Update test and benchmark generation for blockwise kr>2 #6557
Conversation
| struct LoopParams { | ||
| LoopParams() = default; | ||
| explicit LoopParams(size_t from, size_t to, size_t step) | ||
| : is_set(true), from(from), to(to), step(step) {} | ||
| explicit LoopParams(size_t from, size_t to, size_t step, LoopStepType step_type) |
There was a problem hiding this comment.
If preferred, I could change this to a boolean step_prime variable or similar. I figured the enum class approach is slightly more readable, but does feel less idiomatic to the codebase. Feel free to weigh in on this if you prefer it differently.
| @@ -549,23 +583,3 @@ struct GemmTestParams { | |||
| }; | |||
|
|
|||
| using GemmTest = testing::TestWithParam<GemmTestParams>; | |||
|
|
|||
There was a problem hiding this comment.
See above comment - methods are moved to the top of the header.
| @@ -236,93 +236,52 @@ def split_ukernel_name(name): | |||
| , test_func, isa_check) | |||
| .loop_n(1, nr) | |||
| .loop_m(1, mr)); | |||
| if (k_block > 1) { | |||
There was a problem hiding this comment.
The diff looks messy, but all that actually changed here is that block of tests that are incompatible with the blockwise kernel assumptions are wrapped in the $if KERNELTYPE not in ['qb4w']. I did re-generate all tests via scripts/generate-tests.sh and there are no functional changes to existing tests.
|
@alankelly Hey Alan, could you take a look at this PR when you have time? Once this is merged, we can start posting the ISA-specific qb4w kernels. Thanks. |
GregoryComer
force-pushed
the
qb4w-test-multi-kc
branch
from
af15e05
to
9f6c40d
Compare
GregoryComer
force-pushed
the
qb4w-test-multi-kc
branch
from
9f6c40d
to
eac5331
Compare
GregoryComer
force-pushed
the
qb4w-test-multi-kc
branch
from
eac5331
to
01d3ee3
Compare
GregoryComer
force-pushed
the
qb4w-test-multi-kc
branch
from
01d3ee3
to
d974e3c
Compare
|
@gonnet Can you please review? |
This PR updates test generation for blockwise (qb4w) kernels in preparation for ISA-specific kernels with kr > 2. Blockwise kernels currently enforce several constraints:
The majority of the existing gemm tests cover edge cases that are incompatible with the above constraints, so this PR conditionally disables the irrelevant tests for blockwise kernels. Blockwise kernels still have tests for strides, M subtiles, N subtiles, and varying block/k sizes.
Blockwise tests also benefit from linear stepping of nc and kc. Existing test logic always calls NextPrime after incrementing nc and kc values. To allow for linear stepping in certain blockwise tests while also maintaining the existing behavior, I've tentatively added a step_type parameter to LoopParams, allowing for either linear or prime steps. The default value for step_type is set in GemmTestParams to maintain the existing behavior of prime steps for nc and kc.
Additionally, benchmark logic for blockwise kernels now rounds kc up to a multiple of 2*kr*sr.
All tests were regenerated via scripts/generate-tests.sh.