glsl: move lowering linker code out of st
Move all these lowering calls into the linker where they belong. This makes future changes to the linker more flexible and is needed to allow some following patches as we need to call things in a specific order. Reviewed-by: Marek Olšák <marek.olsak@amd.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22846>
This commit is contained in:
Timothy Arceri
committed by
Marge Bot
Marge Bot
parent
a513107424
commit
6786a39940
@@ -786,6 +786,29 @@ nir_build_program_resource_list(const struct gl_constants *consts,
|
||||
_mesa_set_destroy(resource_set, NULL);
|
||||
}
|
||||
|
||||
static void
|
||||
shared_type_info(const struct glsl_type *type, unsigned *size, unsigned *align)
|
||||
{
|
||||
assert(glsl_type_is_vector_or_scalar(type));
|
||||
|
||||
uint32_t comp_size = glsl_type_is_boolean(type)
|
||||
? 4 : glsl_get_bit_size(type) / 8;
|
||||
unsigned length = glsl_get_vector_elements(type);
|
||||
*size = comp_size * length,
|
||||
*align = comp_size * (length == 3 ? 4 : length);
|
||||
}
|
||||
|
||||
static bool
|
||||
can_remove_varying_before_linking(nir_variable *var, void *data)
|
||||
{
|
||||
bool *is_sso = (bool *) data;
|
||||
if (*is_sso) {
|
||||
/* Allow the removal of unused builtins in SSO */
|
||||
return var->data.location > -1 && var->data.location < VARYING_SLOT_VAR0;
|
||||
} else
|
||||
return true;
|
||||
}
|
||||
|
||||
/* - create a gl_PointSize variable
|
||||
* - find every gl_Position write
|
||||
* - store 1.0 to gl_PointSize after every gl_Position write
|
||||
@@ -824,8 +847,220 @@ gl_nir_add_point_size(nir_shader *nir)
|
||||
}
|
||||
}
|
||||
|
||||
static void
|
||||
zero_array_members(nir_builder *b, nir_variable *var)
|
||||
{
|
||||
nir_deref_instr *deref = nir_build_deref_var(b, var);
|
||||
nir_ssa_def *zero = nir_imm_zero(b, 4, 32);
|
||||
for (int i = 0; i < glsl_array_size(var->type); i++) {
|
||||
nir_deref_instr *arr = nir_build_deref_array_imm(b, deref, i);
|
||||
uint32_t mask = BITFIELD_MASK(glsl_get_vector_elements(arr->type));
|
||||
nir_store_deref(b, arr, nir_channels(b, zero, mask), mask);
|
||||
}
|
||||
}
|
||||
|
||||
/* GL has an implicit default of 0 for unwritten gl_ClipDistance members;
|
||||
* to achieve this, write 0 to all members at the start of the shader and
|
||||
* let them be naturally overwritten later
|
||||
*/
|
||||
static bool
|
||||
gl_nir_zero_initialize_clip_distance(nir_shader *nir)
|
||||
{
|
||||
nir_variable *clip_dist0 = nir_find_variable_with_location(nir, nir_var_shader_out, VARYING_SLOT_CLIP_DIST0);
|
||||
nir_variable *clip_dist1 = nir_find_variable_with_location(nir, nir_var_shader_out, VARYING_SLOT_CLIP_DIST1);
|
||||
if (!clip_dist0 && !clip_dist1)
|
||||
return false;
|
||||
|
||||
nir_function_impl *impl = nir_shader_get_entrypoint(nir);
|
||||
nir_builder b = nir_builder_at(nir_before_block(nir_start_block(impl)));
|
||||
if (clip_dist0)
|
||||
zero_array_members(&b, clip_dist0);
|
||||
|
||||
if (clip_dist1)
|
||||
zero_array_members(&b, clip_dist1);
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
static void
|
||||
lower_patch_vertices_in(struct gl_shader_program *shader_prog)
|
||||
{
|
||||
struct gl_linked_shader *linked_tcs =
|
||||
shader_prog->_LinkedShaders[MESA_SHADER_TESS_CTRL];
|
||||
struct gl_linked_shader *linked_tes =
|
||||
shader_prog->_LinkedShaders[MESA_SHADER_TESS_EVAL];
|
||||
|
||||
/* If we have a TCS and TES linked together, lower TES patch vertices. */
|
||||
if (linked_tcs && linked_tes) {
|
||||
nir_shader *tcs_nir = linked_tcs->Program->nir;
|
||||
nir_shader *tes_nir = linked_tes->Program->nir;
|
||||
|
||||
/* The TES input vertex count is the TCS output vertex count,
|
||||
* lower TES gl_PatchVerticesIn to a constant.
|
||||
*/
|
||||
uint32_t tes_patch_verts = tcs_nir->info.tess.tcs_vertices_out;
|
||||
NIR_PASS_V(tes_nir, nir_lower_patch_vertices, tes_patch_verts, NULL);
|
||||
}
|
||||
}
|
||||
|
||||
static void
|
||||
preprocess_shader(const struct gl_constants *consts,
|
||||
const struct gl_extensions *exts,
|
||||
struct gl_program *prog,
|
||||
struct gl_shader_program *shader_program,
|
||||
gl_shader_stage stage)
|
||||
{
|
||||
const struct gl_shader_compiler_options *gl_options =
|
||||
&consts->ShaderCompilerOptions[prog->info.stage];
|
||||
const nir_shader_compiler_options *options = gl_options->NirOptions;
|
||||
assert(options);
|
||||
|
||||
nir_shader *nir = prog->nir;
|
||||
|
||||
if (prog->info.stage == MESA_SHADER_FRAGMENT && consts->HasFBFetch) {
|
||||
nir_shader_gather_info(prog->nir, nir_shader_get_entrypoint(prog->nir));
|
||||
NIR_PASS_V(prog->nir, gl_nir_lower_blend_equation_advanced,
|
||||
exts->KHR_blend_equation_advanced_coherent);
|
||||
nir_lower_global_vars_to_local(prog->nir);
|
||||
NIR_PASS_V(prog->nir, nir_opt_combine_stores, nir_var_shader_out);
|
||||
}
|
||||
|
||||
/* Set the next shader stage hint for VS and TES. */
|
||||
if (!nir->info.separate_shader &&
|
||||
(nir->info.stage == MESA_SHADER_VERTEX ||
|
||||
nir->info.stage == MESA_SHADER_TESS_EVAL)) {
|
||||
|
||||
unsigned prev_stages = (1 << (prog->info.stage + 1)) - 1;
|
||||
unsigned stages_mask =
|
||||
~prev_stages & shader_program->data->linked_stages;
|
||||
|
||||
nir->info.next_stage = stages_mask ?
|
||||
(gl_shader_stage) u_bit_scan(&stages_mask) : MESA_SHADER_FRAGMENT;
|
||||
} else {
|
||||
nir->info.next_stage = MESA_SHADER_FRAGMENT;
|
||||
}
|
||||
|
||||
prog->skip_pointsize_xfb = !(nir->info.outputs_written & VARYING_BIT_PSIZ);
|
||||
if (!consts->PointSizeFixed && prog->skip_pointsize_xfb &&
|
||||
stage < MESA_SHADER_FRAGMENT && stage != MESA_SHADER_TESS_CTRL &&
|
||||
gl_nir_can_add_pointsize_to_program(consts, prog)) {
|
||||
NIR_PASS_V(nir, gl_nir_add_point_size);
|
||||
}
|
||||
|
||||
if (stage < MESA_SHADER_FRAGMENT && stage != MESA_SHADER_TESS_CTRL &&
|
||||
(nir->info.outputs_written & (VARYING_BIT_CLIP_DIST0 | VARYING_BIT_CLIP_DIST1)))
|
||||
NIR_PASS_V(nir, gl_nir_zero_initialize_clip_distance);
|
||||
|
||||
struct nir_remove_dead_variables_options opts;
|
||||
bool is_sso = nir->info.separate_shader;
|
||||
opts.can_remove_var_data = &is_sso;
|
||||
opts.can_remove_var = &can_remove_varying_before_linking;
|
||||
nir_variable_mode mask = nir_var_shader_in | nir_var_shader_out;
|
||||
nir_remove_dead_variables(nir, mask, &opts);
|
||||
|
||||
if (options->lower_all_io_to_temps ||
|
||||
nir->info.stage == MESA_SHADER_VERTEX ||
|
||||
nir->info.stage == MESA_SHADER_GEOMETRY) {
|
||||
NIR_PASS_V(nir, nir_lower_io_to_temporaries,
|
