[Mesa-dev] [PATCH v4] nir: add varying component packing helpers
Dieter Nützel
Dieter at nuetzel-hh.de
Thu Nov 30 02:54:59 UTC 2017
Tested-by: Dieter Nützel <Dieter at nuetzel-hh.de>
Dieter
Am 30.11.2017 01:20, schrieb Timothy Arceri:
> v2: update shader info input/output masks when pack components
> v3: make sure interpolation loc matches, this is required for the
> radeonsi NIR backend.
> v4: 33dca36f4f28 fixed nir_gather_info to update outputs_read
> correct, make sure we also adjust this correctly when
> packing components.
>
> Reviewed-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl> (v1)
> Reviewed-by: Nicolai Hähnle <nicolai.haehnle at amd.com> (v3)
> ---
> src/compiler/nir/nir.h | 2 +
> src/compiler/nir/nir_linking_helpers.c | 330
> +++++++++++++++++++++++++++++++++
> 2 files changed, 332 insertions(+)
>
> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> index 4c5d976a60d..83858afe148 100644
> --- a/src/compiler/nir/nir.h
> +++ b/src/compiler/nir/nir.h
> @@ -2452,20 +2452,22 @@ void nir_lower_io_to_temporaries(nir_shader
> *shader,
> nir_function_impl *entrypoint,
> bool outputs, bool inputs);
>
> void nir_shader_gather_info(nir_shader *shader, nir_function_impl
> *entrypoint);
>
> void nir_assign_var_locations(struct exec_list *var_list, unsigned
> *size,
> int (*type_size)(const struct glsl_type
> *));
>
> /* Some helpers to do very simple linking */
> bool nir_remove_unused_varyings(nir_shader *producer, nir_shader
> *consumer);
> +void nir_compact_varyings(nir_shader *producer, nir_shader *consumer,
> + bool default_to_smooth_interp);
>
> typedef enum {
> /* If set, this forces all non-flat fragment shader inputs to be
> * interpolated as if with the "sample" qualifier. This requires
> * nir_shader_compiler_options::use_interpolated_input_intrinsics.
> */
> nir_lower_io_force_sample_interpolation = (1 << 1),
> } nir_lower_io_options;
> bool nir_lower_io(nir_shader *shader,
> nir_variable_mode modes,
> diff --git a/src/compiler/nir/nir_linking_helpers.c
> b/src/compiler/nir/nir_linking_helpers.c
> index 4d709c1b3c5..9f0122d4519 100644
> --- a/src/compiler/nir/nir_linking_helpers.c
> +++ b/src/compiler/nir/nir_linking_helpers.c
> @@ -166,10 +166,340 @@ nir_remove_unused_varyings(nir_shader
> *producer, nir_shader *consumer)
>
> bool progress = false;
> progress = remove_unused_io_vars(producer, &producer->outputs,
> read,
> patches_read);
>
> progress = remove_unused_io_vars(consumer, &consumer->inputs,
> written,
> patches_written) || progress;
>
> return progress;
> }
> +
> +static uint8_t
> +get_interp_type(nir_variable *var, bool default_to_smooth_interp)
> +{
> + if (var->data.interpolation != INTERP_MODE_NONE)
> + return var->data.interpolation;
> + else if (default_to_smooth_interp)
> + return INTERP_MODE_SMOOTH;
> + else
> + return INTERP_MODE_NONE;
> +}
> +
> +#define INTERPOLATE_LOC_SAMPLE 0
> +#define INTERPOLATE_LOC_CENTROID 1
> +#define INTERPOLATE_LOC_CENTER 2
> +
> +static uint8_t
> +get_interp_loc(nir_variable *var)
> +{
> + if (var->data.sample)
> + return INTERPOLATE_LOC_SAMPLE;
> + else if (var->data.centroid)
> + return INTERPOLATE_LOC_CENTROID;
> + else
> + return INTERPOLATE_LOC_CENTER;
> +}
> +
> +static void
> +get_slot_component_masks_and_interp_types(struct exec_list *var_list,
> + uint8_t *comps,
> + uint8_t *interp_type,
> + uint8_t *interp_loc,
> + gl_shader_stage stage,
> + bool
> default_to_smooth_interp)
> +{
> + nir_foreach_variable_safe(var, var_list) {
> + assert(var->data.location >= 0);
> +
> + /* Only remap things that aren't built-ins.
> + * TODO: add TES patch support.
> + */
> + if (var->data.location >= VARYING_SLOT_VAR0 &&
> + var->data.location - VARYING_SLOT_VAR0 < 32) {
> +
> + const struct glsl_type *type = var->type;
> + if (nir_is_per_vertex_io(var, stage)) {
> + assert(glsl_type_is_array(type));
> + type = glsl_get_array_element(type);
> + }
> +
> + unsigned location = var->data.location - VARYING_SLOT_VAR0;
> + unsigned elements =
> + glsl_get_vector_elements(glsl_without_array(type));
> +
> + bool dual_slot =
> glsl_type_is_dual_slot(glsl_without_array(type));
> + unsigned slots = glsl_count_attribute_slots(type, false);
> + unsigned comps_slot2 = 0;
> + for (unsigned i = 0; i < slots; i++) {
> + interp_type[location + i] =
> + get_interp_type(var, default_to_smooth_interp);
> + interp_loc[location + i] = get_interp_loc(var);
> +
> + if (dual_slot) {
> + if (i & 1) {
> + comps[location + i] |= ((1 << comps_slot2) - 1);
> + } else {
> + unsigned num_comps = 4 - var->data.location_frac;
> + comps_slot2 = (elements * 2) - num_comps;
> +
> + /* Assume ARB_enhanced_layouts packing rules for
> doubles */
> + assert(var->data.location_frac == 0 ||
> + var->data.location_frac == 2);
> + assert(comps_slot2 <= 4);
> +
> + comps[location + i] |=
> + ((1 << num_comps) - 1) <<
> var->data.location_frac;
> + }
> + } else {
> + comps[location + i] |=
> + ((1 << elements) - 1) << var->data.location_frac;
> + }
> + }
> + }
> + }
> +}
> +
> +struct varying_loc
> +{
> + uint8_t component;
> + uint32_t location;
> +};
> +
> +static void
> +remap_slots_and_components(struct exec_list *var_list, gl_shader_stage
> stage,
> + struct varying_loc (*remap)[4],
> + uint64_t *slots_used, uint64_t
> *out_slots_read)
> + {
> + uint64_t out_slots_read_tmp = 0;
> +
> + /* We don't touch builtins so just copy the bitmask */
> + uint64_t slots_used_tmp =
> + *slots_used & (((uint64_t)1 << (VARYING_SLOT_VAR0 - 1)) - 1);
> +
> + nir_foreach_variable(var, var_list) {
> + assert(var->data.location >= 0);
> +
> + /* Only remap things that aren't built-ins */
> + if (var->data.location >= VARYING_SLOT_VAR0 &&
> + var->data.location - VARYING_SLOT_VAR0 < 32) {
> + assert(var->data.location - VARYING_SLOT_VAR0 < 32);
> + assert(remap[var->data.location - VARYING_SLOT_VAR0] >= 0);
> +
> + const struct glsl_type *type = var->type;
> + if (nir_is_per_vertex_io(var, stage)) {
> + assert(glsl_type_is_array(type));
> + type = glsl_get_array_element(type);
> + }
> +
> + unsigned num_slots = glsl_count_attribute_slots(type, false);
> + bool used_across_stages = false;
> + bool outputs_read = false;
> +
> + unsigned location = var->data.location - VARYING_SLOT_VAR0;
> + struct varying_loc *new_loc =
> &remap[location][var->data.location_frac];
> + if (new_loc->location) {
> + uint64_t slots = (((uint64_t)1 << num_slots) - 1) <<
> var->data.location;
> + if (slots & *slots_used)
> + used_across_stages = true;
> +
> + if (slots & *out_slots_read)
> + outputs_read = true;
> +
> + var->data.location = new_loc->location;
> + var->data.location_frac = new_loc->component;
> + }
> +
> + if (var->data.always_active_io) {
> + /* We can't apply link time optimisations (specifically
> array
> + * splitting) to these so we need to copy the existing
> mask
> + * otherwise we will mess up the mask for things like
> partially
> + * marked arrays.
> + */
> + if (used_across_stages) {
> + slots_used_tmp |=
> + *slots_used & (((uint64_t)1 << num_slots) - 1) <<
> var->data.location;
> + }
> +
> + if (outputs_read) {
> + out_slots_read_tmp |=
> + *out_slots_read & (((uint64_t)1 << num_slots) - 1)
> << var->data.location;
> + }
> +
> + } else {
> + for (unsigned i = 0; i < num_slots; i++) {
> + if (used_across_stages)
> + slots_used_tmp |= (uint64_t)1 << (var->data.location
> + i);
> +
> + if (outputs_read)
> + out_slots_read_tmp |= (uint64_t)1 <<
> (var->data.location + i);
> + }
> + }
> + }
> + }
> +
> + *slots_used = slots_used_tmp;
> + *out_slots_read = out_slots_read_tmp;
> +}
> +
> +/* If there are empty components in the slot compact the remaining
> components
> + * as close to component 0 as possible. This will make it easier to
> fill the
> + * empty components with components from a different slot in a
> following pass.
> + */
> +static void
> +compact_components(nir_shader *producer, nir_shader *consumer, uint8_t
> *comps,
> + uint8_t *interp_type, uint8_t *interp_loc,
> + bool default_to_smooth_interp)
> +{
> + struct exec_list *input_list = &consumer->inputs;
> + struct exec_list *output_list = &producer->outputs;
> + struct varying_loc remap[32][4] = {{{0}, {0}}};
> +
> + /* Create a cursor for each interpolation type */
> + unsigned cursor[4] = {0};
> +
> + /* We only need to pass over one stage and we choose the consumer
> as it seems
> + * to cause a larger reduction in instruction counts (tested on
> i965).
> + */
> + nir_foreach_variable(var, input_list) {
> +
> + /* Only remap things that aren't builtins.
> + * TODO: add TES patch support.
> + */
> + if (var->data.location >= VARYING_SLOT_VAR0 &&
> + var->data.location - VARYING_SLOT_VAR0 < 32) {
> +
> + /* We can't repack xfb varyings. */
> + if (var->data.always_active_io)
> + continue;
> +
> + const struct glsl_type *type = var->type;
> + if (nir_is_per_vertex_io(var, consumer->info.stage)) {
> + assert(glsl_type_is_array(type));
> + type = glsl_get_array_element(type);
> + }
> +
> + /* Skip types that require more complex packing handling.
> + * TODO: add support for these types.
> + */
> + if (glsl_type_is_array(type) ||
> + glsl_type_is_dual_slot(type) ||
> + glsl_type_is_matrix(type) ||
> + glsl_type_is_struct(type) ||
> + glsl_type_is_64bit(type))
> + continue;
> +
> + /* We ignore complex types above and all other vector types
> should
> + * have been split into scalar variables by the
> lower_io_to_scalar
> + * pass. The only exeption should by OpenGL xfb varyings.
> + */
> + if (glsl_get_vector_elements(type) != 1)
> + continue;
> +
> + unsigned location = var->data.location - VARYING_SLOT_VAR0;
> + uint8_t used_comps = comps[location];
> +
> + /* If there are no empty components there is nothing more
> for us to do.
