anv, brw: Consolidate ex_bso bits to a static devinfo inline

If we have extended bindless surface offset (ExBSO) support, we want to
use it.  Consolidate the anv_physical_device and brw_compiler bits into
a single static inline that take devinfo.

Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Reviewed-by: Caio Oliveira <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39839>

src/intel/compiler/brw/brw_compiler.c

@@ -91,8 +91,6 @@ brw_compiler_create(void *mem_ctx, const struct intel_device_info *devinfo)
 
    compiler->precise_trig = debug_get_bool_option("INTEL_PRECISE_TRIG", false);
 
-   compiler->extended_bindless_surface_offset = devinfo->verx10 >= 125;
-
    compiler->lower_dpas = !devinfo->has_systolic ||
                           debug_get_bool_option("INTEL_LOWER_DPAS", false);
 

src/intel/compiler/brw/brw_compiler.h

@@ -73,13 +73,6 @@ struct brw_compiler {
     */
    bool precise_trig;
 
-   /**
-    * Gfx12.5+ has a bit in the SEND instruction extending the bindless
-    * surface offset range from 20 to 26 bits, effectively giving us 4Gb of
-    * bindless surface descriptors instead of 64Mb previously.
-    */
-   bool extended_bindless_surface_offset;
-
    /**
     * Gfx11+ has a bit in the dword 3 of the sampler message header that
     * indicates whether the sampler handle is relative to the dynamic state

src/intel/compiler/brw/brw_from_nir.cpp

@@ -6108,7 +6108,7 @@ brw_from_nir_emit_memory_access(nir_to_brw_state &ntb,
          const brw_builder ubld = bld.exec_all().group(8 * reg_unit(devinfo), 0);
          brw_reg bind = ubld.AND(retype(brw_vec1_grf(0, 5), BRW_TYPE_UD),
                                  brw_imm_ud(INTEL_MASK(31, 10)));
-         if (devinfo->ver >= 20 || bld.shader->compiler->extended_bindless_surface_offset)
+         if (devinfo->ver >= 20 || intel_has_extended_bindless(devinfo))
             bind = ubld.SHR(bind, brw_imm_ud(4));
 
          /* load_scratch / store_scratch cannot be is_scalar yet. */

src/intel/compiler/brw/brw_generator.cpp

@@ -1009,7 +1009,7 @@ brw_generator::generate_code(const brw_shader &s,
          generate_send(inst->as_send(), dst, src[SEND_SRC_DESC], src[SEND_SRC_EX_DESC],
                        src[SEND_SRC_PAYLOAD1], src[SEND_SRC_PAYLOAD2],
                        inst->as_send()->bindless_surface &&
-                       compiler->extended_bindless_surface_offset);
+                       intel_has_extended_bindless(devinfo));
          send_count++;
          break;
 
@@ -1018,7 +1018,7 @@ brw_generator::generate_code(const brw_shader &s,
                        src[SEND_GATHER_SRC_DESC], src[SEND_GATHER_SRC_EX_DESC],
                        src[SEND_GATHER_SRC_SCALAR], brw_null_reg(),
                        inst->as_send()->bindless_surface &&
-                       compiler->extended_bindless_surface_offset);
+                       intel_has_extended_bindless(devinfo));
          send_count++;
          break;
 

src/intel/compiler/brw/brw_lower_fill_spill.cpp

@@ -23,7 +23,7 @@ build_ex_desc(const brw_builder &bld, unsigned reg_size, bool unspill)
             brw_imm_ud(INTEL_MASK(31, 10)));
 
    const intel_device_info *devinfo = bld.shader->devinfo;
-   if (devinfo->verx10 >= 200 || bld.shader->compiler->extended_bindless_surface_offset) {
+   if (devinfo->ver >= 20 || intel_has_extended_bindless(devinfo)) {
       ubld.SHR(ex_desc, ex_desc, brw_imm_ud(4));
    } else {
       if (unspill) {

src/intel/compiler/brw/brw_lower_logical_sends.cpp

@@ -2529,8 +2529,7 @@ brw_lower_send_descriptors(brw_shader &s)
           (ex_desc_imm & INTEL_MASK(15, 12)) != 0)
          needs_addr_reg = true;
 
