tu: Disable 128-wide subgroups on No Man's Sky.

It has a lighting compute shader that assumes either 32 or 64, and that
produces vertical banding if the compiler opts for 128.

Cc: mesa-stable
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15423
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41665>

src/freedreno/vulkan/tu_device.cc

@@ -885,8 +885,8 @@ tu_get_physical_device_properties_1_1(struct tu_physical_device *pdevice,
    p->deviceNodeMask = 0;
    p->deviceLUIDValid = false;
 
-   p->subgroupSize = pdevice->info->props.supports_double_threadsize ?
-      pdevice->info->threadsize_base * 2 : pdevice->info->threadsize_base;
+   p->subgroupSize =
+      pdevice->expose_double_threadsize ? pdevice->info->threadsize_base * 2 : pdevice->info->threadsize_base;
    p->subgroupSupportedStages = VK_SHADER_STAGE_COMPUTE_BIT;
    p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
                                     VK_SUBGROUP_FEATURE_VOTE_BIT |
@@ -1034,8 +1034,8 @@ tu_get_physical_device_properties_1_3(struct tu_physical_device *pdevice,
                                       struct vk_properties *p)
 {
    p->minSubgroupSize = pdevice->info->threadsize_base;
-   p->maxSubgroupSize = pdevice->info->props.supports_double_threadsize ?
-      pdevice->info->threadsize_base * 2 : pdevice->info->threadsize_base;
+   p->maxSubgroupSize =
+      pdevice->expose_double_threadsize ? pdevice->info->threadsize_base * 2 : pdevice->info->threadsize_base;
    p->maxComputeWorkgroupSubgroups = pdevice->info->max_waves;
    p->requiredSubgroupSizeStages = VK_SHADER_STAGE_ALL;
 
@@ -1164,9 +1164,9 @@ tu_get_properties(struct tu_physical_device *pdevice,
    props->maxComputeWorkGroupCount[0] =
       props->maxComputeWorkGroupCount[1] =
       props->maxComputeWorkGroupCount[2] = 65535;
-   props->maxComputeWorkGroupInvocations = pdevice->info->props.supports_double_threadsize ?
-      pdevice->info->threadsize_base * 2 * pdevice->info->max_waves :
-      pdevice->info->threadsize_base * pdevice->info->max_waves;
+   props->maxComputeWorkGroupInvocations = pdevice->expose_double_threadsize
+                                              ? pdevice->info->threadsize_base * 2 * pdevice->info->max_waves
+                                              : pdevice->info->threadsize_base * pdevice->info->max_waves;
    if (pdevice->info->props.is_a702) {
       props->maxComputeWorkGroupSize[0] =
          props->maxComputeWorkGroupSize[1] = 512;
@@ -1687,6 +1687,8 @@ tu_physical_device_init(struct tu_physical_device *device,
       goto fail_free_name;
    }
 
+   device->expose_double_threadsize = info.props.supports_double_threadsize && !instance->restrict_subgroup_size_64;
+
    device->level1_dcache_size = util_cache_granularity();
    device->has_cached_non_coherent_memory =
       device->level1_dcache_size > 0 && !DETECT_ARCH_ARM;
@@ -1863,6 +1865,7 @@ static const driOptionDescription tu_dri_options[] = {
       DRI_CONF_TU_EMULATE_ALPHA_TO_COVERAGE(false)
       DRI_CONF_TU_AUTOTUNE_ALGORITHM()
       DRI_CONF_TU_OVERRIDE_UNCACHED_AS_CACHE_COHERENT(false)
+      DRI_CONF_TU_RESTRICT_SUBGROUP_SIZE_64(false)
    DRI_CONF_SECTION_END
 };
 
@@ -1911,6 +1914,7 @@ tu_init_dri_options(struct tu_instance *instance)
    instance->allow_concurrent_binning =
       (driQueryOptionb(&instance->dri_options, "tu_allow_concurrent_binning") && !TU_DEBUG(NO_CONCURRENT_BINNING)) ||
       TU_DEBUG(FORCE_CONCURRENT_BINNING);
+   instance->restrict_subgroup_size_64 = driQueryOptionb(&instance->dri_options, "tu_restrict_subgroup_size_64");
 }
 
 static uint32_t instance_count = 0;

src/freedreno/vulkan/tu_device.h

@@ -145,6 +145,8 @@ struct tu_physical_device
 
    bool has_preemption;
 
+   bool expose_double_threadsize;
+
    /* Whether performance counter selector registers can be written by userspace CSes. */
    bool is_perf_cntr_selectable;
 
@@ -231,6 +233,11 @@ struct tu_instance
     */
    bool enable_d24s8_border_color_workaround;
 
+   /* Various games assume that gl_SubgroupSize is either 32 or 64, and we hide
+    * our 128-invocation subgroup support for them.
+    */
+   bool restrict_subgroup_size_64;
+
    /* When D24S8 is used without enable_d24s8_border_color_workaround, the
     * fast border color HW feature results in an incorrect color being used.
     * However, we want to enable fast border colors for apps that are known

src/freedreno/vulkan/tu_shader.cc

@@ -3471,7 +3471,7 @@ tu_shader_key_subgroup_size(struct tu_shader_key *key,
                             struct tu_device *dev)
 {
    enum ir3_wavesize_option api_wavesize, real_wavesize;
-   if (!dev->physical_device->info->props.supports_double_threadsize) {
+   if (!dev->physical_device->expose_double_threadsize) {
       api_wavesize = IR3_SINGLE_ONLY;
       real_wavesize = IR3_SINGLE_ONLY;
    } else {

src/util/00-mesa-defaults.conf

@@ -1596,6 +1596,14 @@ TODO: document the other workarounds.
         <application name="Half-Life: Alyx" application_name_match="hlvr">
             <option name="tu_emulate_alpha_to_coverage" value="true" />
         </application>
+        <application name='No Man&apos;s Sky' application_name_match='No Man&apos;s Sky'>
+            <!-- A lighting CS reducing 8x8 regions with fmin/fmax does a write once per
+            subgroup (assuming 64), or collects the results from two subgroups in the
+            32 subgroup size case.  Thus, our 128 subgroup size results in vertical banding
+            when half the 8x8 regions don't get written.
+            -->
+            <option name="tu_restrict_subgroup_size_64" value="true" />
+        </application>
     </device>
 
     <device driver="asahi">

src/util/driconf.h

@@ -668,6 +668,10 @@
    DRI_CONF_OPT_B(tu_allow_concurrent_binning, def, \
                   "Allow concurrent binning on A7XX+, the CB is disabled by default because it regresses performance on desktop games")
 
+#define DRI_CONF_TU_RESTRICT_SUBGROUP_SIZE_64(def) \
+   DRI_CONF_OPT_B(tu_restrict_subgroup_size_64, def, \
+                  "Restrict subgroup size to 64 (instead of a max of 128) to work around games assuming desktop GPU 32/64 sizes")
+
 /**
  * \brief Honeykrisp specific configuration options
  */