Instructions and logic to perform floating point and integer operations for machine learning
A processing apparatus is provided comprising a multiprocessor having a multithreaded architecture. The multiprocessor can execute at least one single instruction to perform parallel mixed precision matrix operations. In one embodiment the apparatus includes a memory interface and an array of multiprocessors coupled to the memory interface. At least one multiprocessor in the array of multiprocessors is configured to execute a fused multiply-add instruction in parallel across multiple threads.
1. A graphics processing unit (GPU) comprising:
a plurality of memory controllers;
cache memory coupled with the plurality of memory controllers;
a graphics multiprocessor coupled with the cache memory and the plurality of memory controllers, the graphics multiprocessor having a single instruction, multiple thread (SIMT) architecture, wherein the graphics multiprocessor includes:
a register file; and
circuitry coupled with the register file, the circuitry including a first core to perform a mixed precision matrix operation and a second core to perform, in response to a single instruction, multiple compute operations, wherein the multiple compute operations include a first operation to perform a fused multiply-add and a second operation to apply a rectified linear unit function to a result of the first operation.
2. The GPU as in claim 1 , wherein the first operation and the second operation are single instruction multiple data (SIMD) operations.
3. The GPU as in claim 1 , wherein the multiple compute operations are performed on input in a 16-bit floating-point format having a 1-bit sign and an 8-bit exponent.
4. The GPU as in claim 3 , wherein the second core includes a dynamic precision processing resource that is configurable to automatically convert input in a 32-bit floating point format to the 16-bit floating-point format in conjunction with execution of the single instruction.
5. The GPU as in claim 4 , wherein the dynamic precision processing resource includes configuration circuitry to dynamically configure a precision of functional circuitry within the dynamic precision processing resource.
6. The GPU as in claim 5 , wherein the configuration circuitry is to configure a first stage of the dynamic precision processing resource to operate at a first precision and a second stage of the dynamic precision processing resource to operate at a second precision.
7. The GPU as in claim 1 , wherein the rectified linear unit function is an activation function associated with a first neuron of a neural network.
8. A system comprising:
a memory device;
a graphics processing unit (GPU) comprising a plurality of memory controllers coupled with the memory device, cache memory coupled with the plurality of memory controllers, and a graphics multiprocessor coupled with the cache memory and the plurality of memory controllers, the graphics multiprocessor having a single instruction, multiple thread (SIMT) architecture, wherein the graphics multiprocessor includes:
a register file; and
circuitry coupled with the register file, the circuitry including a first core to perform a mixed precision matrix operation and a second core to perform, in response to a single instruction, multiple compute operations, wherein the multiple compute operations include a first operation to perform a fused multiply-add and a second operation to apply a rectified linear unit function to a result of the first operation.
9. The system as in claim 8 , wherein the first operation and the second operation are single instruction multiple data (SIMD) operations.
10. The system as in claim 8 , wherein the multiple compute operations are performed on input in a 16-bit floating-point format having a 1-bit sign and an 8-bit exponent.
11. The system as in claim 10 , wherein the second core includes a dynamic precision processing resource that is configurable to automatically convert input in a 32-bit floating point format to the 16-bit floating-point format in conjunction with execution of the single instruction.
12. The system as in claim 11 , wherein the dynamic precision processing resource includes configuration circuitry to dynamically configure a precision of functional circuitry within the dynamic precision processing resource.
13. The system as in claim 12 , wherein the configuration circuitry is to configure a first stage of the dynamic precision processing resource to operate at a first precision and a second stage of the dynamic precision processing resource to operate at a second precision.
14. The system as in claim 8 , wherein the rectified linear unit function is an activation function associated with a first neuron of a neural network.
15. A method comprising:
fetching a single instruction from a cache memory of a graphics processing unit (GPU); loading operands associated with the single instruction into a register file of the GPU;
decoding the single instruction into a decoded instruction;
executing the decoded instruction via circuitry coupled with the register file, the circuitry including a first core and a second core, the first core to perform a mixed precision matrix operation and the second core to execute the decoded instruction to perform multiple compute operations; and
performing the multiple compute operations via the second core, wherein the multiple compute operations include a first operation to perform a fused multiply-add and a second operation to apply a rectified linear unit function to a result of the first operation.
16. The method as in claim 15 , wherein the first operation and the second operation are single instruction multiple data (SIMD) operations.
17. The method as in claim 15 , further comprising performing the multiple compute operations on input in a 16-bit floating-point format having a 1-bit sign and an 8-bit exponent.
18. The method as in claim 17 , further comprising automatically converting input in a 32-bit floating point format to the 16-bit floating-point format in conjunction with execution of the decoded instruction.
19. The method as in claim 18 , further comprising dynamically configuring a precision of functional circuitry within the second core in conjunction with performing the fused multiply-add.
20. The method as in claim 19 , wherein dynamically configuring the precision of the functional circuitry includes configuring a first stage of the functional circuitry to operate at a first precision and a second stage of the functional circuitry to operate at a second precision.
21. An apparatus comprising:
means for fetching a single instruction from a cache memory of a processing unit;
means for loading operands associated with the single instruction;
means for decoding the single instruction into a decoded instruction;
means for executing the decoded instruction by performing multiple compute operations; and
means for performing the multiple compute operations via a first operation to perform a fused multiply-add and a second operation to apply a rectified linear unit function to a result of the first operation.
22. The apparatus as in claim 21 , wherein the first operation and the second operation are single instruction multiple data (SIMD) operations.
23. The apparatus as in claim 21 , further comprising means for performing the multiple compute operations on input in a 16-bit floating-point format having a 1-bit sign and an 8-bit exponent.
24. The apparatus as in claim 21 , wherein the apparatus is a general purpose graphics processing unit (GPGPU).
25. The apparatus as in claim 21 , wherein the apparatus is a compute accelerator to accelerate operations associated with a neural network.