mesa/src/compiler/nir/nir_lower_multiview.c

/*
 * Copyright © 2016 Intel Corporation
 * Copyright © 2020 Valve 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 "nir_builder.h"
#include "nir_control_flow.h"

/**
 * This file implements an optimization for multiview. Some GPU's have a
 * special mode which allows the vertex shader (or last stage in the geometry
 * pipeline) to create multiple primitives in different layers of the
 * framebuffer at once by writing multiple copies of gl_Position. The
 * assumption is that in most uses of multiview, the only use of gl_ViewIndex
 * is to change the position to implement the parallax effect, and other
 * varyings will be the same between the different views. We put the body of
 * the original vertex shader in a loop, writing to a different copy of
 * gl_Position each loop iteration, and then let other optimizations clean up
 * the mess. On some hardware it is also possible to write different copies of
 * other varyings, expanding the set of shaders that the optimization is
 * usable for.
 *
 * Indexes in the per-view output arrays generated by this pass do not
 * correspond directly to gl_ViewIndex. Instead, they are compacted, dropping
 * disabled views. For example, with view mask 0b1010, gl_Position[0]
 * is view index 1, and gl_Position[1] is view index 3.
 */

static bool
shader_writes_to_memory(nir_shader *shader)
{
   /* With multiview, we would need to ensure that memory writes happen either
    * once or once per view. Since combination of multiview and memory writes
    * is not expected, we'll just skip this optimization in this case.
    */

   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);

   nir_foreach_block(block, entrypoint) {
      nir_foreach_instr(instr, block) {
         if (instr->type != nir_instr_type_intrinsic)
            continue;
         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);

         switch (intrin->intrinsic) {
         case nir_intrinsic_deref_atomic:
         case nir_intrinsic_deref_atomic_swap:
         case nir_intrinsic_store_ssbo:
         case nir_intrinsic_ssbo_atomic:
         case nir_intrinsic_ssbo_atomic_swap:
         case nir_intrinsic_store_shared:
         case nir_intrinsic_store_shared2_amd:
         case nir_intrinsic_shared_atomic:
         case nir_intrinsic_shared_atomic_swap:
         case nir_intrinsic_shared_append_amd:
         case nir_intrinsic_shared_consume_amd:
         case nir_intrinsic_task_payload_atomic:
         case nir_intrinsic_task_payload_atomic_swap:
         case nir_intrinsic_image_deref_store:
         case nir_intrinsic_image_deref_atomic:
         case nir_intrinsic_image_deref_atomic_swap:
            return true;

         default:
            /* Keep walking. */
            break;
         }
      }
   }

   return false;
}

bool
nir_shader_uses_view_index(nir_shader *shader)
{
   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);

   nir_foreach_block(block, entrypoint) {
      nir_foreach_instr(instr, block) {
         if (instr->type != nir_instr_type_intrinsic)
            continue;

         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
         if (intrin->intrinsic == nir_intrinsic_load_view_index)
            return true;
      }
   }

   return false;
}

static bool
shader_only_allowed_outputs_use_view_index(nir_shader *shader,
                                           uint64_t allowed_outputs)
{
   nir_shader *shader_no_position = nir_shader_clone(NULL, shader);
   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader_no_position);

   /* Remove stores to allowed outputs from a cloned shader. */
   nir_foreach_block(block, entrypoint) {
      nir_foreach_instr_safe(instr, block) {
         if (instr->type != nir_instr_type_intrinsic)
            continue;

         nir_intrinsic_instr *store = nir_instr_as_intrinsic(instr);
         if (store->intrinsic != nir_intrinsic_store_deref)
            continue;

         nir_variable *var = nir_intrinsic_get_var(store, 0);
         if (!(allowed_outputs & BITFIELD64_BIT(var->data.location)))
            continue;

         nir_instr_remove(&store->instr);
      }
   }

   /* Clean up shader so unused load_view_index intrinsics are removed. */
   bool progress;
   do {
      progress = false;
      progress |= nir_opt_dead_cf(shader_no_position);

      /* Peephole select will drop if-blocks that have then and else empty,
       * which will remove the usage of an SSA in the condition.
       */
      progress |= nir_opt_peephole_select(shader_no_position, 0, false, false);

      progress |= nir_opt_dce(shader_no_position);
   } while (progress);

   bool uses_view_index = nir_shader_uses_view_index(shader_no_position);

   ralloc_free(shader_no_position);
   return !uses_view_index;
}

/* Return true if it's safe to call nir_lower_multiview() on this vertex
 * shader. Note that this only handles driver-agnostic checks, i.e. things
 * which would make nir_lower_multiview() incorrect. Any driver-specific
 * checks, e.g. for sufficient varying space or performance considerations,
 * should be handled in the driver.
 *
 * Note that we don't handle the more complex checks needed for lowering
 * pipelines with geometry or tessellation shaders.
 */

bool
nir_can_lower_multiview(nir_shader *shader, nir_lower_multiview_options options)
{
   return !shader_writes_to_memory(shader) &&
          shader_only_allowed_outputs_use_view_index(
             shader, options.allowed_per_view_outputs);
}