> + */
> + if (used_comps == 0xf)
> + continue;
> +
> + bool found_new_offset = false;
> + uint8_t interp = get_interp_type(var,
> default_to_smooth_interp);
> + for (; cursor[interp] < 32; cursor[interp]++) {
> + uint8_t cursor_used_comps = comps[cursor[interp]];
> +
> + /* We couldn't find anywhere to pack the varying continue
> on. */
> + if (cursor[interp] == location &&
> + (var->data.location_frac == 0 ||
> + cursor_used_comps & ((1 << (var->data.location_frac))
> - 1)))
> + break;
> +
> + /* We can only pack varyings with matching interpolation
> types */
> + if (interp_type[cursor[interp]] != interp)
> + continue;
> +
> + /* Interpolation loc must match also.
> + * TODO: i965 can handle these if they don't match, but
> the
> + * radeonsi nir backend handles everything as vec4s and so
> expects
> + * this to be the same for all components. We could make
> this
> + * check driver specfific or drop it if NIR ever become
> the only
> + * radeonsi backend.
> + */
> + if (interp_loc[cursor[interp]] != get_interp_loc(var))
> + continue;
> +
> + /* If the slot is empty just skip it for now,
> compact_var_list()
> + * can be called after this function to remove empty slots
> for us.
> + * TODO: finish implementing compact_var_list() requires
> array and
> + * matrix splitting.
> + */
> + if (!cursor_used_comps)
> + continue;
> +
> + uint8_t unused_comps = ~cursor_used_comps;
> +
> + for (unsigned i = 0; i < 4; i++) {
> + uint8_t new_var_comps = 1 << i;
> + if (unused_comps & new_var_comps) {
> + remap[location][var->data.location_frac].component =
> i;
> + remap[location][var->data.location_frac].location =
> + cursor[interp] + VARYING_SLOT_VAR0;
> +
> + found_new_offset = true;
> +
> + /* Turn off the mask for the component we are
> remapping */
> + if (comps[location] & 1 << var->data.location_frac)
> {
> + comps[location] ^= 1 << var->data.location_frac;
> + comps[cursor[interp]] |= new_var_comps;
> + }
> + break;
> + }
> + }
> +
> + if (found_new_offset)
> + break;
> + }
> + }
> + }
> +
> + uint64_t zero = 0;
> + remap_slots_and_components(input_list, consumer->info.stage, remap,
> + &consumer->info.inputs_read, &zero);
> + remap_slots_and_components(output_list, producer->info.stage,
> remap,
> + &producer->info.outputs_written,
> + &producer->info.outputs_read);
> +}
> +
> +/* We assume that this has been called more-or-less directly after
> + * remove_unused_varyings. At this point, all of the varyings that we
> + * aren't going to be using have been completely removed and the
> + * inputs_read and outputs_written fields in nir_shader_info reflect
> + * this. Therefore, the total set of valid slots is the OR of the two
> + * sets of varyings; this accounts for varyings which one side may
> need
> + * to read/write even if the other doesn't. This can happen if, for
> + * instance, an array is used indirectly from one side causing it to
> be
> + * unsplittable but directly from the other.
> + */
> +void
> +nir_compact_varyings(nir_shader *producer, nir_shader *consumer,
> + bool default_to_smooth_interp)
> +{
> + assert(producer->info.stage != MESA_SHADER_FRAGMENT);
> + assert(consumer->info.stage != MESA_SHADER_VERTEX);
> +
> + uint8_t comps[32] = {0};
> + uint8_t interp_type[32] = {0};
> + uint8_t interp_loc[32] = {0};
> +
> + get_slot_component_masks_and_interp_types(&producer->outputs,
> comps,
> + interp_type, interp_loc,
> + producer->info.stage,
> +
> default_to_smooth_interp);
> + get_slot_component_masks_and_interp_types(&consumer->inputs, comps,
> + interp_type, interp_loc,
> + consumer->info.stage,
> +
> default_to_smooth_interp);
> +
> + compact_components(producer, consumer, comps, interp_type,
> interp_loc,
> + default_to_smooth_interp);
> +}
More information about the mesa-dev
mailing list