Hi,
On 10/13/18 3:08 AM, Jason Ekstrand wrote:
This patch series adds a simple NIR shader cache that sits right after
spirv_to_nir and brw_preprocess_nir and before linking. This should help
alleviate some of the added overhead of link-time optimization since most
of the NIR-level optimization is now cached prior to linking.
I have no numbers to back this series up; just intuition.
Fortunately approximate numbers of where the bottlenecks are, are easy
to come by for shader compilation.
1. Start use-case doing lot of shader compilation (shader-db?)
2. "sudo perf record -a"
3. ^C when shader compilation stops
4. "sudo perf report"
When I profiled GL shader compilation before Mesa switched to NIR,
linking phase was 2/3 of the cycles taken by the whole shader compilation.
(I.e. why shader cache needed to cache linking results to have
significant impact.)
- Eero
Jason Ekstrand (5):
anv/pipeline: Move wpos and input attachment lowering to lower_nir
anv/pipeline: Hash shader modules and spec constants separately
compiler/types: Serialize/deserialize subpass input types correctly
anv/pipeline_cache: Add support for caching NIR
anv/pipeline: Cache the pre-lowered NIR
src/compiler/glsl_types.cpp | 4 +-
src/intel/vulkan/anv_pipeline.c | 118 ++++++++++++++++++--------
src/intel/vulkan/anv_pipeline_cache.c | 100 ++++++++++++++++++++++
src/intel/vulkan/anv_private.h | 18 ++++
4 files changed, 204 insertions(+), 36 deletions(-)
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
