Re: [Mesa-dev] [PATCH] radv: compact varyings after removing unused ones

[Samuel Pitoiset]

On 02/14/2018 11:24 PM, Timothy Arceri wrote:

On 15/02/18 01:10, Samuel Pitoiset wrote:

It makes no sense to compact before, and the description of
nir_compact_varyings() confirms that.


Reviewed-by: Timothy Arceri 

Your right this seems to be in the wrong place. There was a lot of 
restructuring around the time this landed so it seems to have been moved 
to the wrong location at some point. The original commit message (below) 
records a performance improvement it might be good to check if this 
improves things further.


     SaschaWillems Vulkan demo tessellation:

     ~4000fps -> ~4600fps


Mmmh, with this patch 5800->5200 fps on vega, I will have a look.







Polaris10:
Totals from affected shaders:
SGPRS: 106000 -> 105600 (-0.38 %)
VGPRS: 70704 -> 70640 (-0.09 %)
Spilled SGPRs: 892 -> 878 (-1.57 %)
Code Size: 2826240 -> 2812476 (-0.49 %) bytes
Max Waves: 15909 -> 15899 (-0.06 %)

Vega10:
Totals from affected shaders:
SGPRS: 103520 -> 103664 (0.14 %)
VGPRS: 70776 -> 70808 (0.05 %)
Spilled SGPRs: 740 -> 722 (-2.43 %)
Code Size: 2786076 -> 2772004 (-0.51 %) bytes
Max Waves: 18026 -> 18010 (-0.09 %)

Signed-off-by: Samuel Pitoiset 
---
  src/amd/vulkan/radv_pipeline.c | 9 +++--
  1 file changed, 3 insertions(+), 6 deletions(-)

diff --git a/src/amd/vulkan/radv_pipeline.c 
b/src/amd/vulkan/radv_pipeline.c

index 8f872e7c14..06e1ec5ae3 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -1525,6 +1525,9 @@ radv_link_shaders(struct radv_pipeline 
*pipeline, nir_shader **shaders)

 ordered_shaders[i - 1]);
  if (progress) {
+    nir_compact_varyings(ordered_shaders[i],
+ ordered_shaders[i - 1], true);
+
  nir_lower_global_vars_to_local(ordered_shaders[i]);
  radv_optimize_nir(ordered_shaders[i]);
  nir_lower_global_vars_to_local(ordered_shaders[i - 1]);
@@ -1715,7 +1718,6 @@ void radv_create_shaders(struct radv_pipeline 
*pipeline,

  last = i;
  }
-    int prev = -1;
  for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
  const VkPipelineShaderStageCreateInfo *stage = pStages[i];
@@ -1746,11 +1748,6 @@ void radv_create_shaders(struct radv_pipeline 
*pipeline,

  nir_lower_io_to_scalar_early(nir[i], mask);
  radv_optimize_nir(nir[i]);
  }
-
-    if (prev != -1) {
-    nir_compact_varyings(nir[prev], nir[i], true);
-    }
-    prev = i;
  }
  if (nir[MESA_SHADER_TESS_CTRL]) {


___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev

Re: [Mesa-dev] [PATCH] radv: compact varyings after removing unused ones

[Timothy Arceri]

On 15/02/18 01:10, Samuel Pitoiset wrote:

It makes no sense to compact before, and the description of
nir_compact_varyings() confirms that.


Reviewed-by: Timothy Arceri 

Your right this seems to be in the wrong place. There was a lot of 
restructuring around the time this landed so it seems to have been moved 
to the wrong location at some point. The original commit message (below) 
records a performance improvement it might be good to check if this 
improves things further.


SaschaWillems Vulkan demo tessellation:

~4000fps -> ~4600fps





Polaris10:
Totals from affected shaders:
SGPRS: 106000 -> 105600 (-0.38 %)
VGPRS: 70704 -> 70640 (-0.09 %)
Spilled SGPRs: 892 -> 878 (-1.57 %)
Code Size: 2826240 -> 2812476 (-0.49 %) bytes
Max Waves: 15909 -> 15899 (-0.06 %)

Vega10:
Totals from affected shaders:
SGPRS: 103520 -> 103664 (0.14 %)
VGPRS: 70776 -> 70808 (0.05 %)
Spilled SGPRs: 740 -> 722 (-2.43 %)
Code Size: 2786076 -> 2772004 (-0.51 %) bytes
Max Waves: 18026 -> 18010 (-0.09 %)

Signed-off-by: Samuel Pitoiset 
---
  src/amd/vulkan/radv_pipeline.c | 9 +++--
  1 file changed, 3 insertions(+), 6 deletions(-)

diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index 8f872e7c14..06e1ec5ae3 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -1525,6 +1525,9 @@ radv_link_shaders(struct radv_pipeline *pipeline, 
nir_shader **shaders)
   ordered_shaders[i - 
1]);
  
  		if (progress) {

+   nir_compact_varyings(ordered_shaders[i],
+ordered_shaders[i - 1], true);
+
nir_lower_global_vars_to_local(ordered_shaders[i]);
radv_optimize_nir(ordered_shaders[i]);
nir_lower_global_vars_to_local(ordered_shaders[i - 1]);
@@ -1715,7 +1718,6 @@ void radv_create_shaders(struct radv_pipeline *pipeline,
last = i;
}
  
-	int prev = -1;

for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
const VkPipelineShaderStageCreateInfo *stage = pStages[i];
  
@@ -1746,11 +1748,6 @@ void radv_create_shaders(struct radv_pipeline *pipeline,

nir_lower_io_to_scalar_early(nir[i], mask);
radv_optimize_nir(nir[i]);
}
-
-   if (prev != -1) {
-   nir_compact_varyings(nir[prev], nir[i], true);
-   }
-   prev = i;
}
  
  	if (nir[MESA_SHADER_TESS_CTRL]) {



___
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev