hk: fix null FS corner cases

this fixes null FS + cull distance/API sample mask, which require a prolog.
fixes upcoming CTS.

Backport-to: 25.1
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34585>
(cherry picked from commit 3ab8ce8579c967e1d8fb024de298a958511d9ffd)

.pick_status.json

@@ -1424,7 +1424,7 @@
         "description": "hk: fix null FS corner cases",
         "nominated": true,
         "nomination_type": 4,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": null,
         "notes": null

src/asahi/vulkan/hk_cmd_draw.c

@@ -1636,8 +1636,18 @@ hk_cmd_bind_graphics_shader(struct hk_cmd_buffer *cmd,
                             const gl_shader_stage stage,
                             struct hk_api_shader *shader)
 {
+   struct hk_device *dev = hk_cmd_buffer_device(cmd);
    struct vk_dynamic_graphics_state *dyn = &cmd->vk.dynamic_graphics_state;
 
+   /* Null fragment shaders are annoying to handle, because we may still need an
+    * actual fragment shader to attach a prolog to. Rather than adding special
+    * cases, we just bind an empty fragment shader instead of NULL to make
+    * everything work correctly.
+    */
+   if (stage == MESA_SHADER_FRAGMENT && shader == NULL) {
+      shader = dev->null_fs;
+   }
+
    assert(stage < ARRAY_SIZE(cmd->state.gfx.shaders));
    if (cmd->state.gfx.shaders[stage] == shader)
       return;

src/asahi/vulkan/hk_device.c

@@ -20,6 +20,7 @@
 #include "asahi/lib/agx_bo.h"
 #include "asahi/lib/agx_device.h"
 #include "asahi/libagx/geometry.h"
+#include "compiler/nir/nir_builder.h"
 #include "util/hash_table.h"
 #include "util/ralloc.h"
 #include "util/simple_mtx.h"
@@ -466,6 +467,23 @@ hk_CreateDevice(VkPhysicalDevice physicalDevice,
    if (result != VK_SUCCESS)
       goto fail_mem_cache;
 
+   /* Precompile an empty fragment shader that can be used to handle API-level
+    * null fragment shaders. We do this at device-time to make binds cheap.
+    * Regardless, compiling this shader should be fast.
+    */
+   nir_builder b = nir_builder_init_simple_shader(MESA_SHADER_FRAGMENT,
+                                                  &agx_nir_options, "empty FS");
+   struct vk_shader_compile_info info = {
+      .nir = b.shader,
+      .robustness = &vk_robustness_disabled,
+      .stage = MESA_SHADER_FRAGMENT,
+   };
+   hk_compile_shader(dev, &info, NULL, pAllocator, &dev->null_fs);
+   if (!dev->null_fs) {
+      result = VK_ERROR_OUT_OF_HOST_MEMORY;
+      goto fail_meta;
+   }
+
    *pDevice = hk_device_to_handle(dev);
 
    simple_mtx_init(&dev->scratch.lock, mtx_plain);
@@ -479,6 +497,8 @@ hk_CreateDevice(VkPhysicalDevice physicalDevice,
 
    return VK_SUCCESS;
 
+fail_meta:
+   hk_device_finish_meta(dev);
 fail_mem_cache:
    vk_pipeline_cache_destroy(dev->mem_cache, NULL);
 fail_queue:
@@ -534,6 +554,10 @@ hk_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
    agx_scratch_fini(&dev->scratch.cs);
    simple_mtx_destroy(&dev->scratch.lock);
 
+   if (dev->null_fs) {
+      hk_api_shader_destroy(&dev->vk, &dev->null_fs->vk, pAllocator);
+   }
+
    hk_destroy_sampler_heap(dev, &dev->samplers);
    hk_descriptor_table_finish(dev, &dev->images);
    hk_descriptor_table_finish(dev, &dev->occlusion_queries);

src/asahi/vulkan/hk_device.h

@@ -98,7 +98,7 @@ struct hk_device {
 
    struct hk_internal_shaders prolog_epilog;
    struct hk_internal_shaders kernels;
-   struct hk_api_shader *write_shader;
+   struct hk_api_shader *null_fs;
 
    /* Indirected for common secondary emulation */
    struct vk_device_dispatch_table cmd_dispatch;