.lower_fpow = true,
.lower_mul_2x32_64 = true,
.lower_rotate = true,
- .max_unroll_iterations = 32
+ .max_unroll_iterations = 32,
+ .use_interpolated_input_intrinsics = true,
};
VkResult radv_CreateShaderModule(
.int64_atomics = true,
.multiview = true,
.physical_storage_buffer_address = true,
+ .post_depth_coverage = true,
.runtime_descriptor_array = true,
.shader_viewport_index_layer = true,
.stencil_export = true,
.phys_ssbo_addr_format = nir_address_format_64bit_global,
.push_const_addr_format = nir_address_format_logical,
.shared_addr_format = nir_address_format_32bit_offset,
+ .frag_coord_is_sysval = true,
};
nir = spirv_to_nir(spirv, module->size / 4,
spec_entries, num_spec_entries,
NIR_PASS_V(nir, nir_split_var_copies);
NIR_PASS_V(nir, nir_split_per_member_structs);
+ if (nir->info.stage == MESA_SHADER_FRAGMENT)
+ NIR_PASS_V(nir, nir_lower_input_attachments, true);
+
NIR_PASS_V(nir, nir_remove_dead_variables,
nir_var_shader_in | nir_var_shader_out | nir_var_system_value);
nir_lower_vars_to_ssa(nir);
if (nir->info.stage == MESA_SHADER_VERTEX ||
- nir->info.stage == MESA_SHADER_GEOMETRY) {
+ nir->info.stage == MESA_SHADER_GEOMETRY ||
+ nir->info.stage == MESA_SHADER_FRAGMENT) {
NIR_PASS_V(nir, nir_lower_io_to_temporaries,
nir_shader_get_entrypoint(nir), true, true);
- } else if (nir->info.stage == MESA_SHADER_TESS_EVAL||
- nir->info.stage == MESA_SHADER_FRAGMENT) {
+ } else if (nir->info.stage == MESA_SHADER_TESS_EVAL) {
NIR_PASS_V(nir, nir_lower_io_to_temporaries,
nir_shader_get_entrypoint(nir), true, false);
}
return nir;
}
+static void mark_16bit_fs_input(struct radv_shader_variant_info *shader_info,
+ const struct glsl_type *type,
+ int location)
+{
+ if (glsl_type_is_scalar(type) || glsl_type_is_vector(type) || glsl_type_is_matrix(type)) {
+ unsigned attrib_count = glsl_count_attribute_slots(type, false);
+ if (glsl_type_is_16bit(type)) {
+ shader_info->fs.float16_shaded_mask |= ((1ull << attrib_count) - 1) << location;
+ }
+ } else if (glsl_type_is_array(type)) {
+ unsigned stride = glsl_count_attribute_slots(glsl_get_array_element(type), false);
+ for (unsigned i = 0; i < glsl_get_length(type); ++i) {
+ mark_16bit_fs_input(shader_info, glsl_get_array_element(type), location + i * stride);
+ }
+ } else {
+ assert(glsl_type_is_struct_or_ifc(type));
+ for (unsigned i = 0; i < glsl_get_length(type); i++) {
+ mark_16bit_fs_input(shader_info, glsl_get_struct_field(type, i), location);
+ location += glsl_count_attribute_slots(glsl_get_struct_field(type, i), false);
+ }
+ }
+}
+
+static void
+handle_fs_input_decl(struct radv_shader_variant_info *shader_info,
+ struct nir_variable *variable)
+{
+ unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
+
+ if (variable->data.compact) {
+ unsigned component_count = variable->data.location_frac +
+ glsl_get_length(variable->type);
+ attrib_count = (component_count + 3) / 4;
+ } else {
+ mark_16bit_fs_input(shader_info, variable->type,
+ variable->data.driver_location);
+ }
+
+ uint64_t mask = ((1ull << attrib_count) - 1);
+
+ if (variable->data.interpolation == INTERP_MODE_FLAT)
+ shader_info->fs.flat_shaded_mask |= mask << variable->data.driver_location;
+
+ if (variable->data.location >= VARYING_SLOT_VAR0)
+ shader_info->fs.input_mask |= mask << (variable->data.location - VARYING_SLOT_VAR0);
+}
+
+static int
+type_size_vec4(const struct glsl_type *type, bool bindless)
+{
+ return glsl_count_attribute_slots(type, false);
+}
+
+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");
+ 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.
+ */
+
+static bool
+lower_view_index(nir_shader *nir)
+{
+ bool progress = false;
+ nir_function_impl *entry = nir_shader_get_entrypoint(nir);
+ nir_builder b;
+ nir_builder_init(&b, entry);
+
+ nir_variable *layer = NULL;
+ nir_foreach_block(block, entry) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ 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)
+ continue;
+
+ if (!layer)
+ layer = find_layer_in_var(nir);
+
+ b.cursor = nir_before_instr(instr);
+ nir_ssa_def *def = nir_load_var(&b, layer);
+ nir_ssa_def_rewrite_uses(&load->dest.ssa,
+ nir_src_for_ssa(def));
+
+ nir_instr_remove(instr);
+ progress = true;
+ }
+ }
+
+ return progress;
+}
+
+/* Gather information needed to setup the vs<->ps linking registers in
+ * radv_pipeline_generate_ps_inputs().
+ */
+
+static void
+handle_fs_inputs(nir_shader *nir, struct radv_shader_variant_info *shader_info)
+{
+ shader_info->fs.num_interp = nir->num_inputs;
+
+ nir_foreach_variable(variable, &nir->inputs)
+ handle_fs_input_decl(shader_info, variable);
+}
+
+static void
+lower_fs_io(nir_shader *nir, struct radv_shader_variant_info *shader_info)
+{
+ NIR_PASS_V(nir, lower_view_index);
+ nir_assign_io_var_locations(&nir->inputs, &nir->num_inputs,
+ MESA_SHADER_FRAGMENT);
+
+ handle_fs_inputs(nir, shader_info);
+
+ NIR_PASS_V(nir, nir_lower_io, nir_var_shader_in, type_size_vec4, 0);
+
+ /* This pass needs actual constants */
+ nir_opt_constant_folding(nir);
+
+ NIR_PASS_V(nir, nir_io_add_const_offset_to_base, nir_var_shader_in);
+}
+
+
void *
radv_alloc_shader_memory(struct radv_device *device,
struct radv_shader_variant *shader)
config_out->float_mode |= V_00B028_FP_64_DENORMS;
config_out->rsrc2 = S_00B12C_USER_SGPR(info->num_user_sgprs) |
- S_00B12C_USER_SGPR_MSB_GFX9(info->num_user_sgprs >> 5) |
- S_00B12C_SCRATCH_EN(scratch_enabled) |
- S_00B12C_SO_BASE0_EN(!!info->info.so.strides[0]) |
- S_00B12C_SO_BASE1_EN(!!info->info.so.strides[1]) |
- S_00B12C_SO_BASE2_EN(!!info->info.so.strides[2]) |
- S_00B12C_SO_BASE3_EN(!!info->info.so.strides[3]) |
- S_00B12C_SO_EN(!!info->info.so.num_outputs);
+ S_00B12C_SCRATCH_EN(scratch_enabled);
config_out->rsrc1 = S_00B848_VGPRS((num_vgprs - 1) / 4) |
- S_00B848_SGPRS((num_sgprs - 1) / 8) |
S_00B848_DX10_CLAMP(1) |
S_00B848_FLOAT_MODE(config_out->float_mode);
+ if (pdevice->rad_info.chip_class >= GFX10) {
+ config_out->rsrc2 |= S_00B22C_USER_SGPR_MSB_GFX10(info->num_user_sgprs >> 5);
+ } else {
+ config_out->rsrc1 |= S_00B228_SGPRS((num_sgprs - 1) / 8);
+ config_out->rsrc2 |= S_00B22C_USER_SGPR_MSB_GFX9(info->num_user_sgprs >> 5) |
+ S_00B12C_SO_BASE0_EN(!!info->info.so.strides[0]) |
+ S_00B12C_SO_BASE1_EN(!!info->info.so.strides[1]) |
+ S_00B12C_SO_BASE2_EN(!!info->info.so.strides[2]) |
+ S_00B12C_SO_BASE3_EN(!!info->info.so.strides[3]) |
+ S_00B12C_SO_EN(!!info->info.so.num_outputs);
+ }
+
switch (stage) {
case MESA_SHADER_TESS_EVAL:
- if (info->tes.as_es) {
+ if (info->is_ngg) {
+ config_out->rsrc1 |= S_00B228_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
+ config_out->rsrc2 |= S_00B22C_OC_LDS_EN(1);
+ } else if (info->tes.as_es) {
assert(pdevice->rad_info.chip_class <= GFX8);
vgpr_comp_cnt = info->info.uses_prim_id ? 3 : 2;
+
+ config_out->rsrc2 |= S_00B12C_OC_LDS_EN(1);
} else {
bool enable_prim_id = info->tes.export_prim_id || info->info.uses_prim_id;
vgpr_comp_cnt = enable_prim_id ? 3 : 2;
+
+ config_out->rsrc1 |= S_00B128_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
+ config_out->rsrc2 |= S_00B12C_OC_LDS_EN(1);
}
- config_out->rsrc2 |= S_00B12C_OC_LDS_EN(1);
break;
case MESA_SHADER_TESS_CTRL:
if (pdevice->rad_info.chip_class >= GFX9) {
* 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->info.vs.needs_instance_id ? 2 : 1;
+ if (pdevice->rad_info.chip_class >= GFX10) {
+ vgpr_comp_cnt = info->info.vs.needs_instance_id ? 3 : 1;
+ } else {
+ vgpr_comp_cnt = info->info.vs.needs_instance_id ? 2 : 1;
+ }
} else {
config_out->rsrc2 |= S_00B12C_OC_LDS_EN(1);
}
+ config_out->rsrc1 |= S_00B428_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10) |
+ S_00B848_WGP_MODE(pdevice->rad_info.chip_class >= GFX10);
break;
case MESA_SHADER_VERTEX:
- if (info->vs.as_ls) {
+ if (info->is_ngg) {
+ config_out->rsrc1 |= S_00B228_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
+ } else if (info->vs.as_ls) {
assert(pdevice->rad_info.chip_class <= GFX8);
/* We need at least 2 components for LS.
* VGPR0-3: (VertexID, RelAutoindex, InstanceID / StepRate0, InstanceID).
} else {
vgpr_comp_cnt = 0;
}
+
+ config_out->rsrc1 |= S_00B128_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
}
break;
case MESA_SHADER_FRAGMENT:
+ config_out->rsrc1 |= S_00B028_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10);
+ break;
case MESA_SHADER_GEOMETRY:
+ config_out->rsrc1 |= S_00B228_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10) |
+ S_00B848_WGP_MODE(pdevice->rad_info.chip_class >= GFX10);
break;
case MESA_SHADER_COMPUTE:
+ config_out->rsrc1 |= S_00B848_MEM_ORDERED(pdevice->rad_info.chip_class >= GFX10) |
+ S_00B848_WGP_MODE(pdevice->rad_info.chip_class >= GFX10);
config_out->rsrc2 |=
S_00B84C_TGID_X_EN(info->info.cs.uses_block_id[0]) |
S_00B84C_TGID_Y_EN(info->info.cs.uses_block_id[1]) |
break;
}
- if (pdevice->rad_info.chip_class >= GFX9 &&
- stage == MESA_SHADER_GEOMETRY) {
+ if (pdevice->rad_info.chip_class >= GFX10 && info->is_ngg &&
+ (stage == MESA_SHADER_VERTEX || stage == MESA_SHADER_TESS_EVAL || stage == MESA_SHADER_GEOMETRY)) {
+ unsigned gs_vgpr_comp_cnt, es_vgpr_comp_cnt;
+ gl_shader_stage es_stage = stage;
+ if (stage == MESA_SHADER_GEOMETRY)
+ es_stage = info->gs.es_type;
+
+ /* VGPR5-8: (VertexID, UserVGPR0, UserVGPR1, UserVGPR2 / InstanceID) */
+ if (es_stage == MESA_SHADER_VERTEX) {
+ es_vgpr_comp_cnt = info->info.vs.needs_instance_id ? 3 : 0;
+ } else if (es_stage == MESA_SHADER_TESS_EVAL) {
+ bool enable_prim_id = info->tes.export_prim_id || info->info.uses_prim_id;
+ es_vgpr_comp_cnt = enable_prim_id ? 3 : 2;
+ }
+
+ bool tes_triangles = stage == MESA_SHADER_TESS_EVAL &&
+ info->tes.primitive_mode >= 4; /* GL_TRIANGLES */
+ if (info->info.uses_invocation_id || stage == MESA_SHADER_VERTEX) {
+ gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID. */
+ } else if (info->info.uses_prim_id) {
+ gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
+ } else if (info->gs.vertices_in >= 3 || tes_triangles) {
+ gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
+ } else {
+ gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
+ }
+
+ config_out->rsrc1 |= S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt) |
+ S_00B228_WGP_MODE(1);
+ config_out->rsrc2 |= S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
+ S_00B22C_LDS_SIZE(config_in->lds_size) |
+ S_00B22C_OC_LDS_EN(es_stage == MESA_SHADER_TESS_EVAL);
+ } else if (pdevice->rad_info.chip_class >= GFX9 &&
+ stage == MESA_SHADER_GEOMETRY) {
unsigned es_type = info->gs.es_type;
unsigned gs_vgpr_comp_cnt, es_vgpr_comp_cnt;
unsigned num_lds_symbols = 0;
const char *elf_data = (const char *)((struct radv_shader_binary_rtld *)binary)->data;
size_t elf_size = ((struct radv_shader_binary_rtld *)binary)->elf_size;
+ unsigned esgs_ring_size = 0;
if (device->physical_device->rad_info.chip_class >= GFX9 &&
binary->stage == MESA_SHADER_GEOMETRY && !binary->is_gs_copy_shader) {
+ /* TODO: Do not hardcode this value */
+ esgs_ring_size = 32 * 1024;
+ }
+
+ if (binary->variant_info.is_ngg) {
+ /* GS stores Primitive IDs into LDS at the address
+ * corresponding to the ES thread of the provoking
+ * vertex. All ES threads load and export PrimitiveID
+ * for their thread.
+ */
+ if (binary->stage == MESA_SHADER_VERTEX &&
+ binary->variant_info.vs.export_prim_id) {
+ /* TODO: Do not harcode this value */
+ esgs_ring_size = 256 /* max_out_verts */ * 4;
+ }
+ }
+
+ if (esgs_ring_size) {
/* We add this symbol even on LLVM <= 8 to ensure that
* shader->config.lds_size is set correctly below.
*/
struct ac_rtld_symbol *sym = &lds_symbols[num_lds_symbols++];
sym->name = "esgs_ring";
- sym->size = 32 * 1024;
+ sym->size = esgs_ring_size;
sym->align = 64 * 1024;
+
+ /* Make sure to have LDS space for NGG scratch. */
+ /* TODO: Compute this correctly somehow? */
+ if (binary->variant_info.is_ngg)
+ sym->size -= 32;
}
struct ac_rtld_open_info open_info = {
.info = &device->physical_device->rad_info,
return NULL;
}
- const char *disasm_data;
- size_t disasm_size;
- if (!ac_rtld_get_section_by_name(&rtld_binary, ".AMDGPU.disasm", &disasm_data, &disasm_size)) {
- radv_shader_variant_destroy(device, variant);
- ac_rtld_close(&rtld_binary);
- return NULL;
- }
+ if (device->keep_shader_info) {
+ const char *disasm_data;
+ size_t disasm_size;
+ if (!ac_rtld_get_section_by_name(&rtld_binary, ".AMDGPU.disasm", &disasm_data, &disasm_size)) {
+ radv_shader_variant_destroy(device, variant);
+ ac_rtld_close(&rtld_binary);
+ return NULL;
+ }
- variant->llvm_ir_string = bin->llvm_ir_size ? strdup((const char*)(bin->data + bin->elf_size)) : NULL;
- variant->disasm_string = malloc(disasm_size + 1);
- memcpy(variant->disasm_string, disasm_data, disasm_size);
- variant->disasm_string[disasm_size] = 0;
+ variant->llvm_ir_string = bin->llvm_ir_size ? strdup((const char*)(bin->data + bin->elf_size)) : NULL;
+ variant->disasm_string = malloc(disasm_size + 1);
+ memcpy(variant->disasm_string, disasm_data, disasm_size);
+ variant->disasm_string[disasm_size] = 0;
+ }
ac_rtld_close(&rtld_binary);
} else {
struct radv_shader_variant_info variant_info = {0};
bool thread_compiler;
+ if (shaders[0]->info.stage == MESA_SHADER_FRAGMENT)
+ lower_fs_io(shaders[0], &variant_info);
+
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);
}
const char *
-radv_get_shader_name(struct radv_shader_variant *var, gl_shader_stage stage)
+radv_get_shader_name(struct radv_shader_variant_info *info,
+ gl_shader_stage stage)
{
switch (stage) {
- case MESA_SHADER_VERTEX: return var->info.vs.as_ls ? "Vertex Shader as LS" : var->info.vs.as_es ? "Vertex Shader as ES" : "Vertex Shader as VS";
- case MESA_SHADER_GEOMETRY: return "Geometry Shader";
- case MESA_SHADER_FRAGMENT: return "Pixel Shader";
- case MESA_SHADER_COMPUTE: return "Compute Shader";
- case MESA_SHADER_TESS_CTRL: return "Tessellation Control Shader";
- case MESA_SHADER_TESS_EVAL: return var->info.tes.as_es ? "Tessellation Evaluation Shader as ES" : "Tessellation Evaluation Shader as VS";
+ case MESA_SHADER_VERTEX:
+ if (info->vs.as_ls)
+ return "Vertex Shader as LS";
+ else if (info->vs.as_es)
+ return "Vertex Shader as ES";
+ else if (info->is_ngg)
+ return "Vertex Shader as ESGS";
+ else
+ return "Vertex Shader as VS";
+ case MESA_SHADER_TESS_CTRL:
+ return "Tessellation Control Shader";
+ case MESA_SHADER_TESS_EVAL:
+ if (info->tes.as_es)
+ return "Tessellation Evaluation Shader as ES";
+ else if (info->is_ngg)
+ return "Tessellation Evaluation Shader as ESGS";
+ else
+ return "Tessellation Evaluation Shader as VS";
+ case MESA_SHADER_GEOMETRY:
+ return "Geometry Shader";
+ case MESA_SHADER_FRAGMENT:
+ return "Pixel Shader";
+ case MESA_SHADER_COMPUTE:
+ return "Compute Shader";
default:
return "Unknown shader";
};
generate_shader_stats(device, variant, stage, buf);
- fprintf(file, "\n%s:\n", radv_get_shader_name(variant, stage));
+ fprintf(file, "\n%s:\n", radv_get_shader_name(&variant->info, stage));
fprintf(file, "%s", buf->buf);
_mesa_string_buffer_destroy(buf);
case VK_SHADER_INFO_TYPE_DISASSEMBLY_AMD:
buf = _mesa_string_buffer_create(NULL, 1024);
- _mesa_string_buffer_printf(buf, "%s:\n", radv_get_shader_name(variant, stage));
+ _mesa_string_buffer_printf(buf, "%s:\n", radv_get_shader_name(&variant->info, stage));
_mesa_string_buffer_printf(buf, "%s\n\n", variant->llvm_ir_string);
_mesa_string_buffer_printf(buf, "%s\n\n", variant->disasm_string);
generate_shader_stats(device, variant, stage, buf);