||||
nir_shader_get_entrypoint(nir),
|
||||
true, true);
|
||||
} else if (nir->info.stage == MESA_SHADER_FRAGMENT ||
|
||||
!consts->SupportsReadingOutputs) {
|
||||
NIR_PASS_V(nir, nir_lower_io_to_temporaries,
|
||||
nir_shader_get_entrypoint(nir),
|
||||
true, false);
|
||||
}
|
||||
|
||||
NIR_PASS_V(nir, nir_lower_global_vars_to_local);
|
||||
NIR_PASS_V(nir, nir_split_var_copies);
|
||||
NIR_PASS_V(nir, nir_lower_var_copies);
|
||||
|
||||
if (gl_options->LowerPrecisionFloat16 && gl_options->LowerPrecisionInt16) {
|
||||
NIR_PASS_V(nir, nir_lower_mediump_vars, nir_var_function_temp | nir_var_shader_temp | nir_var_mem_shared);
|
||||
}
|
||||
|
||||
if (options->lower_to_scalar) {
|
||||
NIR_PASS_V(nir, nir_remove_dead_variables,
|
||||
nir_var_function_temp | nir_var_shader_temp |
|
||||
nir_var_mem_shared, NULL);
|
||||
NIR_PASS_V(nir, nir_opt_copy_prop_vars);
|
||||
NIR_PASS_V(nir, nir_lower_alu_to_scalar,
|
||||
options->lower_to_scalar_filter, NULL);
|
||||
}
|
||||
|
||||
/* before buffers and vars_to_ssa */
|
||||
NIR_PASS_V(nir, gl_nir_lower_images, true);
|
||||
|
||||
if (prog->nir->info.stage == MESA_SHADER_COMPUTE) {
|
||||
NIR_PASS_V(prog->nir, nir_lower_vars_to_explicit_types,
|
||||
nir_var_mem_shared, shared_type_info);
|
||||
NIR_PASS_V(prog->nir, nir_lower_explicit_io,
|
||||
nir_var_mem_shared, nir_address_format_32bit_offset);
|
||||
}
|
||||
|
||||
/* Do a round of constant folding to clean up address calculations */
|
||||
NIR_PASS_V(nir, nir_opt_constant_folding);
|
||||
}
|
||||
|
||||
static bool
|
||||
prelink_lowering(const struct gl_constants *consts,
|
||||
const struct gl_extensions *exts,
|
||||
struct gl_shader_program *shader_program,
|
||||
struct gl_linked_shader **linked_shader, unsigned num_shaders)
|
||||
{
|
||||
for (unsigned i = 0; i < num_shaders; i++) {
|
||||
struct gl_linked_shader *shader = linked_shader[i];
|
||||
const nir_shader_compiler_options *options =
|
||||
consts->ShaderCompilerOptions[shader->Stage].NirOptions;
|
||||
struct gl_program *prog = shader->Program;
|
||||
|
||||
preprocess_shader(consts, exts, prog, shader_program, shader->Stage);
|
||||
|
||||
if (prog->nir->info.shared_size > consts->MaxComputeSharedMemorySize) {
|
||||
linker_error(shader_program, "Too much shared memory used (%u/%u)\n",
|
||||
prog->nir->info.shared_size,
|
||||
consts->MaxComputeSharedMemorySize);
|
||||
return false;
|
||||
}
|
||||
|
||||
if (options->lower_to_scalar) {
|
||||
NIR_PASS_V(shader->Program->nir, nir_lower_load_const_to_scalar);
|
||||
}
|
||||
}
|
||||
|
||||
lower_patch_vertices_in(shader_program);
|
||||
|
||||
/* Linking shaders also optimizes them. Separate shaders, compute shaders
|
||||
* and shaders with a fixed-func VS or FS that don't need linking are
|
||||
* optimized here.
