardalan 
20170139710  May 18, 2017  Zbiciak 
20170153890  June 1, 2017  Anderson et al. 
20170153959  June 1, 2017  Zbiciak 
20170168898  June 15, 2017  Zbiciak et al. 
20170249150  August 31, 2017  Zbiciak et al. 
20170308381  October 26, 2017  Zbiciak 
20190266218  August 29, 2019  Scott et al. 
20200051551  February 13, 2020  Sainath 
 International Search Report and Written Opinion dated Jan. 31, 2019, International Application No. PCT/US2018/019746, 7 pages.
 International Search Report and Written Opinion dated Aug. 9, 2018, International Application No. PCT/US2018/020462, 5 pages.
 International Search Report and Written Opinion dated Aug. 23, 2018, International Application No. PCT/US2018/020283, 7 pages.
 Notification of Transmittal odf the International Search Report and te Written Opinion of the International Searching Authority, or the Declaration; PCT/US 2018/020334; dated Jun. 7, 2018; 12 pages.
 Extended European Search Report; App. Pat.No. 18760733.81231/3590048 PCT/US2018020462; dated Feb. 4, 2020; 8 pages.
 Utku Aydoinat, et al.“An Open CL(TM) Deep Learning Accelerator on Arria 10”, arxiv.org,Cornell University Library, 201 Olin Library, Cornell University Ithaca, NY 14853, Jan. 13, 2017, XP080741254.
 Hussain Tassadaq, et al.“PPMC: A Program Pattern Based Mwmory Controller”, Mar. 19, 2012, International Conference on Financial Cryptography and Data Security; [Lecture Notes in Computer Science; Lect. Noptes Computer], Springer, Berlin,Heidelberg, pp. 89101, XP047379471, ISBN: 9783642173189.
 Chen Zhang et al.“Optimizing FPGAbased Accelerator Design for Deep Convolutional Neural Networks”, Proceedings of the 2015 ACM/SIGDA International Symposium of FieldProgrammable Gate Arrays, FPGA, Feb. 15, 22, 2015, pp. 161170, XP055265150, New, NY, USA, DOI: 10.1145/2684746.2689060, ISBN:978450333153.
Type: Grant
Filed: Feb 28, 2018
Date of Patent: Aug 10, 2021
Patent Publication Number: 20180253402
Assignee: TEXAS INSTRUMENTS INCORPORATED (Dallas, TX)
Inventors: Arthur John Redfern (Plano, TX), Timothy David Anderson (University Park, TX), Kai Chirca (Dallas, TX), Chenchi Luo (Plano, TX), Zhenhua Yu (Santa Clara, CA)
Primary Examiner: Michael D. Yaary
Application Number: 15/907,356