Esempio n. 1
0
static nir_shader *
build_resolve_compute_shader(struct radv_device *dev, bool is_integer, int samples)
{
	nir_builder b;
	char name[64];
	nir_if *outer_if = NULL;
	const struct glsl_type *sampler_type = glsl_sampler_type(GLSL_SAMPLER_DIM_MS,
								 false,
								 false,
								 GLSL_TYPE_FLOAT);
	const struct glsl_type *img_type = glsl_sampler_type(GLSL_SAMPLER_DIM_2D,
							     false,
							     false,
							     GLSL_TYPE_FLOAT);
	snprintf(name, 64, "meta_resolve_cs-%d-%s", samples, is_integer ? "int" : "float");
	nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
	b.shader->info->name = ralloc_strdup(b.shader, name);
	b.shader->info->cs.local_size[0] = 16;
	b.shader->info->cs.local_size[1] = 16;
	b.shader->info->cs.local_size[2] = 1;

	nir_variable *input_img = nir_variable_create(b.shader, nir_var_uniform,
						      sampler_type, "s_tex");
	input_img->data.descriptor_set = 0;
	input_img->data.binding = 0;

	nir_variable *output_img = nir_variable_create(b.shader, nir_var_uniform,
						       img_type, "out_img");
	output_img->data.descriptor_set = 0;
	output_img->data.binding = 1;
	nir_ssa_def *invoc_id = nir_load_system_value(&b, nir_intrinsic_load_local_invocation_id, 0);
	nir_ssa_def *wg_id = nir_load_system_value(&b, nir_intrinsic_load_work_group_id, 0);
	nir_ssa_def *block_size = nir_imm_ivec4(&b,
						b.shader->info->cs.local_size[0],
						b.shader->info->cs.local_size[1],
						b.shader->info->cs.local_size[2], 0);

	nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);

	nir_intrinsic_instr *src_offset = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
	src_offset->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
	src_offset->num_components = 2;
	nir_ssa_dest_init(&src_offset->instr, &src_offset->dest, 2, 32, "src_offset");
	nir_builder_instr_insert(&b, &src_offset->instr);

	nir_intrinsic_instr *dst_offset = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
	dst_offset->src[0] = nir_src_for_ssa(nir_imm_int(&b, 8));
	dst_offset->num_components = 2;
	nir_ssa_dest_init(&dst_offset->instr, &dst_offset->dest, 2, 32, "dst_offset");
	nir_builder_instr_insert(&b, &dst_offset->instr);

	nir_ssa_def *img_coord = nir_iadd(&b, global_id, &src_offset->dest.ssa);
	/* do a txf_ms on each sample */
	nir_ssa_def *tmp;

	nir_tex_instr *tex = nir_tex_instr_create(b.shader, 2);
	tex->sampler_dim = GLSL_SAMPLER_DIM_MS;
	tex->op = nir_texop_txf_ms;
	tex->src[0].src_type = nir_tex_src_coord;
	tex->src[0].src = nir_src_for_ssa(img_coord);
	tex->src[1].src_type = nir_tex_src_ms_index;
	tex->src[1].src = nir_src_for_ssa(nir_imm_int(&b, 0));
	tex->dest_type = nir_type_float;
	tex->is_array = false;
	tex->coord_components = 2;
	tex->texture = nir_deref_var_create(tex, input_img);
	tex->sampler = NULL;

	nir_ssa_dest_init(&tex->instr, &tex->dest, 4, 32, "tex");
	nir_builder_instr_insert(&b, &tex->instr);

	tmp = &tex->dest.ssa;
	nir_variable *color =
		nir_local_variable_create(b.impl, glsl_vec4_type(), "color");

	if (!is_integer && samples > 1) {
		nir_tex_instr *tex_all_same = nir_tex_instr_create(b.shader, 1);
		tex_all_same->sampler_dim = GLSL_SAMPLER_DIM_MS;
		tex_all_same->op = nir_texop_samples_identical;
		tex_all_same->src[0].src_type = nir_tex_src_coord;
		tex_all_same->src[0].src = nir_src_for_ssa(img_coord);
		tex_all_same->dest_type = nir_type_float;
		tex_all_same->is_array = false;
		tex_all_same->coord_components = 2;
		tex_all_same->texture = nir_deref_var_create(tex_all_same, input_img);
		tex_all_same->sampler = NULL;

		nir_ssa_dest_init(&tex_all_same->instr, &tex_all_same->dest, 1, 32, "tex");
		nir_builder_instr_insert(&b, &tex_all_same->instr);

		nir_ssa_def *all_same = nir_ine(&b, &tex_all_same->dest.ssa, nir_imm_int(&b, 0));
		nir_if *if_stmt = nir_if_create(b.shader);
		if_stmt->condition = nir_src_for_ssa(all_same);
		nir_cf_node_insert(b.cursor, &if_stmt->cf_node);