|
||||
*/
|
||||
if (num_shaders == 1)
|
||||
gl_nir_opts(linked_shader[0]->Program->nir);
|
||||
|
||||
/* nir_opt_access() needs to run before linking so that ImageAccess[]
|
||||
* and BindlessImage[].access are filled out with the correct modes.
|
||||
*/
|
||||
for (unsigned i = 0; i < num_shaders; i++) {
|
||||
nir_shader *nir = linked_shader[i]->Program->nir;
|
||||
|
||||
nir_opt_access_options opt_access_options;
|
||||
opt_access_options.is_vulkan = false;
|
||||
NIR_PASS_V(nir, nir_opt_access, &opt_access_options);
|
||||
|
||||
/* Combine clip and cull outputs into one array and set:
|
||||
* - shader_info::clip_distance_array_size
|
||||
* - shader_info::cull_distance_array_size
|
||||
*/
|
||||
if (consts->CombinedClipCullDistanceArrays)
|
||||
NIR_PASS_V(nir, nir_lower_clip_cull_distance_arrays);
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
bool
|
||||
gl_nir_link_spirv(const struct gl_constants *consts,
|
||||
const struct gl_extensions *exts,
|
||||
struct gl_shader_program *prog,
|
||||
const struct gl_nir_linker_options *options)
|
||||
{
|
||||
@@ -839,6 +1074,9 @@ gl_nir_link_spirv(const struct gl_constants *consts,
|
||||
linked_shader[num_shaders++] = prog->_LinkedShaders[i];
|
||||
}
|
||||
|
||||
if (!prelink_lowering(consts, exts, prog, linked_shader, num_shaders))
|
||||
return false;
|
||||
|
||||
/* Linking the stages in the opposite order (from fragment to vertex)
|
||||
* ensures that inter-shader outputs written to in an earlier stage
|
||||
* are eliminated if they are (transitively) not used in a later
|
||||
@@ -1000,6 +1238,17 @@ gl_nir_link_glsl(const struct gl_constants *consts,
|
||||
|
||||
MESA_TRACE_FUNC();
|
||||
|
||||
struct gl_linked_shader *linked_shader[MESA_SHADER_STAGES];
|
||||
unsigned num_shaders = 0;
|
||||
|
||||
for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
|
||||
if (prog->_LinkedShaders[i])
|
||||
linked_shader[num_shaders++] = prog->_LinkedShaders[i];
|
||||
}
|
||||
|
||||
if (!prelink_lowering(consts, exts, prog, linked_shader, num_shaders))
|
||||
return false;
|
||||
|
||||
if (!gl_nir_link_varyings(consts, exts, api, prog))
|
||||
return false;
|
||||
|
||||
@@ -1016,14 +1265,6 @@ gl_nir_link_glsl(const struct gl_constants *consts,
|
||||
if (prog->data->LinkStatus == LINKING_FAILURE)
|
||||
return false;
|
||||
|
||||
struct gl_linked_shader *linked_shader[MESA_SHADER_STAGES];
|
||||
unsigned num_shaders = 0;
|
||||
|
||||
for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
|
||||
if (prog->_LinkedShaders[i])
|
||||
linked_shader[num_shaders++] = prog->_LinkedShaders[i];
|
||||
}
|
||||
|
||||
/* Linking the stages in the opposite order (from fragment to vertex)
|
||||
* ensures that inter-shader outputs written to in an earlier stage
|
||||
* are eliminated if they are (transitively) not used in a later
|
||||
|
||||
@@ -56,6 +56,7 @@ struct gl_nir_linker_options {
|
||||
void gl_nir_opts(nir_shader *nir);
|
||||
|
||||
bool gl_nir_link_spirv(const struct gl_constants *consts,
|
||||
const struct gl_extensions *exts,
|
||||
struct gl_shader_program *prog,
|
||||
const struct gl_nir_linker_options *options);
|
||||
|
||||
|
||||
Reference in New Issue
Block a user