anv: Implement VK_KHR_zero_initialize_workgroup_memory

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8708>

docs/relnotes/new_features.txt

@@ -15,3 +15,4 @@ Sparse memory support on RADV
 Rapid packed math (16bit-vectorization) on RADV
 VK_KHR_workgroup_memory_explicit_layout on Intel, RADV
 DRM format modifiers for AMD.
+VK_KHR_zero_initialize_workgroup_memory on Intel

src/intel/vulkan/anv_device.c

@@ -1451,6 +1451,13 @@ void anv_GetPhysicalDeviceFeatures2(
          break;
       }
 
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ZERO_INITIALIZE_WORKGROUP_MEMORY_FEATURES_KHR: {
+         VkPhysicalDeviceZeroInitializeWorkgroupMemoryFeaturesKHR *features =
+            (VkPhysicalDeviceZeroInitializeWorkgroupMemoryFeaturesKHR *)ext;
+         features->shaderZeroInitializeWorkgroupMemory = true;
+         break;
+      }
+
       default:
          anv_debug_ignored_stype(ext->sType);
          break;

src/intel/vulkan/anv_extensions.py

@@ -122,6 +122,7 @@ EXTENSIONS = [
     Extension('VK_KHR_workgroup_memory_explicit_layout',  1, True),
     Extension('VK_KHR_xcb_surface',                       6, 'VK_USE_PLATFORM_XCB_KHR'),
     Extension('VK_KHR_xlib_surface',                      6, 'VK_USE_PLATFORM_XLIB_KHR'),
+    Extension('VK_KHR_zero_initialize_workgroup_memory',  1, True),
     Extension('VK_EXT_4444_formats',                      1, True),
     Extension('VK_EXT_acquire_xlib_display',              1, 'VK_USE_PLATFORM_XLIB_XRANDR_EXT'),
     Extension('VK_EXT_buffer_device_address',             2, 'device->has_a64_buffer_access'),

src/intel/vulkan/anv_pipeline.c

@@ -1748,6 +1748,22 @@ anv_pipeline_compile_cs(struct anv_compute_pipeline *pipeline,
 
       NIR_PASS_V(stage.nir, nir_lower_explicit_io,
                  nir_var_mem_shared, nir_address_format_32bit_offset);
+
+      if (stage.nir->info.cs.zero_initialize_shared_memory &&
+          stage.nir->info.cs.shared_size > 0) {
+         /* The effective Shared Local Memory size is at least 1024 bytes and
+          * is always rounded to a power of two, so it is OK to align the size
+          * used by the shader to chunk_size -- which does simplify the logic.
+          */
+         const unsigned chunk_size = 16;
+         const unsigned shared_size = ALIGN(stage.nir->info.cs.shared_size, chunk_size);
+         assert(shared_size <=
+                calculate_gen_slm_size(compiler->devinfo->gen, stage.nir->info.cs.shared_size));
+
+         NIR_PASS_V(stage.nir, nir_zero_initialize_shared_memory,
+                    shared_size, chunk_size);
+      }
+
       NIR_PASS_V(stage.nir, brw_nir_lower_cs_intrinsics);
 
       stage.num_stats = 1;