[Mesa-dev] [PATCH 0/6] nir: Implement a load-combine pass

[Eduardo Lima Mitev]

Hi,

Any chance to get this looked at?
Thanks!

Eduardo

On 04/14/2016 06:52 PM, Eduardo Lima Mitev wrote:
> Hi,
> 
> This is a series adding a new NIR pass that will combine redundant SSBO, shared variable and image load instructions. It is based on a previous series that Iago Toral [1] sent a few months ago, which I have updated to account for changes in NIR since then, and also added support for shared variables and image load/store.
> 
> Currently, the emitted load instructions for these types of variables are highly redundant. For example, assuming 'value' is an SSBO float, doing:
> 
> float next = value + 1.0;
> float prev = value - 1.0;
> 
> will generate two redundant load_ssbo instructions. The extreme case of this is:
> 
> mat4 mult = ssbo_mat4 * ssbo_mat4;
> 
> which will spwan 16 load_ssbo instructions.
> 
> Compute shader's shared variables and image loads (from ARB_image_load_store) suffer from the same problem. This pass coaleses the redundant loads, and is also able to combine a load with a previous store to the same address.
> 
> I'm still working on a piglit test series that tied up the different corner cases, but it doesn't hurt to get a few eyes on this already. The WIP piglit branch is at: https://github.com/Igalia/piglit/commits/nir-opt-load-combine-rc1
> You can see there a test example for the matrix multiplication cases: https://github.com/Igalia/piglit/commit/b7ea6d2553891c7eec993c39352a10925ebcefbc. This pass is able to reduce the emitted instructions from 210 to 134 (~64%), comparing to Mesa master.
> 
> No piglit or dEQP-GLES31 regressions observed.
> 
> cheers,
> Eduardo
> 
> [1] https://lists.freedesktop.org/archives/mesa-dev/2015-October/098426.html
> 
> Eduardo Lima Mitev (3):
>   nir/opt_load_combine: Extend the pass to include shared variables
>   nir/opt_load_combine: Extend the pass to include image load/store
> 
> Iago Toral (3):
>   nir: add a load-combine pass
>   nir/load_combine: expand the pass to support load-after-store
>   i965: use the load-combine pass
> 
>  src/compiler/Makefile.sources           |   1 +
>  src/compiler/nir/nir.h                  |   2 +
>  src/compiler/nir/nir_opt_load_combine.c | 807 ++++++++++++++++++++++++++++++++
>  src/mesa/drivers/dri/i965/brw_nir.c     |   1 +
>  4 files changed, 811 insertions(+)
>  create mode 100644 src/compiler/nir/nir_opt_load_combine.c
>