.lower_fpow = true,
.lower_mul_2x32_64 = true,
.lower_rotate = true,
+ .use_scoped_barrier = true,
.max_unroll_iterations = 32,
.use_interpolated_input_intrinsics = true,
/* nir_lower_int64() isn't actually called for the LLVM backend, but
nir_lower_divmod64 |
nir_lower_minmax64 |
nir_lower_iabs64,
+ .lower_doubles_options = nir_lower_drcp |
+ nir_lower_dsqrt |
+ nir_lower_drsq |
+ nir_lower_ddiv,
};
static const struct nir_shader_compiler_options nir_options_aco = {
.lower_fpow = true,
.lower_mul_2x32_64 = true,
.lower_rotate = true,
+ .use_scoped_barrier = true,
.max_unroll_iterations = 32,
.use_interpolated_input_intrinsics = true,
.lower_int64_options = nir_lower_imul64 |
nir_lower_imul_high64 |
nir_lower_imul_2x32_64 |
nir_lower_divmod64 |
- nir_lower_logic64 |
nir_lower_minmax64 |
nir_lower_iabs64,
+ .lower_doubles_options = nir_lower_drcp |
+ nir_lower_dsqrt |
+ nir_lower_drsq |
+ nir_lower_ddiv,
};
bool
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
assert(pCreateInfo->flags == 0);
- module = vk_alloc2(&device->alloc, pAllocator,
+ module = vk_alloc2(&device->vk.alloc, pAllocator,
sizeof(*module) + pCreateInfo->codeSize, 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (module == NULL)
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+ vk_object_base_init(&device->vk, &module->base,
+ VK_OBJECT_TYPE_SHADER_MODULE);
+
module->nir = NULL;
module->size = pCreateInfo->codeSize;
memcpy(module->data, pCreateInfo->pCode, module->size);
if (!module)
return;
- vk_free2(&device->alloc, pAllocator, module);
+ vk_object_base_finish(&module->base);
+ vk_free2(&device->vk.alloc, pAllocator, module);
}
void
NIR_PASS(progress, shader, nir_opt_copy_prop_vars);
NIR_PASS(progress, shader, nir_opt_dead_write_vars);
NIR_PASS(progress, shader, nir_remove_dead_variables,
- nir_var_function_temp | nir_var_shader_in | nir_var_shader_out);
+ nir_var_function_temp | nir_var_shader_in | nir_var_shader_out,
+ NULL);
NIR_PASS_V(shader, nir_lower_alu_to_scalar, NULL, NULL);
NIR_PASS_V(shader, nir_lower_phis_to_scalar);
}
NIR_PASS(progress, shader, nir_opt_undef);
+ NIR_PASS(progress, shader, nir_opt_shrink_vectors);
if (shader->options->max_unroll_iterations) {
NIR_PASS(progress, shader, nir_opt_loop_unroll, 0);
}
} while (progress && !optimize_conservatively);
NIR_PASS(progress, shader, nir_opt_conditional_discard);
- NIR_PASS(progress, shader, nir_opt_shrink_load);
NIR_PASS(progress, shader, nir_opt_move, nir_move_load_ubo);
}
*align = comp_size;
}
+struct radv_shader_debug_data {
+ struct radv_device *device;
+ const struct radv_shader_module *module;
+};
+
+static void radv_spirv_nir_debug(void *private_data,
+ enum nir_spirv_debug_level level,
+ size_t spirv_offset,
+ const char *message)
+{
+ struct radv_shader_debug_data *debug_data = private_data;
+ struct radv_instance *instance = debug_data->device->instance;
+
+ static const VkDebugReportFlagsEXT vk_flags[] = {
+ [NIR_SPIRV_DEBUG_LEVEL_INFO] = VK_DEBUG_REPORT_INFORMATION_BIT_EXT,
+ [NIR_SPIRV_DEBUG_LEVEL_WARNING] = VK_DEBUG_REPORT_WARNING_BIT_EXT,
+ [NIR_SPIRV_DEBUG_LEVEL_ERROR] = VK_DEBUG_REPORT_ERROR_BIT_EXT,
+ };
+ char buffer[256];
+
+ snprintf(buffer, sizeof(buffer), "SPIR-V offset %lu: %s",
