Methods and apparatus to perform weight and activation compression and decompression
Methods, apparatus, systems, and articles of manufacture to perform weight and activation compression and decompression are disclosed. An example apparatus includes memory, instructions in the apparatus, and processor circuitry to execute the instructions to execute a compression operation to obtain compressed data corresponding to weights in a weight matrix, and determine meta-data associated with the weight matrix, a first portion of the meta-data indicative of whether the weight matrix is compressed, a second portion of the meta-data indicative of a cache size of the compressed data, and a third portion of the meta-data indicative of the compression operation executed to obtain the compressed data.
1. An apparatus comprising:
memory;
instructions in the apparatus; and
processor circuitry to execute the instructions to:
execute a compression operation on a first portion of a matrix of weights for a neural network to obtain first compressed data, the first portion of the matrix is a sparse portion;
execute the compression operation on a second portion of the matrix of weights for a neural network to obtain second compressed data, the second portion of the matrix is a sparse portion;
determine first meta-data associated with the first compressed data and second meta-data associated with the second compressed data, a first portion of the first meta-data indicative of whether the first compressed data is compressed, a second portion of the first meta-data indicative of a cache size of the first compressed data, and a third portion of the first meta-data indicative of the compression operation executed to obtain the first compressed data;
cause the first meta-data and the second meta-data to be stored contiguously in a first row of a memory;
cause the first compressed data to be stored in a second row in the memory; and
cause the second compressed data to be stored in a third row in the memory, wherein the second row and the third row are consecutive rows.
2. The apparatus of claim 1 , wherein the first meta-data is a byte.
3. The apparatus of claim 1 , wherein the first portion of the first meta-data is of a first size, the third portion of the first meta-data is of a second size larger than the first size, and the second portion of the first meta-data is of a third size larger than the second size.
4. The apparatus of claim 1 , wherein the first portion of the first meta-data is a bit, wherein the processor circuitry is to:
assign a first value to the bit in response to the first compressed data being compressed; and
assign a second value to the bit in response to the first compressed data being uncompressed.
5. The apparatus of claim 1 , wherein the processor circuitry is to record a quantity of cache lines occupied by the first compressed data in the second portion of the first meta-data.
6. The apparatus of claim 1 , wherein the processor circuitry is to:
pack non-zero weights from the matrix of weights into a compressed tile; and
generate a bitmap indicative of respective locations of the non-zero weights in the matrix of weights.
7. A non-transitory machine readable medium comprising instructions which, when executed, cause one or more processors to:
execute a compression operation on a first portion of a matrix of weights for a neural network to obtain first compressed data, the first portion of the matrix is a sparse portion;
execute the compression operation on a second portion of the matrix of weights for a neural network to obtain second compressed data, the second portion of the matrix is a sparse portion;
determine first meta-data associated with the first compressed data and second meta-data associated with the second compressed data, a first portion of the first meta-data indicative of whether the first compressed data is compressed, a second portion of the first meta-data indicative of a cache size of the first compressed data, and a third portion of the first meta-data indicative of the compression operation executed to obtain the first compressed data;
cause the first meta-data and the second meta-data to be stored contiguously in a first row of a memory;
cause the first compressed data to be stored in a second row in the memory; and
cause the second compressed data to be stored in a third row in the memory, wherein the second row and the third row are consecutive rows.
8. The non-transitory machine readable medium of claim 7 , wherein the first meta-data is a byte.
9. The non-transitory machine readable medium of claim 7 , wherein the first portion of the first meta-data is of a first size, the third portion of the first meta-data is of a second size larger than the first size, and the second portion of the first meta-data is of a third size larger than the second size.
10. The non-transitory machine readable medium of claim 7 , wherein the first portion of the first meta-data is a bit, wherein the instructions, when executed, cause the one or more processors to:
assign a first value to the bit in response to the first compressed data being compressed; and
assign a second value to the bit in response to the first compressed data being uncompressed.
11. The non-transitory machine readable medium of claim 7 , wherein the instructions, when executed, cause the one or more processors to record a quantity of cache lines occupied by the first compressed data in the second portion of the first meta-data.
12. The non-transitory machine readable medium of claim 7 , wherein the compressed data includes non-zero weights from the matrix of weights and a bitmap indicative of respective locations of the non-zero weights in the matrix of weights.
13. A method comprising:
executing a compression operation on a first portion of a matrix of weights for a neural network to obtain first compressed data, the first portion of the matrix is a sparse portion;
executing the compression operation on a second portion of the matrix of weights for a neural network to obtain second compressed data, the second portion of the matrix is a sparse portion;
determining first meta-data associated with the first compressed data and second meta-data associated with the second compressed data, a first portion of the first meta-data indicative of whether the first compressed data is compressed, a second portion of the first meta-data indicative of a cache size of the first compressed data, and a third portion of the first meta-data indicative of the compression operation executed to obtain the first compressed data;
causing the first meta-data and the second meta-data to be stored contiguously in a first row of a memory;
causing the first compressed data to be stored in a second row in the memory; and
causing the second compressed data to be stored in a third row in the memory, wherein the second row and the third row are consecutive rows.
14. The method of claim 13 , wherein the first meta-data is a byte.
15. The method of claim 13 , wherein the first portion of the first meta-data is of a first size, the third portion of the first meta-data is of a second size larger than the first size, and the second portion of the first meta-data is of a third size larger than the second size.
16. The method of claim 13 , wherein the first portion of the first meta-data is a bit, wherein determining the first meta-data associated with the first compressed data includes:
assigning a first value to the bit in response to the first compressed data being compressed; and
assigning a second value to the bit in response to the first compressed data being uncompressed.
17. The method of claim 13 , wherein determining the first meta-data includes a quantity of cache lines occupied by the first compressed data in the second portion of the second meta-data.
18. The method of claim 13 , wherein executing the compression operation includes:
packing non-zero weights from the matrix of weights into a compressed array; and
generating a bitmap indicative of respective positions of the non-zero weights in the matrix of weights.