-      if (send->bindless_surface &&
-          s.compiler->extended_bindless_surface_offset) {
+      if (send->bindless_surface && intel_has_extended_bindless(devinfo)) {
          needs_addr_reg = true;
          /* When using the extended bindless offset, the whole extended
           * descriptor is the surface handle.

src/intel/dev/intel_device_info.h

@@ -214,6 +214,19 @@ enum intel_wa_steppings intel_device_info_wa_stepping(struct intel_device_info *
 uint32_t intel_device_info_get_max_slm_size(const struct intel_device_info *devinfo);
 uint32_t intel_device_info_get_max_preferred_slm_size(const struct intel_device_info *devinfo);
 
+/**
+ * True if this device supports the Extended Bindless Surface Offset mode,
+ * which offers 26-bit surface handles, instead of 20-bit.  This effectively
+ * gives us 4GB of bindless surface descriptors instead of only 64MB.
+ *
+ * On Gfx12.5 this is enabled via an "ExBSO" bit in the SEND instruction.
+ */
+static inline bool
+intel_has_extended_bindless(const struct intel_device_info *devinfo)
+{
+   return devinfo->verx10 >= 125;
+}
+
 /**
  * Whether indirect UBO loads should use the sampler or go through the
  * data/constant cache.  For the sampler, UBO surface states have to be set

src/intel/vulkan/anv_descriptor_set.c

@@ -198,7 +198,7 @@ anv_direct_descriptor_data_for_type(const struct anv_physical_device *device,
           * table. On previous generations, they are only reachable through
           * the binding table.
           */
-         if (device->uses_ex_bso) {
+         if (intel_has_extended_bindless(&device->info)) {
             data &= ~(ANV_DESCRIPTOR_BTI_SURFACE_STATE |
                       ANV_DESCRIPTOR_BTI_SAMPLER_STATE);
          }
@@ -414,7 +414,7 @@ anv_descriptor_data_supports_bindless(const struct anv_physical_device *pdevice,
        * bindless offset, all push descriptors have to go through the binding
        * tables.
        */
-      if (!pdevice->uses_ex_bso &&
+      if (!intel_has_extended_bindless(&pdevice->info) &&
           (set_flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR)) {
          return data & (ANV_DESCRIPTOR_INDIRECT_ADDRESS_RANGE |
                         ANV_DESCRIPTOR_INDIRECT_SAMPLED_IMAGE |
@@ -1852,10 +1852,10 @@ anv_push_descriptor_set_init(struct anv_cmd_buffer *cmd_buffer,
       uint64_t push_base_address;
 
       if (layout->vk.flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT) {
-         push_stream = pdevice->uses_ex_bso ?
+         push_stream = intel_has_extended_bindless(&pdevice->info) ?
             &cmd_buffer->push_descriptor_buffer_stream :
             &cmd_buffer->surface_state_stream;
-         push_base_address = pdevice->uses_ex_bso ?
+         push_base_address = intel_has_extended_bindless(&pdevice->info) ?
             pdevice->va.push_descriptor_buffer_pool.addr :
             pdevice->va.internal_surface_state_pool.addr;
       } else {

src/intel/vulkan/anv_nir_apply_pipeline_layout.c

@@ -601,7 +601,7 @@ build_desc_address64(nir_builder *b, nir_def *set_idx, unsigned set_idx_imm,
       anv_load_driver_uniform(b, 1, desc_surface_offsets[set_idx_imm]);
    desc_offset = nir_iand_imm(b, desc_offset, ANV_DESCRIPTOR_SET_OFFSET_MASK);
    if (state->bind_map->layout_type == ANV_PIPELINE_DESCRIPTOR_SET_LAYOUT_TYPE_BUFFER &&
-       !state->pdevice->uses_ex_bso) {
+       !intel_has_extended_bindless(&state->pdevice->info)) {
       nir_def *bindless_base_offset =
          anv_load_driver_uniform(b, 1, surfaces_base_offset);
       desc_offset = nir_iadd(b, bindless_base_offset, desc_offset);

src/intel/vulkan/anv_nir_lower_resource_intel.c

@@ -122,7 +122,7 @@ lower_resource_intel(nir_builder *b, nir_intrinsic_instr *intrin, void *data)
     */
    if (state->desc_type == ANV_PIPELINE_DESCRIPTOR_SET_LAYOUT_TYPE_DIRECT ||
        state->desc_type == ANV_PIPELINE_DESCRIPTOR_SET_LAYOUT_TYPE_BUFFER) {
-      if (!state->device->uses_ex_bso) {
+      if (!intel_has_extended_bindless(&state->device->info)) {
          /* We're trying to reduce the number of instructions in the shaders
           * to compute surface handles. The assumption is that we're using
           * more surface handles than sampler handles (UBO, SSBO, images,

src/intel/vulkan/anv_physical_device.c

@@ -1675,7 +1675,7 @@ get_properties(const struct anv_physical_device *pdevice,
       props->allowSamplerImageViewPostSubmitCreation = true;
       props->descriptorBufferOffsetAlignment = ANV_SURFACE_STATE_SIZE;
 
-      if (pdevice->uses_ex_bso) {
+      if (intel_has_extended_bindless(devinfo)) {
          props->maxDescriptorBufferBindings = MAX_SETS;
          props->maxResourceDescriptorBufferBindings = MAX_SETS;
          props->maxSamplerDescriptorBufferBindings = MAX_SETS;
@@ -2780,10 +2780,8 @@ anv_physical_device_try_create(struct vk_instance *vk_instance,
 
    device->vk.pipeline_cache_import_ops = anv_cache_import_ops;
 
-   device->uses_ex_bso = device->info.verx10 >= 125;
-
    device->indirect_descriptors =
-      !device->uses_ex_bso ||
+      !intel_has_extended_bindless(&devinfo) ||
       driQueryOptionb(&instance->dri_options, "force_indirect_descriptors");
 
    device->alloc_aux_tt_mem =
@@ -2823,7 +2821,6 @@ anv_physical_device_try_create(struct vk_instance *vk_instance,
    }
    device->compiler->shader_debug_log = compiler_debug_log;
    device->compiler->shader_perf_log = compiler_perf_log;
-   device->compiler->extended_bindless_surface_offset = device->uses_ex_bso;
    device->compiler->use_bindless_sampler_offset = false;
    device->compiler->spilling_rate =
       driQueryOptioni(&instance->dri_options, "shader_spilling_rate");

src/intel/vulkan/anv_private.h

@@ -1523,7 +1523,6 @@ struct anv_physical_device {
     /* true if FCV optimization should be disabled. */
     bool                                        disable_fcv;
     /**/
-    bool                                        uses_ex_bso;
 
     bool                                        always_flush_cache;
 
@@ -1735,7 +1734,7 @@ anv_physical_device_bindless_heap_size(const struct anv_physical_device *device,
     * but we have some workarounds that require 2 heaps to overlap, so the
     * size is dictated by our VA allocation.
     */
-   return device->uses_ex_bso ?
+   return intel_has_extended_bindless(&device->info) ?
       (descriptor_buffer ?
        device->va.dynamic_visible_pool.size :
        device->va.bindless_surface_state_pool.size) :
@@ -2841,7 +2840,7 @@ anv_surface_state_to_handle(struct anv_physical_device *device,
     */
    assert(state.offset >= 0);
    uint32_t offset = state.offset;
-   if (device->uses_ex_bso) {
+   if (intel_has_extended_bindless(&device->info)) {
       assert(util_is_aligned(offset, 64));
       return offset;
    } else {

src/intel/vulkan/genX_cmd_buffer.c

@@ -3284,7 +3284,7 @@ compute_descriptor_set_surface_offset(const struct anv_cmd_buffer *cmd_buffer,
 {
    const struct anv_physical_device *device = cmd_buffer->device->physical;
 
-   if (device->uses_ex_bso) {
+   if (intel_has_extended_bindless(&device->info)) {
       int32_t buffer_index =
          pipe_state->descriptor_buffers[set_idx].buffer_index;
       uint64_t buffer_address =