		b.cursor = nir_after_cf_list(&if_stmt->then_list);
		for (int i = 1; i < samples; i++) {
			nir_tex_instr *tex_add = nir_tex_instr_create(b.shader, 2);
			tex_add->sampler_dim = GLSL_SAMPLER_DIM_MS;
			tex_add->op = nir_texop_txf_ms;
			tex_add->src[0].src_type = nir_tex_src_coord;
			tex_add->src[0].src = nir_src_for_ssa(img_coord);
			tex_add->src[1].src_type = nir_tex_src_ms_index;
			tex_add->src[1].src = nir_src_for_ssa(nir_imm_int(&b, i));
			tex_add->dest_type = nir_type_float;
			tex_add->is_array = false;
			tex_add->coord_components = 2;
			tex_add->texture = nir_deref_var_create(tex_add, input_img);
			tex_add->sampler = NULL;

			nir_ssa_dest_init(&tex_add->instr, &tex_add->dest, 4, 32, "tex");
			nir_builder_instr_insert(&b, &tex_add->instr);

			tmp = nir_fadd(&b, tmp, &tex_add->dest.ssa);
		}

		tmp = nir_fdiv(&b, tmp, nir_imm_float(&b, samples));
		nir_store_var(&b, color, tmp, 0xf);
		b.cursor = nir_after_cf_list(&if_stmt->else_list);
		outer_if = if_stmt;
	}
	nir_store_var(&b, color, &tex->dest.ssa, 0xf);

	if (outer_if)
		b.cursor = nir_after_cf_node(&outer_if->cf_node);

	nir_ssa_def *newv = nir_load_var(&b, color);
	nir_ssa_def *coord = nir_iadd(&b, global_id, &dst_offset->dest.ssa);
	nir_intrinsic_instr *store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_image_store);
	store->src[0] = nir_src_for_ssa(coord);
	store->src[1] = nir_src_for_ssa(nir_ssa_undef(&b, 1, 32));
	store->src[2] = nir_src_for_ssa(newv);
	store->variables[0] = nir_deref_var_create(store, output_img);
	nir_builder_instr_insert(&b, &store->instr);
	return b.shader;
}
static void
emit_indirect_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr,
                         nir_deref_var *deref, nir_deref *arr_parent,
                         int start, int end,
                         nir_ssa_def **dest, nir_ssa_def *src)
{
   assert(arr_parent->child &&
          arr_parent->child->deref_type == nir_deref_type_array);
   nir_deref_array *arr = nir_deref_as_array(arr_parent->child);
   assert(arr->deref_array_type == nir_deref_array_type_indirect);
   assert(arr->indirect.is_ssa);

   assert(start < end);
   if (start == end - 1) {
      /* Base case.  Just emit the load/store op */
      nir_deref_array direct = *arr;
      direct.deref_array_type = nir_deref_array_type_direct;
      direct.base_offset += start;
      direct.indirect = NIR_SRC_INIT;

      arr_parent->child = &direct.deref;
      emit_load_store(b, orig_instr, deref, &arr->deref, dest, src);
      arr_parent->child = &arr->deref;
   } else {
      int mid = start + (end - start) / 2;

      nir_ssa_def *then_dest, *else_dest;

      nir_if *if_stmt = nir_if_create(b->shader);
      if_stmt->condition = nir_src_for_ssa(nir_ilt(b, arr->indirect.ssa,
                                                      nir_imm_int(b, mid)));
      nir_cf_node_insert(b->cursor, &if_stmt->cf_node);

      b->cursor = nir_after_cf_list(&if_stmt->then_list);
      emit_indirect_load_store(b, orig_instr, deref, arr_parent,
                               start, mid, &then_dest, src);

      b->cursor = nir_after_cf_list(&if_stmt->else_list);
      emit_indirect_load_store(b, orig_instr, deref, arr_parent,
                               mid, end, &else_dest, src);

      b->cursor = nir_after_cf_node(&if_stmt->cf_node);

      if (src == NULL) {
         /* We're a load.  We need to insert a phi node */
         nir_phi_instr *phi = nir_phi_instr_create(b->shader);
         unsigned bit_size = then_dest->bit_size;
         nir_ssa_dest_init(&phi->instr, &phi->dest,
                           then_dest->num_components, bit_size, NULL);

         nir_phi_src *src0 = ralloc(phi, nir_phi_src);
         src0->pred = nir_cf_node_as_block(nir_if_last_then_node(if_stmt));
         src0->src = nir_src_for_ssa(then_dest);
         exec_list_push_tail(&phi->srcs, &src0->node);

         nir_phi_src *src1 = ralloc(phi, nir_phi_src);
         src1->pred = nir_cf_node_as_block(nir_if_last_else_node(if_stmt));
         src1->src = nir_src_for_ssa(else_dest);
         exec_list_push_tail(&phi->srcs, &src1->node);

         nir_builder_instr_insert(b, &phi->instr);
         *dest = &phi->dest.ssa;
      }
   }
}