/**
 * The lowering. Call with the last active geometry stage.
 */

bool
nir_lower_multiview(nir_shader *shader, nir_lower_multiview_options options)
{
   assert(shader->info.stage != MESA_SHADER_FRAGMENT);
   int view_count = util_bitcount(options.view_mask);

   shader->info.view_mask = options.view_mask;

   nir_function_impl *entrypoint = nir_shader_get_entrypoint(shader);

   /* Update per-view outputs to refer to arrays. */
   nir_foreach_shader_out_variable(var, shader) {
      if (options.allowed_per_view_outputs & BITFIELD64_BIT(var->data.location)) {
         var->type = glsl_array_type(var->type, view_count, 0);
         var->data.per_view = true;
         shader->info.per_view_outputs |= BITFIELD64_BIT(var->data.location);
      }
   }

   nir_cf_list body;
   nir_cf_list_extract(&body, &entrypoint->body);

   nir_builder b = nir_builder_at(nir_after_impl(entrypoint));

   /* Loop Index will go from 0 to view_count. */
   nir_variable *loop_index_var =
      nir_local_variable_create(entrypoint, glsl_uint_type(), "loop_index");
   nir_deref_instr *loop_index_deref = nir_build_deref_var(&b, loop_index_var);
   nir_store_deref(&b, loop_index_deref, nir_imm_int(&b, 0), 1);

   /* Array of view index values that are active in the loop.  Note that the
    * loop index only matches the view index if there are no gaps in the
    * view_mask.
    */
   nir_variable *view_index_var = nir_local_variable_create(
      entrypoint, glsl_array_type(glsl_uint_type(), view_count, 0), "view_index");
   nir_deref_instr *view_index_deref = nir_build_deref_var(&b, view_index_var);
   {
      int array_position = 0;
      uint32_t view_mask_temp = options.view_mask;
      while (view_mask_temp) {
         uint32_t view_index = u_bit_scan(&view_mask_temp);
         nir_store_deref(&b, nir_build_deref_array_imm(&b, view_index_deref, array_position),
                         nir_imm_int(&b, view_index), 1);
         array_position++;
      }
   }

   /* Create the equivalent of
    *
    *    while (true):
    *       if (loop_index >= view_count):
    *          break
    *
    *       view_index = active_indices[loop_index]
    *
    *       out1_deref = &out1[loop_index]
    *       out2_deref = &out2[loop_index]
    *       ...
    *
    *       # Placeholder for the body to be reinserted.
    *
    *       loop_index += 1
    *
    * Later both `view_index` and `outN_deref` will be used to rewrite the
    * original shader body.
    */

   nir_loop *loop = nir_push_loop(&b);

   nir_def *loop_index = nir_load_deref(&b, loop_index_deref);
   nir_def *cmp = nir_ige_imm(&b, loop_index, view_count);
   nir_if *loop_check = nir_push_if(&b, cmp);
   nir_jump(&b, nir_jump_break);
   nir_pop_if(&b, loop_check);

   nir_def *view_index =
      nir_load_deref(&b, nir_build_deref_array(&b, view_index_deref, loop_index));

   struct hash_table *out_derefs = _mesa_pointer_hash_table_create(NULL);
   nir_cursor body_cursor = b.cursor;
   nir_foreach_shader_out_variable(var, shader) {
      if (var->data.per_view) {
         nir_deref_instr *deref =
            nir_build_deref_array(&b, nir_build_deref_var(&b, var), loop_index);
         _mesa_hash_table_insert(out_derefs, var, (void *)deref);
         body_cursor = nir_after_instr(&deref->instr);
      }
   }

   nir_store_deref(&b, loop_index_deref, nir_iadd_imm(&b, loop_index, 1), 1);
   nir_pop_loop(&b, loop);

   /* Reinsert the body. */
   b.cursor = body_cursor;
   nir_cf_reinsert(&body, b.cursor);

   nir_foreach_block(block, entrypoint) {
      nir_foreach_instr_safe(instr, block) {
         if (instr->type != nir_instr_type_intrinsic)
            continue;

         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);

         switch (intrin->intrinsic) {
         case nir_intrinsic_load_view_index: {
            nir_def_rewrite_uses(&intrin->def, view_index);
            break;
         }

         case nir_intrinsic_store_deref: {
            nir_variable *var = nir_intrinsic_get_var(intrin, 0);
            struct hash_entry *entry = _mesa_hash_table_search(out_derefs, var);
            if (entry) {
               nir_deref_instr *new_deref = entry->data;
               nir_deref_instr *old_deref = nir_src_as_deref(intrin->src[0]);

               nir_src_rewrite(&intrin->src[0], &new_deref->def);

               /* Remove old deref since it has the wrong type. */
               nir_deref_instr_remove_if_unused(old_deref);
            }
            break;
         }

         case nir_intrinsic_load_deref: {
            nir_variable *var = nir_intrinsic_get_var(intrin, 0);
            if (_mesa_hash_table_search(out_derefs, var)) {
               unreachable("Should have lowered I/O to temporaries "
                           "so no load_deref on output is expected.");
            }
            break;
         }

         case nir_intrinsic_copy_deref:
            unreachable("Should have lowered copy_derefs at this point");
            break;

         default:
            /* Do nothing. */
            break;
         }
      }
   }

   _mesa_hash_table_destroy(out_derefs, NULL);

   nir_metadata_preserve(entrypoint, nir_metadata_none);
   return true;
}