+ (unsigned long)spirv_offset, message);
+
+ vk_debug_report(&instance->debug_report_callbacks,
+ vk_flags[level],
+ VK_DEBUG_REPORT_OBJECT_TYPE_SHADER_MODULE_EXT,
+ (uint64_t)(uintptr_t)debug_data->module,
+ 0, 0, "radv", buffer);
+}
+
+static void radv_compiler_debug(void *private_data,
+ enum radv_compiler_debug_level level,
+ const char *message)
+{
+ struct radv_shader_debug_data *debug_data = private_data;
+ struct radv_instance *instance = debug_data->device->instance;
+
+ static const VkDebugReportFlagsEXT vk_flags[] = {
+ [RADV_COMPILER_DEBUG_LEVEL_PERFWARN] = VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT,
+ [RADV_COMPILER_DEBUG_LEVEL_ERROR] = VK_DEBUG_REPORT_ERROR_BIT_EXT,
+ };
+
+ /* VK_DEBUG_REPORT_DEBUG_BIT_EXT specifies diagnostic information
+ * from the implementation and layers.
+ */
+ vk_debug_report(&instance->debug_report_callbacks,
+ vk_flags[level] | VK_DEBUG_REPORT_DEBUG_BIT_EXT,
+ VK_DEBUG_REPORT_OBJECT_TYPE_SHADER_MODULE_EXT,
+ (uint64_t)(uintptr_t)debug_data->module,
+ 0, 0, "radv", message);
+}
+
nir_shader *
radv_shader_compile_to_nir(struct radv_device *device,
struct radv_shader_module *module,
{
nir_shader *nir;
const nir_shader_compiler_options *nir_options =
- device->physical_device->use_aco ? &nir_options_aco :
- &nir_options_llvm;
+ radv_use_llvm_for_stage(device, stage) ? &nir_options_llvm : &nir_options_aco;
if (module->nir) {
/* Some things such as our meta clear/blit code will give us a NIR
struct nir_spirv_specialization *spec_entries = NULL;
if (spec_info && spec_info->mapEntryCount > 0) {
num_spec_entries = spec_info->mapEntryCount;
- spec_entries = malloc(num_spec_entries * sizeof(*spec_entries));
+ spec_entries = calloc(num_spec_entries, sizeof(*spec_entries));
for (uint32_t i = 0; i < num_spec_entries; i++) {
VkSpecializationMapEntry entry = spec_info->pMapEntries[i];
const void *data = spec_info->pData + entry.offset;
assert(data + entry.size <= spec_info->pData + spec_info->dataSize);
spec_entries[i].id = spec_info->pMapEntries[i].constantID;
- if (spec_info->dataSize == 8)
- spec_entries[i].data64 = *(const uint64_t *)data;
- else
- spec_entries[i].data32 = *(const uint32_t *)data;
+ switch (entry.size) {
+ case 8:
+ spec_entries[i].value.u64 = *(const uint64_t *)data;
+ break;
+ case 4:
+ spec_entries[i].value.u32 = *(const uint32_t *)data;
+ break;
+ case 2:
+ spec_entries[i].value.u16 = *(const uint16_t *)data;
+ break;
+ case 1:
+ spec_entries[i].value.u8 = *(const uint8_t *)data;
+ break;
+ default:
+ assert(!"Invalid spec constant size");
+ break;
+ }
}
}
+
+ struct radv_shader_debug_data spirv_debug_data = {
+ .device = device,
+ .module = module,
+ };
const struct spirv_to_nir_options spirv_options = {
.lower_ubo_ssbo_access_to_offsets = true,
.caps = {
.amd_fragment_mask = true,
.amd_gcn_shader = true,
+ .amd_image_gather_bias_lod = true,
.amd_image_read_write_lod = true,
- .amd_shader_ballot = device->physical_device->use_shader_ballot,
+ .amd_shader_ballot = true,
.amd_shader_explicit_vertex_parameter = true,
.amd_trinary_minmax = true,
- .demote_to_helper_invocation = device->physical_device->use_aco,
+ .demote_to_helper_invocation = true,
.derivative_group = true,
.descriptor_array_dynamic_indexing = true,
.descriptor_array_non_uniform_indexing = true,
.device_group = true,
.draw_parameters = true,
.float_controls = true,
- .float16 = !device->physical_device->use_aco,
+ .float16 = device->physical_device->rad_info.has_packed_math_16bit,
+ .float32_atomic_add = true,
.float64 = true,
.geometry_streams = true,
.image_ms_array = true,
.image_read_without_format = true,
.image_write_without_format = true,
- .int8 = !device->physical_device->use_aco,
- .int16 = !device->physical_device->use_aco,
+ .int8 = true,
+ .int16 = true,
.int64 = true,
.int64_atomics = true,
+ .min_lod = true,
.multiview = true,
.physical_storage_buffer_address = true,
.post_depth_coverage = true,
.shader_clock = true,
.shader_viewport_index_layer = true,
.stencil_export = true,
- .storage_8bit = !device->physical_device->use_aco,
- .storage_16bit = !device->physical_device->use_aco,
+ .storage_8bit = true,
+ .storage_16bit = true,
.storage_image_ms = true,
.subgroup_arithmetic = true,
.subgroup_ballot = true,
.tessellation = true,
.transform_feedback = true,
.variable_pointers = true,
+ .vk_memory_model = true,
+ .vk_memory_model_device_scope = true,
},
.ubo_addr_format = nir_address_format_32bit_index_offset,
.ssbo_addr_format = nir_address_format_32bit_index_offset,
.push_const_addr_format = nir_address_format_logical,
.shared_addr_format = nir_address_format_32bit_offset,
.frag_coord_is_sysval = true,
+ .debug = {
+ .func = radv_spirv_nir_debug,
+ .private_data = &spirv_debug_data,
+ },
};
nir = spirv_to_nir(spirv, module->size / 4,
spec_entries, num_spec_entries,
NIR_PASS_V(nir, nir_lower_variable_initializers, nir_var_function_temp);
NIR_PASS_V(nir, nir_lower_returns);
NIR_PASS_V(nir, nir_inline_functions);
+ NIR_PASS_V(nir, nir_copy_prop);
NIR_PASS_V(nir, nir_opt_deref);
/* Pick off the single entrypoint that we want */
NIR_PASS_V(nir, nir_split_per_member_structs);
if (nir->info.stage == MESA_SHADER_FRAGMENT &&
- device->physical_device->use_aco)
+ !radv_use_llvm_for_stage(device, nir->info.stage))
NIR_PASS_V(nir, nir_lower_io_to_vector, nir_var_shader_out);
if (nir->info.stage == MESA_SHADER_FRAGMENT)
- NIR_PASS_V(nir, nir_lower_input_attachments, true);
+ NIR_PASS_V(nir, nir_lower_input_attachments,
+ &(nir_input_attachment_options) {
+ .use_fragcoord_sysval = true,
+ .use_layer_id_sysval = false,
+ });
NIR_PASS_V(nir, nir_remove_dead_variables,
- nir_var_shader_in | nir_var_shader_out | nir_var_system_value | nir_var_mem_shared);
+ nir_var_shader_in | nir_var_shader_out | nir_var_system_value | nir_var_mem_shared,
+ NULL);
NIR_PASS_V(nir, nir_propagate_invariant);
NIR_PASS_V(nir, nir_lower_system_values);
NIR_PASS_V(nir, nir_lower_clip_cull_distance_arrays);
- NIR_PASS_V(nir, radv_nir_lower_ycbcr_textures, layout);
+
if (device->instance->debug_flags & RADV_DEBUG_DISCARD_TO_DEMOTE)
NIR_PASS_V(nir, nir_lower_discard_to_demote);
+
+ nir_lower_doubles_options lower_doubles =
+ nir->options->lower_doubles_options;
+
+ if (device->physical_device->rad_info.chip_class == GFX6) {
+ /* GFX6 doesn't support v_floor_f64 and the precision
+ * of v_fract_f64 which is used to implement 64-bit
+ * floor is less than what Vulkan requires.
+ */
+ lower_doubles |= nir_lower_dfloor;
+ }
+
+ NIR_PASS_V(nir, nir_lower_doubles, NULL, lower_doubles);
}
/* Vulkan uses the separate-shader linking model */
nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
- if (nir->info.stage == MESA_SHADER_GEOMETRY &&
- device->physical_device->use_aco)
+ if (nir->info.stage == MESA_SHADER_GEOMETRY)
nir_lower_gs_intrinsics(nir, true);
static const nir_lower_tex_options tex_options = {
nir_split_var_copies(nir);
nir_lower_global_vars_to_local(nir);
- nir_remove_dead_variables(nir, nir_var_function_temp);
+ nir_remove_dead_variables(nir, nir_var_function_temp, NULL);
bool gfx7minus = device->physical_device->rad_info.chip_class <= GFX7;
nir_lower_subgroups(nir, &(struct nir_lower_subgroups_options) {
.subgroup_size = subgroup_size,
.lower_vote_eq_to_ballot = 1,
.lower_quad_broadcast_dynamic = 1,
.lower_quad_broadcast_dynamic_to_const = gfx7minus,
+ .lower_shuffle_to_swizzle_amd = 1,
});
nir_lower_load_const_to_scalar(nir);
if (!(flags & VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT))
radv_optimize_nir(nir, false, true);
+ /* call radv_nir_lower_ycbcr_textures() late as there might still be
+ * tex with undef texture/sampler before first optimization */
+ NIR_PASS_V(nir, radv_nir_lower_ycbcr_textures, layout);
+
/* We call nir_lower_var_copies() after the first radv_optimize_nir()
* to remove any copies introduced by nir_opt_find_array_copies().
*/
static nir_variable *
find_layer_in_var(nir_shader *nir)
{
- nir_foreach_variable(var, &nir->inputs) {
- if (var->data.location == VARYING_SLOT_LAYER) {
- return var;
- }
- }
-
nir_variable *var =
- nir_variable_create(nir, nir_var_shader_in, glsl_int_type(), "layer id");
+ nir_find_variable_with_location(nir, nir_var_shader_in, VARYING_SLOT_LAYER);
+ if (var != NULL)
+ return var;
+
+ var = nir_variable_create(nir, nir_var_shader_in, glsl_int_type(), "layer id");
var->data.location = VARYING_SLOT_LAYER;
var->data.interpolation = INTERP_MODE_FLAT;
return var;
}
/* We use layered rendering to implement multiview, which means we need to map
- * view_index to gl_Layer. The attachment lowering also uses needs to know the
- * layer so that it can sample from the correct layer. The code generates a
- * load from the layer_id sysval, but since we don't have a way to get at this
- * information from the fragment shader, we also need to lower this to the
- * gl_Layer varying. This pass lowers both to a varying load from the LAYER
- * slot, before lowering io, so that nir_assign_var_locations() will give the
- * LAYER varying the correct driver_location.
+ * view_index to gl_Layer. The code generates a load from the layer_id sysval,
+ * but since we don't have a way to get at this information from the fragment
+ * shader, we also need to lower this to the gl_Layer varying. This pass
+ * lowers both to a varying load from the LAYER slot, before lowering io, so
+ * that nir_assign_var_locations() will give the LAYER varying the correct
+ * driver_location.
*/
static bool
continue;
nir_intrinsic_instr *load = nir_instr_as_intrinsic(instr);
- if (load->intrinsic != nir_intrinsic_load_view_index &&
- load->intrinsic != nir_intrinsic_load_layer_id)
+ if (load->intrinsic != nir_intrinsic_load_view_index)
continue;
if (!layer)
radv_lower_fs_io(nir_shader *nir)
{
NIR_PASS_V(nir, lower_view_index);
- nir_assign_io_var_locations(&nir->inputs, &nir->num_inputs,
+ nir_assign_io_var_locations(nir, nir_var_shader_in, &nir->num_inputs,
MESA_SHADER_FRAGMENT);
NIR_PASS_V(nir, nir_lower_io, nir_var_shader_in, type_size_vec4, 0);
}
-void *
+static void *
radv_alloc_shader_memory(struct radv_device *device,
struct radv_shader_variant *shader)
{
}
offset = align_u64(s->bo_offset + s->code_size, 256);
}
- if (slab->size - offset >= shader->code_size) {
+ if (offset <= slab->size && slab->size - offset >= shader->code_size) {
shader->bo = slab->bo;
shader->bo_offset = offset;
list_addtail(&shader->slab_list, &slab->shaders);
mtx_unlock(&device->shader_slab_mutex);
struct radv_shader_slab *slab = calloc(1, sizeof(struct radv_shader_slab));
- slab->size = 256 * 1024;
+ slab->size = MAX2(256 * 1024, shader->code_size);
slab->bo = device->ws->buffer_create(device->ws, slab->size, 256,
RADEON_DOMAIN_VRAM,
RADEON_FLAG_NO_INTERPROCESS_SHARING |
(device->physical_device->rad_info.cpdma_prefetch_writes_memory ?
0 : RADEON_FLAG_READ_ONLY),
RADV_BO_PRIORITY_SHADER);
+ if (!slab->bo) {
+ free(slab);
+ return NULL;
+ }
+
slab->ptr = (char*)device->ws->buffer_map(slab->bo);
+ if (!slab->ptr) {
+ device->ws->buffer_destroy(slab->bo);
+ free(slab);
+ return NULL;
+ }
+
list_inithead(&slab->shaders);
mtx_lock(&device->shader_slab_mutex);
*/
if (pdevice->rad_info.chip_class >= GFX10) {
vgpr_comp_cnt = info->vs.needs_instance_id ? 3 : 1;
+ config_out->rsrc2 |= S_00B42C_LDS_SIZE_GFX10(info->tcs.num_lds_blocks);
} else {
vgpr_comp_cnt = info->vs.needs_instance_id ? 2 : 1;
+ config_out->rsrc2 |= S_00B42C_LDS_SIZE_GFX9(info->tcs.num_lds_blocks);
}
} else {
config_out->rsrc2 |= S_00B12C_OC_LDS_EN(1);
}
config_out->rsrc1 |= S_00B128_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
- config_out->rsrc2 |= S_00B12C_SHARED_VGPR_CNT(num_shared_vgpr_blocks);
}
+ config_out->rsrc2 |= S_00B12C_SHARED_VGPR_CNT(num_shared_vgpr_blocks);
break;
case MESA_SHADER_FRAGMENT:
config_out->rsrc1 |= S_00B028_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
return NULL;
}
- if (!ac_rtld_read_config(&rtld_binary, &config)) {
+ if (!ac_rtld_read_config(&device->physical_device->rad_info,
+ &rtld_binary, &config)) {
ac_rtld_close(&rtld_binary);
free(variant);
return NULL;
radv_postprocess_config(device->physical_device, &config, &binary->info,
binary->stage, &variant->config);
- if (radv_device_use_secure_compile(device->instance)) {
+ void *dest_ptr = radv_alloc_shader_memory(device, variant);
+ if (!dest_ptr) {
if (binary->type == RADV_BINARY_TYPE_RTLD)
ac_rtld_close(&rtld_binary);
-
- return variant;
+ free(variant);
+ return NULL;
}
- void *dest_ptr = radv_alloc_shader_memory(device, variant);
-
if (binary->type == RADV_BINARY_TYPE_RTLD) {
struct radv_shader_binary_rtld* bin = (struct radv_shader_binary_rtld *)binary;
struct ac_rtld_upload_info info = {
enum radeon_family chip_family = device->physical_device->rad_info.family;
struct radv_shader_binary *binary = NULL;
+ struct radv_shader_debug_data debug_data = {
+ .device = device,
+ .module = module,
+ };
+
options->family = chip_family;
options->chip_class = device->physical_device->rad_info.chip_class;
options->dump_shader = radv_can_dump_shader(device, module, gs_copy_shader);
options->address32_hi = device->physical_device->rad_info.address32_hi;
options->has_ls_vgpr_init_bug = device->physical_device->rad_info.has_ls_vgpr_init_bug;
options->use_ngg_streamout = device->physical_device->use_ngg_streamout;
+ options->enable_mrt_output_nan_fixup = device->instance->enable_mrt_output_nan_fixup;
+ options->debug.func = radv_compiler_debug;
+ options->debug.private_data = &debug_data;
struct radv_shader_args args = {};
args.options = options;
shader_count >= 2 ? shaders[shader_count - 2]->info.stage
: MESA_SHADER_VERTEX);
- if (!device->physical_device->use_aco ||
+ if (radv_use_llvm_for_stage(device, stage) ||
options->dump_shader || options->record_ir)
ac_init_llvm_once();
- if (device->physical_device->use_aco) {
- aco_compile_shader(shader_count, shaders, &binary, &args);
- } else {
+ if (radv_use_llvm_for_stage(device, stage)) {
llvm_compile_shader(device, shader_count, shaders, &binary, &args);
+ } else {
+ aco_compile_shader(shader_count, shaders, &binary, &args);
}
binary->info = *info;
}
if (options->dump_shader) {
- fprintf(stderr, "disasm:\n%s\n", variant->disasm_string);
+ fprintf(stderr, "%s", radv_get_shader_name(info, shaders[0]->info.stage));
+ for (int i = 1; i < shader_count; ++i)
+ fprintf(stderr, " + %s", radv_get_shader_name(info, shaders[i]->info.stage));
+
+ fprintf(stderr, "\ndisasm:\n%s\n", variant->disasm_string);
}
bool keep_shader_info, bool keep_statistic_info,
struct radv_shader_binary **binary_out)
{
+ gl_shader_stage stage = shaders[shader_count - 1]->info.stage;
struct radv_nir_compiler_options options = {0};
options.layout = layout;
if (key)
options.key = *key;
- options.explicit_scratch_args = device->physical_device->use_aco;
+ options.explicit_scratch_args = !radv_use_llvm_for_stage(device, stage);
options.robust_buffer_access = device->robust_buffer_access;
- return shader_variant_compile(device, module, shaders, shader_count, shaders[shader_count - 1]->info.stage, info,
+ return shader_variant_compile(device, module, shaders, shader_count, stage, info,
&options, false, keep_shader_info, keep_statistic_info, binary_out);
}
bool multiview)
{
struct radv_nir_compiler_options options = {0};
+ gl_shader_stage stage = MESA_SHADER_VERTEX;
- options.explicit_scratch_args = device->physical_device->use_aco;
+ options.explicit_scratch_args = !radv_use_llvm_for_stage(device, stage);
options.key.has_multiview_view_index = multiview;
- return shader_variant_compile(device, NULL, &shader, 1, MESA_SHADER_VERTEX,
+ return shader_variant_compile(device, NULL, &shader, 1, stage,
info, &options, true, keep_shader_info, keep_statistic_info, binary_out);
}