#include <llvm-c/Support.h>
#include "sid.h"
-#include "gfx9d.h"
#include "ac_binary.h"
#include "ac_llvm_util.h"
#include "ac_nir_to_llvm.h"
.lower_device_index_to_zero = true,
.lower_fsat = true,
.lower_fdiv = true,
+ .lower_bitfield_insert_to_bitfield_select = true,
+ .lower_bitfield_extract = true,
.lower_sub = true,
.lower_pack_snorm_2x16 = true,
.lower_pack_snorm_4x8 = true,
.lower_ffma = true,
.lower_fpow = true,
.lower_mul_2x32_64 = true,
+ .lower_rotate = true,
.max_unroll_iterations = 32
};
const struct radv_pipeline_layout *layout)
{
nir_shader *nir;
- nir_function *entry_point;
if (module->nir) {
/* Some things such as our meta clear/blit code will give us a NIR
* shader directly. In that case, we just ignore the SPIR-V entirely
nir_validate_shader(nir, "in internal shader");
assert(exec_list_length(&nir->functions) == 1);
- struct exec_node *node = exec_list_get_head(&nir->functions);
- entry_point = exec_node_data(nir_function, node, node);
} else {
uint32_t *spirv = (uint32_t *) module->data;
assert(module->size % 4 == 0);
const struct spirv_to_nir_options spirv_options = {
.lower_ubo_ssbo_access_to_offsets = true,
.caps = {
+ .amd_gcn_shader = true,
+ .amd_shader_ballot = device->instance->perftest_flags & RADV_PERFTEST_SHADER_BALLOT,
+ .amd_trinary_minmax = true,
.derivative_group = true,
.descriptor_array_dynamic_indexing = true,
.descriptor_array_non_uniform_indexing = true,
.draw_parameters = true,
.float16 = true,
.float64 = true,
- .gcn_shader = true,
.geometry_streams = true,
.image_read_without_format = true,
.image_write_without_format = true,
.subgroup_vote = true,
.tessellation = true,
.transform_feedback = true,
- .trinary_minmax = true,
.variable_pointers = true,
},
.ubo_addr_format = nir_address_format_32bit_index_offset,
.push_const_addr_format = nir_address_format_logical,
.shared_addr_format = nir_address_format_32bit_offset,
};
- entry_point = spirv_to_nir(spirv, module->size / 4,
- spec_entries, num_spec_entries,
- stage, entrypoint_name,
- &spirv_options, &nir_options);
- nir = entry_point->shader;
+ nir = spirv_to_nir(spirv, module->size / 4,
+ spec_entries, num_spec_entries,
+ stage, entrypoint_name,
+ &spirv_options, &nir_options);
assert(nir->info.stage == stage);
nir_validate_shader(nir, "after spirv_to_nir");
/* Pick off the single entrypoint that we want */
foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
- if (func != entry_point)
+ if (func->is_entrypoint)
+ func->name = ralloc_strdup(func, "main");
+ else
exec_node_remove(&func->node);
}
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
/* Vulkan uses the separate-shader linking model */
nir->info.separate_shader = true;
- nir_shader_gather_info(nir, entry_point->impl);
+ nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
static const nir_lower_tex_options tex_options = {
.lower_txp = ~0,
static void
radv_fill_shader_variant(struct radv_device *device,
struct radv_shader_variant *variant,
+ struct radv_nir_compiler_options *options,
struct ac_shader_binary *binary,
gl_shader_stage stage)
{
switch (stage) {
case MESA_SHADER_TESS_EVAL:
- vgpr_comp_cnt = 3;
+ if (options->key.tes.as_es) {
+ assert(device->physical_device->rad_info.chip_class <= GFX8);
+ vgpr_comp_cnt = info->uses_prim_id ? 3 : 2;
+ } else {
+ bool enable_prim_id = options->key.tes.export_prim_id || info->uses_prim_id;
+ vgpr_comp_cnt = enable_prim_id ? 3 : 2;
+ }
variant->rsrc2 |= S_00B12C_OC_LDS_EN(1);
break;
case MESA_SHADER_TESS_CTRL:
if (device->physical_device->rad_info.chip_class >= GFX9) {
- vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
+ /* We need at least 2 components for LS.
+ * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
+ * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
+ */
+ vgpr_comp_cnt = info->vs.needs_instance_id ? 2 : 1;
} else {
variant->rsrc2 |= S_00B12C_OC_LDS_EN(1);
}
break;
case MESA_SHADER_VERTEX:
- case MESA_SHADER_GEOMETRY:
- vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
+ if (variant->info.vs.as_ls) {
+ assert(device->physical_device->rad_info.chip_class <= GFX8);
+ /* We need at least 2 components for LS.
+ * VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
+ * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
+ */
+ vgpr_comp_cnt = info->vs.needs_instance_id ? 2 : 1;
+ } else if (variant->info.vs.as_es) {
+ assert(device->physical_device->rad_info.chip_class <= GFX8);
+ /* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
+ vgpr_comp_cnt = info->vs.needs_instance_id ? 1 : 0;
+ } else {
+ /* VGPR0-3: (VertexID, InstanceID / StepRate0, PrimID, InstanceID)
+ * If PrimID is disabled. InstanceID / StepRate1 is loaded instead.
+ * StepRate0 is set to 1. so that VGPR3 doesn't have to be loaded.
+ */
+ if (options->key.vs.export_prim_id) {
+ vgpr_comp_cnt = 2;
+ } else if (info->vs.needs_instance_id) {
+ vgpr_comp_cnt = 1;
+ } else {
+ vgpr_comp_cnt = 0;
+ }
+ }
break;
case MESA_SHADER_FRAGMENT:
+ case MESA_SHADER_GEOMETRY:
break;
case MESA_SHADER_COMPUTE:
variant->rsrc2 |=
unsigned gs_vgpr_comp_cnt, es_vgpr_comp_cnt;
if (es_type == MESA_SHADER_VERTEX) {
- es_vgpr_comp_cnt = variant->info.vs.vgpr_comp_cnt;
+ /* VGPR0-3: (VertexID, InstanceID / StepRate0, ...) */
+ es_vgpr_comp_cnt = info->vs.needs_instance_id ? 1 : 0;
} else if (es_type == MESA_SHADER_TESS_EVAL) {
- es_vgpr_comp_cnt = 3;
+ es_vgpr_comp_cnt = info->uses_prim_id ? 3 : 2;
} else {
unreachable("invalid shader ES type");
}
radv_destroy_llvm_compiler(&ac_llvm, thread_compiler);
- radv_fill_shader_variant(device, variant, &binary, stage);
+ radv_fill_shader_variant(device, variant, options, &binary, stage);
if (code_out) {
*code_out = binary.code;