turnip/ci: Increase coverage of a660-vk job

Use 5 devices instead of 3, skip redundant tests, increase
tests_per_group value, and decrease the fraction used.

Signed-off-by: Valentine Burley <valentine.burley@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38239>

src/freedreno/ci/deqp-freedreno-a660-vk-full.toml

@@ -14,7 +14,7 @@ caselists = ["/deqp-vk/mustpass/vk-main.txt"]
 timeout = 300
 prefix = "gmem-"
 fraction = 10
-tests_per_group = 5000
+tests_per_group = 10000
 [deqp.env]
 TU_DEBUG = "gmem,forcebin"
 

src/freedreno/ci/deqp-freedreno-a660-vk.toml

@@ -3,6 +3,7 @@
 deqp = "/deqp-vk/external/vulkancts/modules/vulkan/deqp-vk"
 caselists = ["/deqp-vk/mustpass/vk-main.txt"]
 renderer_check = "Turnip Adreno .* 660"
+fraction = 3
 tests_per_group = 10000
 
 # force-gmem testing
@@ -12,7 +13,7 @@ deqp = "/deqp-vk/external/vulkancts/modules/vulkan/deqp-vk"
 caselists = ["/deqp-vk/mustpass/vk-main.txt"]
 prefix = "gmem-"
 fraction = 20
-tests_per_group = 5000
+tests_per_group = 10000
 [deqp.env]
 TU_DEBUG = "gmem,forcebin"
 

src/freedreno/ci/freedreno-a660-fails.txt

@@ -40,7 +40,9 @@ dEQP-VK.transform_feedback.simple_optimized_gpl.draw_indirect_counter_offset_508
 dEQP-VK.transform_feedback.simple_optimized_gpl.draw_indirect_multiview_counter_offset_16,Fail
 dEQP-VK.transform_feedback.simple_optimized_gpl.draw_indirect_multiview_counter_offset_244,Fail
 dEQP-VK.transform_feedback.simple_optimized_gpl.draw_indirect_multiview_counter_offset_508,Fail
+gmem-dEQP-VK.pipeline.fast_linked_library.vertex_input.misc.unused_binding_dynamic,Crash
 gmem-dEQP-VK.transform_feedback.simple.draw_indirect_multiview_counter_offset_16,Fail
+gmem-dEQP-VK.transform_feedback.simple.draw_indirect_multiview_counter_offset_244,Fail
 nobin-dEQP-VK.transform_feedback.simple.draw_indirect_multiview_counter_offset_16,Fail
 
 # Bad assumptions in piglit about layout of multiplanar formats,

src/freedreno/ci/freedreno-a660-slow-skips.txt

@@ -13,3 +13,9 @@ dEQP-VK.subgroups.clustered.compute.subgroupclusteredmin_uint64_t_requiredsubgro
 # used in ir3 create_variant, but no single pass there dominating its time.
 # Just use the scalar cases, which are fast enough for CI.
 dEQP-VK.subgroups.clustered.compute.subgroupclustered(min|max|mul)_[iu]64vec[234].*
+
+# Speed up pre-merge testing by removing redundant coverage in force-gmem mode
+gmem-dEQP-VK.api.*
+gmem-dEQP-VK.renderpass\..*
+gmem-dEQP-VK.synchronization\..*
+gmem-dEQP-VK.wsi.*

src/freedreno/ci/gitlab-ci.yml

@@ -266,11 +266,10 @@ a660-vk:
     - .lava-arm64-test-vk
     - .collabora-turnip-rules
     - .lava-sm8350-hdk:arm64
-  parallel: 3
+  parallel: 5
   variables:
     DEQP_SUITE: freedreno-a660-vk
     FLAKES_CHANNEL: "#freedreno-ci"
-    DEQP_FRACTION: 6
 
 a660-vk-full:
   extends: