Automatic loop vectorization using hardware transactional memory
Technologies for automatic loop vectorization include a computing device with an optimizing compiler. During an optimization pass, the compiler identifies a loop and generates a transactional code segment including a vectorized implementation of the loop body including one or more vector memory read instructions capable of generating an exception. The compiler also generates a non-transactional fallback code segment including a scalar implementation of the loop body that is executed in response to an exception generated within the transactional code segment. The compiler may detect whether the loop contains a memory read dependent on a condition that may be updated in a previous iteration or whether the loop contains a potential data dependence between two iterations. The compiler may generate a dynamic check for an actual data dependence and an explicit transactional abort instruction to be executed when an actual data dependence exists. Other embodiments are described and claimed.
1. A computing device for loop vectorization, the computing device comprising:
a processor; and
one or more non-transitory, computer-readable storage media comprising a plurality of instructions that in response to being executed by the processor cause the computing device to:
detect a loop of a source program, the loop to define one or more scalar iterations and have a loop body for execution during each scalar iteration;
generate a transactional code segment, wherein to generate the transactional code segment comprises to generate a vectorized implementation of the loop body within the transactional code segment as a function of the loop of the source program, the vectorized implementation to define multiple tiled vector iterations of the vectorized loop body within a transactional envelope and to include a vector memory read instruction capable of generation of an exception; and
generate a non-transactional fallback code segment associated with the transactional code segment, the non-transactional fallback code to be executed in response to generation of an exception within the transactional code segment, wherein the exception is based on a detection of a data dependence between two scalar iterations included in the same vector iteration of the loop and wherein the non-transactional fallback code segment comprises a scalar implementation of the loop body to execute the scalar iterations sequentially in response to generation of the exception.
2. The computing device of claim 1 , wherein:
to detect the loop of the source program comprises to detect a loop that includes a memory read dependent on a condition that may be updated in a previous scalar iteration included in the same vector iteration; and
to generate the transactional code segment comprises to generate the transactional code segment in response to detection of the loop that includes the memory read.
3. The computing device of claim 1 , wherein the one or more non-transitory, computer-readable storage media further comprises a plurality of instructions that in response to being executed by the processor cause the computing device to:
detect a potential data dependence between two scalar iterations included in the same vector iteration, and
wherein to generate a transactional code segment comprises to generate, in response to the detection of the potential data dependence, the transactional envelope surrounding vectorized loop code associated with the vectorized implementation of the loop.
4. The computing device of claim 3 , wherein to detect the loop of the source program comprises to detect a loop that has a memory load dependent on a condition that may be updated in a previous scalar iteration included in the same vector iteration.
5. The computing device of claim 1 , wherein:
to generate the vectorized implementation comprises to generate a partially vectorized implementation that includes one or more vector instructions, each vector instruction to define a vector length; and
wherein each vector iteration corresponds to up to the vector length number of scalar iterations of the loop body.
6. The computing device of claim 5 , wherein to generate the partially vectorized implementation comprises to generate a partially vectorized implementation that includes a hardware partial vectorization support instruction, wherein the hardware partial vectorization support instruction comprises a set up to and including first true in mask instruction, a set up to but not including first true in mask instruction, or a masked broadcast of last set element instruction.
7. The computing device of claim 1 , wherein to generate the vectorized implementation comprises to generate a vectorized implementation of the loop body that includes a vector load instruction that is capable of generation of an exception or a vector gather instruction that is capable of generation of an exception.
8. The computing device of claim 1 , wherein to generate the transactional code segment comprises to include a plurality of vector iterations of the vectorized implementation of the loop body per transactional code segment.
9. A method for loop vectorization, the method comprising:
detecting, by a computing device, a loop of a source program, the loop defining one or more scalar iterations and having a loop body for execution during each scalar iteration;
generating, by the computing device, a transactional code segment, wherein generating the transactional code segment comprises generating a vectorized implementation of the loop body within the transactional code segment as a function of the loop of the source program, the vectorized implementation defining multiple tiled vector iterations of the vectorized loop body within a transactional envelope and including a vector memory read instruction that is capable of generating an exception; and
generating, by the computing device, a non-transactional fallback code segment associated with the transactional code segment, the non-transactional fallback code to be executed in response to generation of an exception within the transactional code segment, wherein the exception is based on a detection of a data dependence between two scalar iterations included in the same vector iteration of the loop and wherein the non-transactional fallback code segment comprises a scalar implementation of the loop body to execute the scalar iterations sequentially in response to generation of the exception.
10. The method of claim 9 , wherein:
detecting the loop of the source program comprises detecting a loop including a memory read dependent on a condition that may be updated in a previous scalar iteration included in the same vector iteration; and
generating the transactional code segment comprises generating the transactional code segment in response to detecting the loop including the memory read.
11. The method of claim 9 , further comprising:
detecting, by the computing device, a potential data dependence between two scalar iterations included in the same vector iteration,
wherein generating a transactional code segment comprises generating, in response to the detection of the potential data dependence, the transactional envelope surrounding vectorized loop code associated with the vectorized implementation of the loop.
12. The method of claim 11 , wherein detecting the loop of the source program comprises detecting a loop having a memory load dependent on a condition that may be updated in a previous scalar iteration included in the same vector iteration.
13. The method of claim 9 , wherein:
generating the vectorized implementation comprises generating a partially vectorized implementation including one or more vector instructions, each vector instruction defining a vector length; and
wherein each vector iteration corresponds to up to the vector length number of scalar iterations of the loop body.
14. The method of claim 9 , wherein generating the transactional code segment comprises including a plurality of vector iterations of the vectorized implementation of the loop body per transactional code segment.
15. One or more non-transitory, computer-readable storage media comprising a plurality of instructions that in response to being executed cause a computing device to:
detect a loop of a source program, the loop defining one or more scalar iterations and having a loop body for execution during each scalar iteration;
generate a transactional code segment, wherein to generate the transactional code segment comprises to generate a vectorized implementation of the loop body within the transactional code segment as a function of the loop of the source program, the vectorized implementation to define multiple tiled vector iterations of the vectorized loop body within a transactional envelope and including a vector memory read instruction that is capable of generating an exception; and
generate a non-transactional fallback code segment associated with the transactional code segment, the non-transactional fallback code to be executed in response to generation of an exception within the transactional code segment, wherein the exception is based on a detection of a data dependence between two scalar iterations included in the same vector iteration of the loop and wherein the non-transactional fallback code segment comprises a scalar implementation of the loop body to execute the scalar iterations sequentially in response to generation of the exception.
16. The one or more non-transitory, computer-readable storage media of claim 15 , wherein:
to detect the loop of the source program comprises to detect a loop including a memory read dependent on a condition that may be updated in a previous scalar iteration included in the same vector iteration; and
to generate the transactional code segment comprises to generate the transactional code segment in response to detecting the loop including the memory read.
17. The one or more non-transitory, computer-readable storage media of claim 15 , further comprising a plurality of instructions that in response to being executed cause the computing device to:
detect a potential data dependence between two scalar iterations included in the same vector iteration, and
wherein to generate a transactional code segment comprises to generate, in response to the detection of the potential data dependence, the transactional envelope surrounding vectorized loop code associated with the vectorized implementation of the loop.
18. The one or more non-transitory, computer-readable storage media of claim 17 , wherein to detect the loop of the source program comprises to detect a loop having a memory load dependent on a condition that may be updated in a previous scalar iteration included in the same vector iteration.
19. The one or more non-transitory, computer-readable storage media of claim 15 , wherein:
to generate the vectorized implementation comprises to generate a partially vectorized implementation including one or more vector instructions, each vector instruction defining a vector length; and
wherein each vector iteration corresponds to up to the vector length number of scalar iterations of the loop body.