anv: add an option for using indirect descriptors

This is the default for now. It needs to be part the pipeline hashing
as we will allow this to be tweaked per application.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21645>

diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
index 1c07c45..73ba463 100644 (file)
--- a/src/intel/vulkan/anv_device.c
+++ b/src/intel/vulkan/anv_device.c
@@ -77,6 +77,7 @@ static const driOptionDescription anv_dri_options[] = {
       DRI_CONF_ANV_GENERATED_INDIRECT_THRESHOLD(4)
       DRI_CONF_NO_16BIT(false)
       DRI_CONF_ANV_QUERY_CLEAR_WITH_BLORP_THRESHOLD(6)
+      DRI_CONF_ANV_FORCE_INDIRECT_DESCRIPTORS(false)
    DRI_CONF_SECTION_END
 
    DRI_CONF_SECTION_DEBUG
@@ -1325,6 +1326,13 @@ anv_physical_device_try_create(struct vk_instance *vk_instance,
 
    device->uses_ex_bso = device->info.verx10 >= 125;
 
+   /* For now always use indirect descriptors. We'll update this
+    * to !uses_ex_bso when all the infrastructure is built up.
+    */
+   device->indirect_descriptors =
+      true ||
+      driQueryOptionb(&instance->dri_options, "force_indirect_descriptors");
+
    /* Check if we can read the GPU timestamp register from the CPU */
    uint64_t u64_ignore;
    device->has_reg_timestamp = intel_gem_read_render_timestamp(fd,

diff --git a/src/intel/vulkan/anv_pipeline.c b/src/intel/vulkan/anv_pipeline.c
index 4e3a2c2..14887aa 100644 (file)
--- a/src/intel/vulkan/anv_pipeline.c
+++ b/src/intel/vulkan/anv_pipeline.c
@@ -698,23 +698,33 @@ anv_graphics_pipeline_stage_fragment_dynamic(const struct anv_pipeline_stage *st
 }
 
 static void
+anv_pipeline_hash_common(struct mesa_sha1 *ctx,
+                         const struct anv_pipeline *pipeline)
+{
+   struct anv_device *device = pipeline->device;
+
+   _mesa_sha1_update(ctx, pipeline->layout.sha1, sizeof(pipeline->layout.sha1));
+
+   const bool indirect_descriptors = device->physical->indirect_descriptors;
+   _mesa_sha1_update(ctx, &indirect_descriptors, sizeof(indirect_descriptors));
+
+   const bool rba = device->robust_buffer_access;
+   _mesa_sha1_update(ctx, &rba, sizeof(rba));
+}
+
+static void
 anv_pipeline_hash_graphics(struct anv_graphics_base_pipeline *pipeline,
                            struct anv_pipeline_stage *stages,
                            uint32_t view_mask,
                            unsigned char *sha1_out)
 {
+   const struct anv_device *device = pipeline->base.device;
    struct mesa_sha1 ctx;
    _mesa_sha1_init(&ctx);
 
-   _mesa_sha1_update(&ctx, &view_mask, sizeof(view_mask));
-
-   _mesa_sha1_update(&ctx, pipeline->base.layout.sha1,
-                     sizeof(pipeline->base.layout.sha1));
-
-   const struct anv_device *device = pipeline->base.device;
+   anv_pipeline_hash_common(&ctx, &pipeline->base);
 
-   const bool rba = device->robust_buffer_access;
-   _mesa_sha1_update(&ctx, &rba, sizeof(rba));
+   _mesa_sha1_update(&ctx, &view_mask, sizeof(view_mask));
 
    for (uint32_t s = 0; s < ANV_GRAPHICS_SHADER_STAGE_COUNT; s++) {
       if (pipeline->base.active_stages & BITFIELD_BIT(s)) {
@@ -737,16 +747,11 @@ anv_pipeline_hash_compute(struct anv_compute_pipeline *pipeline,
                           struct anv_pipeline_stage *stage,
                           unsigned char *sha1_out)
 {
+   const struct anv_device *device = pipeline->base.device;
    struct mesa_sha1 ctx;
    _mesa_sha1_init(&ctx);
 
-   _mesa_sha1_update(&ctx, pipeline->base.layout.sha1,
-                     sizeof(pipeline->base.layout.sha1));
-
-   const struct anv_device *device = pipeline->base.device;
-
-   const bool rba = device->robust_buffer_access;
-   _mesa_sha1_update(&ctx, &rba, sizeof(rba));
+   anv_pipeline_hash_common(&ctx, &pipeline->base);
 
    const bool afs = device->physical->instance->assume_full_subgroups;
    _mesa_sha1_update(&ctx, &afs, sizeof(afs));
@@ -766,11 +771,7 @@ anv_pipeline_hash_ray_tracing_shader(struct anv_ray_tracing_pipeline *pipeline,
    struct mesa_sha1 ctx;
    _mesa_sha1_init(&ctx);
 
-   _mesa_sha1_update(&ctx, pipeline->base.layout.sha1,
-                     sizeof(pipeline->base.layout.sha1));
-
-   const bool rba = pipeline->base.device->robust_buffer_access;
-   _mesa_sha1_update(&ctx, &rba, sizeof(rba));
+   anv_pipeline_hash_common(&ctx, &pipeline->base);
 
    _mesa_sha1_update(&ctx, stage->shader_sha1, sizeof(stage->shader_sha1));
    _mesa_sha1_update(&ctx, &stage->key, sizeof(stage->key.bs));

diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h
index 824b16b..9d294e8 100644 (file)
--- a/src/intel/vulkan/anv_private.h
+++ b/src/intel/vulkan/anv_private.h
@@ -912,6 +912,26 @@ struct anv_physical_device {
      */
     bool                                        generated_indirect_draws;
 
+    /**
+     * True if the descriptors buffers are holding one of the following :
+     *    - anv_sampled_image_descriptor
+     *    - anv_storage_image_descriptor
+     *    - anv_address_range_descriptor
+     *
+     * Accessing the descriptors in a bindless fashion from the shader
+     * requires an indirection in the shader, first fetch one of the structure
+     * listed above from the descriptor buffer, then emit the send message to
+     * the fixed function (sampler, dataport, etc...) with the handle fetched
+     * above.
+     *
+     * We need to do things this way prior to DG2 because the bindless surface
+     * state space is limited to 64Mb and some application will allocate more
+     * than what HW can support. On DG2+ we get 4Gb of bindless surface state
+     * and so we can reference directly RENDER_SURFACE_STATE/SAMPLER_STATE
+     * structures instead.
+     */
+    bool                                        indirect_descriptors;
+
     struct {
       uint32_t                                  family_count;
       struct anv_queue_family                   families[ANV_MAX_QUEUE_FAMILIES];

diff --git a/src/util/driconf.h b/src/util/driconf.h
index d1a000e..ac51622 100644 (file)
--- a/src/util/driconf.h
+++ b/src/util/driconf.h
@@ -677,6 +677,10 @@
    DRI_CONF_OPT_I(query_clear_with_blorp_threshold, def, 0, INT32_MAX, \
                   "Indirect threshold count above which we start generating commands")
 
+#define DRI_CONF_ANV_FORCE_INDIRECT_DESCRIPTORS(def) \
+   DRI_CONF_OPT_B(force_indirect_descriptors, def, \
+                  "Use an indirection to access buffer/image/texture/sampler handles")
+
 /**
  * \brief DZN specific configuration options
  */