[Mesa-dev] [PATCH] radv: do not lower UBO/SSBO access to offsets

[Jason Ekstrand]

I did some benchmarking of Assassin's Creed Odyssey today on ANV.
Disabling the SSBO optimization possibilities in ANV that moving to derefs
unlocks (I just moved the lowering super-early; moving back to index/offset
would be insane) drops the perf of ACO by 20%.  You want SSBO derefs. :D

On Thu, Mar 28, 2019 at 9:08 AM Jason Ekstrand <jason at jlekstrand.net> wrote:

> On Wed, Mar 27, 2019 at 4:13 AM Samuel Pitoiset <samuel.pitoiset at gmail.com>
> wrote:
>
>> This helps few compute shaders, mostly for F12017.
>>
>> 27670 shaders in 14347 tests
>> Totals:
>> SGPRS: 1231173 -> 1231173 (0.00 %)
>> VGPRS: 866056 -> 865928 (-0.01 %)
>> Spilled SGPRs: 24201 -> 24201 (0.00 %)
>> Code Size: 46137040 -> 46144868 (0.02 %) bytes
>> Max Waves: 232287 -> 232302 (0.01 %)
>>
>> Totals from affected shaders:
>> SGPRS: 24624 -> 24624 (0.00 %)
>> VGPRS: 25960 -> 25832 (-0.49 %)
>> Spilled SGPRs: 0 -> 0 (0.00 %)
>> Code Size: 2922632 -> 2930460 (0.27 %) bytes
>> Max Waves: 1216 -> 1231 (1.23 %)
>>
>> Suggested-by: <Jason Ekstrand jason at jlekstrand.net>
>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>> ---
>>  src/amd/common/ac_nir_to_llvm.c | 8 ++++++++
>>  src/amd/vulkan/radv_shader.c    | 5 ++++-
>>  2 files changed, 12 insertions(+), 1 deletion(-)
>>
>> diff --git a/src/amd/common/ac_nir_to_llvm.c
>> b/src/amd/common/ac_nir_to_llvm.c
>> index b25cc6a0a84..c46d98e6dd9 100644
>> --- a/src/amd/common/ac_nir_to_llvm.c
>> +++ b/src/amd/common/ac_nir_to_llvm.c
>> @@ -3275,6 +3275,14 @@ static void visit_intrinsic(struct ac_nir_context
>> *ctx,
>>         case nir_intrinsic_vulkan_resource_reindex:
>>                 result = visit_vulkan_resource_reindex(ctx, instr);
>>                 break;
>> +       case nir_intrinsic_load_vulkan_descriptor: {
>> +               LLVMValueRef values[2] = {
>> +                       get_src(ctx, instr->src[0]),
>> +                       ctx->ac.i32_0,
>> +               };
>> +               result = ac_build_gather_values(&ctx->ac, values, 2);
>> +               break;
>> +       }
>>         case nir_intrinsic_store_ssbo:
>>                 visit_store_ssbo(ctx, instr);
>>                 break;
>> diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
>> index 19a807df199..2751302e8b9 100644
>> --- a/src/amd/vulkan/radv_shader.c
>> +++ b/src/amd/vulkan/radv_shader.c
>> @@ -220,7 +220,6 @@ radv_shader_compile_to_nir(struct radv_device *device,
>>                         }
>>                 }
>>                 const struct spirv_to_nir_options spirv_options = {
>> -                       .lower_ubo_ssbo_access_to_offsets = true,
>>                         .caps = {
>>                                 .descriptor_array_dynamic_indexing = true,
>>                                 .device_group = true,
>> @@ -306,6 +305,10 @@ radv_shader_compile_to_nir(struct radv_device
>> *device,
>>                 NIR_PASS_V(nir, nir_lower_system_values);
>>                 NIR_PASS_V(nir, nir_lower_clip_cull_distance_arrays);
>>                 NIR_PASS_V(nir, nir_lower_frexp);
>> +
>> +               NIR_PASS_V(nir, nir_lower_explicit_io,
>> +                         nir_var_mem_ubo | nir_var_mem_ssbo,
>> +                         nir_address_format_32bit_index_offset);
>>
>
> If you actually want to get SSBO access optimization, you need to call
> this *after* your first call to your main optimization loop.  Otherwise,
> the change is basically just a no-op which just shuffles around the way
> address calculations are done a bit.
>
> --Jason
>
>
>>         }
>>
>>         /* Vulkan uses the separate-shader linking model */
>> --
>> 2.21.0
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20190328/59f8072a/attachment-0001.html>