.lower_scmp = true,
.lower_flrp32 = true,
.lower_flrp64 = true,
+ .lower_device_index_to_zero = true,
.lower_fsat = true,
.lower_fdiv = true,
.lower_sub = true,
.lower_extract_byte = true,
.lower_extract_word = true,
.lower_ffma = true,
+ .lower_fpow = true,
.vs_inputs_dual_locations = true,
.max_unroll_iterations = 32
};
} while (progress);
NIR_PASS(progress, shader, nir_opt_shrink_load);
+ NIR_PASS(progress, shader, nir_opt_move_load_ubo);
}
nir_shader *
}
const struct spirv_to_nir_options spirv_options = {
.caps = {
+ .device_group = true,
.draw_parameters = true,
.float64 = true,
.image_read_without_format = true,
.tessellation = true,
.int64 = true,
.multiview = true,
+ .subgroup_basic = true,
.variable_pointers = true,
+ .gcn_shader = true,
+ .trinary_minmax = true,
+ .shader_viewport_index_layer = true,
},
};
entry_point = spirv_to_nir(spirv, module->size / 4,
assert(exec_list_length(&nir->functions) == 1);
entry_point->name = ralloc_strdup(entry_point, "main");
+ /* Make sure we lower constant initializers on output variables so that
+ * nir_remove_dead_variables below sees the corresponding stores
+ */
+ NIR_PASS_V(nir, nir_lower_constant_initializers, nir_var_shader_out);
+
NIR_PASS_V(nir, nir_remove_dead_variables,
nir_var_shader_in | nir_var_shader_out | nir_var_system_value);
nir_shader_gather_info(nir, entry_point->impl);
- /* While it would be nice not to have this flag, we are constrained
- * by the reality that LLVM 5.0 doesn't have working VGPR indexing
- * on GFX9.
- */
- bool llvm_has_working_vgpr_indexing =
- device->physical_device->rad_info.chip_class <= VI;
-
- /* TODO: Indirect indexing of GS inputs is unimplemented.
- *
- * TCS and TES load inputs directly from LDS or offchip memory, so
- * indirect indexing is trivial.
- */
- nir_variable_mode indirect_mask = 0;
- if (nir->info.stage == MESA_SHADER_GEOMETRY ||
- (nir->info.stage != MESA_SHADER_TESS_CTRL &&
- nir->info.stage != MESA_SHADER_TESS_EVAL &&
- !llvm_has_working_vgpr_indexing)) {
- indirect_mask |= nir_var_shader_in;
- }
- if (!llvm_has_working_vgpr_indexing &&
- nir->info.stage != MESA_SHADER_TESS_CTRL)
- indirect_mask |= nir_var_shader_out;
-
- /* TODO: We shouldn't need to do this, however LLVM isn't currently
- * smart enough to handle indirects without causing excess spilling
- * causing the gpu to hang.
- *
- * See the following thread for more details of the problem:
- * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
- */
- indirect_mask |= nir_var_local;
-
- nir_lower_indirect_derefs(nir, indirect_mask);
-
static const nir_lower_tex_options tex_options = {
.lower_txp = ~0,
};
nir_lower_var_copies(nir);
nir_lower_global_vars_to_local(nir);
nir_remove_dead_variables(nir, nir_var_local);
+ nir_lower_subgroups(nir, &(struct nir_lower_subgroups_options) {
+ .subgroup_size = 64,
+ .ballot_bit_size = 64,
+ .lower_to_scalar = 1,
+ .lower_subgroup_masks = 1,
+ .lower_shuffle = 1,
+ .lower_quad = 1,
+ });
+
+ radv_optimize_nir(nir);
+
+ /* Indirect lowering must be called after the radv_optimize_nir() loop
+ * has been called at least once. Otherwise indirect lowering can
+ * bloat the instruction count of the loop and cause it to be
+ * considered too large for unrolling.
+ */
+ ac_lower_indirect_derefs(nir, device->physical_device->rad_info.chip_class);
radv_optimize_nir(nir);
return nir;
case MESA_SHADER_FRAGMENT:
break;
case MESA_SHADER_COMPUTE: {
- struct ac_shader_info *info = &variant->info.info;
+ struct radv_shader_info *info = &variant->info.info;
variant->rsrc2 |=
S_00B84C_TGID_X_EN(info->cs.uses_block_id[0]) |
S_00B84C_TGID_Y_EN(info->cs.uses_block_id[1]) |
if (device->physical_device->rad_info.chip_class >= GFX9 &&
stage == MESA_SHADER_GEOMETRY) {
- struct ac_shader_info *info = &variant->info.info;
+ struct radv_shader_info *info = &variant->info.info;
unsigned es_type = variant->info.gs.es_type;
unsigned gs_vgpr_comp_cnt, es_vgpr_comp_cnt;
struct nir_shader * const *shaders,
int shader_count,
gl_shader_stage stage,
- struct ac_nir_compiler_options *options,
+ struct radv_nir_compiler_options *options,
bool gs_copy_shader,
void **code_out,
unsigned *code_size_out)
{
enum radeon_family chip_family = device->physical_device->rad_info.family;
- bool dump_shaders = radv_can_dump_shader(device, module);
enum ac_target_machine_options tm_options = 0;
struct radv_shader_variant *variant;
struct ac_shader_binary binary;
options->family = chip_family;
options->chip_class = device->physical_device->rad_info.chip_class;
- options->dump_preoptir = radv_can_dump_shader(device, module) &&
+ options->dump_shader = radv_can_dump_shader(device, module);
+ options->dump_preoptir = options->dump_shader &&
device->instance->debug_flags & RADV_DEBUG_PREOPTIR;
+ options->record_llvm_ir = device->keep_shader_info;
+ options->tess_offchip_block_dw_size = device->tess_offchip_block_dw_size;
if (options->supports_spill)
tm_options |= AC_TM_SUPPORTS_SPILL;
if (gs_copy_shader) {
assert(shader_count == 1);
- ac_create_gs_copy_shader(tm, *shaders, &binary, &variant->config,
- &variant->info, options, dump_shaders);
+ radv_compile_gs_copy_shader(tm, *shaders, &binary,
+ &variant->config, &variant->info,
+ options);
} else {
- ac_compile_nir_shader(tm, &binary, &variant->config,
- &variant->info, shaders, shader_count, options,
- dump_shaders);
+ radv_compile_nir_shader(tm, &binary, &variant->config,
+ &variant->info, shaders, shader_count,
+ options);
}
LLVMDisposeTargetMachine(tm);
if (device->keep_shader_info) {
variant->disasm_string = binary.disasm_string;
+ variant->llvm_ir_string = binary.llvm_ir_string;
if (!gs_copy_shader && !module->nir) {
variant->nir = *shaders;
variant->spirv = (uint32_t *)module->data;
struct nir_shader *const *shaders,
int shader_count,
struct radv_pipeline_layout *layout,
- const struct ac_shader_variant_key *key,
+ const struct radv_shader_variant_key *key,
void **code_out,
unsigned *code_size_out)
{
- struct ac_nir_compiler_options options = {0};
+ struct radv_nir_compiler_options options = {0};
options.layout = layout;
if (key)
unsigned *code_size_out,
bool multiview)
{
- struct ac_nir_compiler_options options = {0};
+ struct radv_nir_compiler_options options = {0};
options.key.has_multiview_view_index = multiview;
ralloc_free(variant->nir);
free(variant->disasm_string);
+ free(variant->llvm_ir_string);
free(variant);
}
"VGPRS: %d\n"
"Spilled SGPRs: %d\n"
"Spilled VGPRs: %d\n"
+ "PrivMem VGPRS: %d\n"
"Code Size: %d bytes\n"
"LDS: %d blocks\n"
"Scratch: %d bytes per wave\n"
"Max Waves: %d\n"
"********************\n\n\n",
conf->num_sgprs, conf->num_vgprs,
- conf->spilled_sgprs, conf->spilled_vgprs, variant->code_size,
+ conf->spilled_sgprs, conf->spilled_vgprs,
+ variant->info.private_mem_vgprs, variant->code_size,
conf->lds_size, conf->scratch_bytes_per_wave,
max_simd_waves);
}