Skip to content

Bf16 replacement kernels (develop branch)#715

Merged
zaliu merged 2 commits into
ROCm:developfrom
zaliu:bf16_replacement_kernels
Sep 24, 2019
Merged

Bf16 replacement kernels (develop branch)#715
zaliu merged 2 commits into
ROCm:developfrom
zaliu:bf16_replacement_kernels

Conversation

@zaliu
Copy link
Copy Markdown
Contributor

@zaliu zaliu commented Sep 24, 2019

full rocblas-test on gfx900 and gfx906 resulted in the 4 known nrm2 failures.
full rocblas-test on gfx908 resulted in the 4 known nrm2 failures and 4 failures to be labeled "known bugs".
JIRA ticket will be filed to get C/D separation in rocblas-bench properly dealt with.

* Revert "Switch to using separate D for gemm_ex benchmark calls (ROCm#667)"
This reverts commit 402d231.
* bf16 kernels for gfx908
* use bf16 UseBeta=0 replacement kernels
* update tensile_tag to use bf16 UseBeta=0 replacement kernels
* restore UseBeta=1 logic for arcturus BF16 TN
@leekillough
Copy link
Copy Markdown
Contributor

Why are C and D being merged in rocblas-bench? Do we need to change rocblas-test/rocblas-bench to use a YAML flag to indicate C==D or C != D, to separate the two cases out?

@leekillough
Copy link
Copy Markdown
Contributor

C and D can be separated out this way:

ldd <= 0: C and D are the same (exactly overlapping)
ldd > 0: C and D are separate (non-overlapping)

@zaliu zaliu merged commit 8e48a65 into ROCm:develop Sep 24, 2019
@zaliu zaliu deleted the bf16_replacement_kernels branch September 24, 2019 18:20
mlse-lib-jenkins pushed a commit that referenced this pull request Jun 4, 2021
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants