anv: add a pass to partially lower resource_intel

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>

src/intel/vulkan/anv_nir.h

@@ -84,6 +84,12 @@ void anv_nir_compute_push_layout(nir_shader *nir,
 void anv_nir_validate_push_layout(struct brw_stage_prog_data *prog_data,
                                   struct anv_pipeline_bind_map *map);
 
+bool anv_nir_update_resource_intel_block(nir_shader *shader);
+
+bool anv_nir_lower_resource_intel(nir_shader *shader,
+                                  const struct anv_physical_device *device,
+                                  enum anv_descriptor_set_layout_type desc_type);
+
 bool anv_nir_add_base_work_group_id(nir_shader *shader);
 
 uint32_t anv_nir_compute_used_push_descriptors(nir_shader *shader,

src/intel/vulkan/anv_nir_compute_push_layout.c

@@ -55,7 +55,7 @@ anv_nir_compute_push_layout(nir_shader *nir,
             nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
             switch (intrin->intrinsic) {
             case nir_intrinsic_load_ubo:
-               if (nir_src_is_const(intrin->src[0]) &&
+               if (brw_nir_ubo_surface_index_is_pushable(intrin->src[0]) &&
                    nir_src_is_const(intrin->src[1]))
                   has_const_ubo = true;
                break;

src/intel/vulkan/anv_nir_lower_resource_intel.c

@@ -0,0 +1,203 @@
+/*
+ * Copyright © 2022 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "anv_nir.h"
+#include "nir_builder.h"
+
+/* This pass updates the block index in the resource_intel intrinsics if the
+ * array index is constant.
+ *
+ * This pass must be run before anv_nir_compute_push_layout().
+ */
+static bool
+update_resource_intel_block(nir_builder *b, nir_instr *instr, UNUSED void *data)
+{
+   if (instr->type != nir_instr_type_intrinsic)
+      return false;
+
+   nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+   if (intrin->intrinsic != nir_intrinsic_resource_intel)
+      return false;
+
+   /* If the array index in the descriptor binding is not const, we won't be
+    * able to turn this load_ubo into a push constant.
+    *
+    * Also if not pushable, set the block to 0xffffffff.
+    *
+    * Otherwise we need to update the block index by adding the array index so
+    * that when anv_nir_compute_push_layout() uses the block value it uses the
+    * right surface in the array of the binding.
+    */
+   if (!nir_src_is_const(intrin->src[2]) ||
+       !(nir_intrinsic_resource_access_intel(intrin) &
+         nir_resource_intel_pushable)) {
+      nir_intrinsic_set_resource_block_intel(intrin, 0xffffffff);
+      nir_intrinsic_set_resource_access_intel(
+         intrin,
+         nir_intrinsic_resource_access_intel(intrin) &
+         ~nir_resource_intel_pushable);
+   } else {
+      nir_intrinsic_set_resource_block_intel(
+         intrin,
+         nir_intrinsic_resource_block_intel(intrin) +
+         nir_src_as_uint(intrin->src[2]));
+   }
+
+   return true;
+}
+
+bool
+anv_nir_update_resource_intel_block(nir_shader *shader)
+{
+   return nir_shader_instructions_pass(shader, update_resource_intel_block,
+                                       nir_metadata_all,
+                                       NULL);
+}
+
+static bool
+intrinsic_dont_need_rewrite(nir_intrinsic_instr *instr)
+{
+   switch (instr->intrinsic) {
+   case nir_intrinsic_load_ubo:
+   case nir_intrinsic_load_ssbo:
+   case nir_intrinsic_store_ssbo:
+      return true;
+   case nir_intrinsic_image_load:
+   case nir_intrinsic_image_store:
+   case nir_intrinsic_image_atomic:
+   case nir_intrinsic_image_atomic_swap:
+   case nir_intrinsic_image_size:
+   case nir_intrinsic_image_load_raw_intel:
+   case nir_intrinsic_image_store_raw_intel:
+   case nir_intrinsic_image_samples:
+   case nir_intrinsic_bindless_image_load:
+   case nir_intrinsic_bindless_image_store:
+   case nir_intrinsic_bindless_image_atomic:
+   case nir_intrinsic_bindless_image_atomic_swap:
+   case nir_intrinsic_bindless_image_size:
+      return true;
+
+   default:
+      return false;
+   }
+}
+
+struct lower_resource_state {
+   enum anv_descriptor_set_layout_type desc_type;
+   const struct anv_physical_device *device;
+};
+
+/* This pass lower resource_intel surface_index source, combining the
+ * descriptor set offset with the surface offset in the descriptor set.
+ *
+ * This pass must be run after anv_nir_compute_push_layout() because we want
+ * the push constant selection to tell if the surface offset is constant. Once
+ * combined the constant detection does not work anymore.
+ */
+static bool
+lower_resource_intel(nir_builder *b, nir_instr *instr, void *data)
+{
+   if (instr->type != nir_instr_type_intrinsic)
+      return false;
+
+   nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+   if (intrin->intrinsic != nir_intrinsic_resource_intel)
+      return false;
+
+   const bool is_bindless =
+      (nir_intrinsic_resource_access_intel(intrin) &
+       nir_resource_intel_bindless) != 0;
+   const bool is_sampler =
+      (nir_intrinsic_resource_access_intel(intrin) &
+       nir_resource_intel_sampler) != 0;
+   const struct lower_resource_state *state = data;
+
+   if (!is_bindless)
+      return true;
+
+   b->cursor = nir_before_instr(instr);
+
+   nir_ssa_def *set_offset = intrin->src[0].ssa;
+   nir_ssa_def *binding_offset = intrin->src[1].ssa;
+
+   /* When using indirect descriptor, the surface handles are loaded from the
+    * descriptor buffer and do not need any offset.
+    */
+   if (state->desc_type == ANV_PIPELINE_DESCRIPTOR_SET_LAYOUT_TYPE_DIRECT) {
+      if (!state->device->uses_ex_bso) {
+         /* 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,
+          * etc...) so it's worth optimizing that case.
+          *
+          * Surface handles in the extended descriptor message have to be
+          * shifted left by 6 prior to ex_bso (bits 31:12 in extended
+          * descriptor, match bits 25:6 of the surface handle). We have to
+          * combine 2 parts in the shader to build the final surface handle,
+          * base offset of the descriptor set (in the push constant, located
+          * in resource_intel::src[0]) and the relative descriptor offset
+          * (resource_intel::src[1]).
+          *
+          * For convenience, up to here, resource_intel::src[1] is in bytes.
+          * We now have to shift it left by 6 to match the shifted left by 6
+          * done for the push constant value provided in
+          * resource_intel::src[0]. That way the shader can just do a single
+          * ADD and get the surface handle.
+          *
+          * Samplers have a 4Gb heap and in the message they're in bits 31:6
+          * of the component 3 of the sampler message header. But since we
+          * push only a single offset for the base offset of the descriptor
+          * set, resource_intel::src[0] has to be shifted right by 6 (bringing
+          * it back in bytes).
+          */
+         if (is_sampler)
+            set_offset = nir_ushr_imm(b, set_offset, 6);
+         else
+            binding_offset = nir_ishl_imm(b, binding_offset, 6);
+      }
+
+      nir_instr_rewrite_src_ssa(instr, &intrin->src[1],
+                                nir_iadd(b, set_offset, binding_offset));
+   }
+
+   /* Now unused values : set offset, array index */
+   nir_instr_rewrite_src_ssa(instr, &intrin->src[0], nir_imm_int(b, 0xdeaddeed));
+   nir_instr_rewrite_src_ssa(instr, &intrin->src[2], nir_imm_int(b, 0xdeaddeed));
+
+   return true;
+}
+
+bool
+anv_nir_lower_resource_intel(nir_shader *shader,
+                             const struct anv_physical_device *device,
+                             enum anv_descriptor_set_layout_type desc_type)
+{
+   struct lower_resource_state state = {
+      .desc_type = desc_type,
+      .device = device,
+   };
+   return nir_shader_instructions_pass(shader, lower_resource_intel,
+                                       nir_metadata_block_index |
+                                       nir_metadata_dominance,
+                                       &state);
+}

src/intel/vulkan/anv_nir_push_descriptor_analysis.c

@@ -23,6 +23,8 @@
 
 #include "anv_nir.h"
 
+#include "compiler/brw_nir.h"
+
 const struct anv_descriptor_set_layout *
 anv_pipeline_layout_get_push_set(const struct anv_pipeline_sets_layout *layout,
                                  uint8_t *set_idx)
@@ -191,12 +193,11 @@ anv_nir_push_desc_ubo_fully_promoted(nir_shader *nir,
             if (intrin->intrinsic != nir_intrinsic_load_ubo)
                continue;
 
-            const nir_const_value *const_bt_idx =
-               nir_src_as_const_value(intrin->src[0]);
-            if (const_bt_idx == NULL)
+            if (!brw_nir_ubo_surface_index_is_pushable(intrin->src[0]))
                continue;
 
-            const unsigned bt_idx = const_bt_idx[0].u32;
+            const unsigned bt_idx =
+               brw_nir_ubo_surface_index_get_bti(intrin->src[0]);
 
             /* Skip if this isn't a load from push descriptor buffer. */
             const struct anv_pipeline_binding *binding =

src/intel/vulkan/anv_pipeline.c

@@ -1025,8 +1025,14 @@ anv_pipeline_lower_nir(struct anv_pipeline *pipeline,
                   .types = lower_non_uniform_access_types,
                   .callback = NULL,
                });
+
+      NIR_PASS(_, nir, brw_nir_lower_non_uniform_resource_intel);
+      NIR_PASS(_, nir, brw_nir_cleanup_resource_intel);
+      NIR_PASS(_, nir, nir_opt_dce);
    }
 
+   NIR_PASS_V(nir, anv_nir_update_resource_intel_block);
+
    stage->dynamic_push_values = anv_nir_compute_dynamic_push_bits(nir);
 
    NIR_PASS_V(nir, anv_nir_compute_push_layout,
@@ -1034,6 +1040,9 @@ anv_pipeline_lower_nir(struct anv_pipeline *pipeline,
               anv_graphics_pipeline_stage_fragment_dynamic(stage),
               prog_data, &stage->bind_map, mem_ctx);
 
+   NIR_PASS_V(nir, anv_nir_lower_resource_intel, pdevice,
+              pipeline->layout.type);
+
    if (gl_shader_stage_uses_workgroup(nir->info.stage)) {
       if (!nir->info.shared_memory_explicit_layout) {
          NIR_PASS(_, nir, nir_lower_vars_to_explicit_types,

src/intel/vulkan/meson.build

@@ -170,6 +170,7 @@ libanv_files = files(
   'anv_nir_lower_multiview.c',
   'anv_nir_lower_load_patch_vertices_in.c',
   'anv_nir_lower_ubo_loads.c',
+  'anv_nir_lower_resource_intel.c',
   'anv_nir_push_descriptor_analysis.c',
   'anv_perf.c',
   'anv_pipeline.c',