#include "broadcom/common/v3d_device_info.h"
#include "v3d_compiler.h"
+#include "util/u_prim.h"
+#include "compiler/nir/nir_schedule.h"
int
vir_get_nsrc(struct qinst *inst)
if (inst->qpu.sig.ldtmu ||
inst->qpu.sig.ldvary ||
+ inst->qpu.sig.ldtlbu ||
+ inst->qpu.sig.ldtlb ||
inst->qpu.sig.wrtmuc ||
inst->qpu.sig.thrsw) {
return true;
}
}
+ /* CS textures may not have return_size reflecting the shadow state. */
+ nir_foreach_variable(var, &c->s->uniforms) {
+ const struct glsl_type *type = glsl_without_array(var->type);
+ unsigned array_len = MAX2(glsl_get_length(var->type), 1);
+
+ if (!glsl_type_is_sampler(type) ||
+ !glsl_sampler_type_is_shadow(type))
+ continue;
+
+ for (int i = 0; i < array_len; i++) {
+ tex_options.lower_tex_packing[var->data.binding + i] =
+ nir_lower_tex_packing_16;
+ }
+ }
+
NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
NIR_PASS_V(c->s, nir_lower_system_values);
+
+ NIR_PASS_V(c->s, nir_lower_vars_to_scratch,
+ nir_var_function_temp,
+ 0,
+ glsl_get_natural_size_align_bytes);
+ NIR_PASS_V(c->s, v3d_nir_lower_scratch);
}
static void
prog_data->vcm_cache_size = CLAMP(vpm_output_batches - 1, 2, 4);
}
+static void
+v3d_gs_set_prog_data(struct v3d_compile *c,
+ struct v3d_gs_prog_data *prog_data)
+{
+ prog_data->num_inputs = c->num_inputs;
+ memcpy(prog_data->input_slots, c->input_slots,
+ c->num_inputs * sizeof(*c->input_slots));
+
+ /* gl_PrimitiveIdIn is written by the GBG into the first word of the
+ * VPM output header automatically and the shader will overwrite
+ * it after reading it if necessary, so it doesn't add to the VPM
+ * size requirements.
+ */
+ prog_data->uses_pid = (c->s->info.system_values_read &
+ (1ull << SYSTEM_VALUE_PRIMITIVE_ID));
+
+ /* Output segment size is in sectors (8 rows of 32 bits per channel) */
+ prog_data->vpm_output_size = align(c->vpm_output_size, 8) / 8;
+
+ /* Compute SIMD dispatch width and update VPM output size accordingly
+ * to ensure we can fit our program in memory. Available widths are
+ * 16, 8, 4, 1.
+ *
+ * Notice that at draw time we will have to consider VPM memory
+ * requirements from other stages and choose a smaller dispatch
+ * width if needed to fit the program in VPM memory.
+ */
+ prog_data->simd_width = 16;
+ while ((prog_data->simd_width > 1 && prog_data->vpm_output_size > 16) ||
+ prog_data->simd_width == 2) {
+ prog_data->simd_width >>= 1;
+ prog_data->vpm_output_size =
+ align(prog_data->vpm_output_size, 2) / 2;
+ }
+ assert(prog_data->vpm_output_size <= 16);
+ assert(prog_data->simd_width != 2);
+
+ prog_data->out_prim_type = c->s->info.gs.output_primitive;
+ prog_data->num_invocations = c->s->info.gs.invocations;
+}
+
static void
v3d_set_fs_prog_data_inputs(struct v3d_compile *c,
struct v3d_fs_prog_data *prog_data)
prog_data->writes_z = c->writes_z;
prog_data->disable_ez = !c->s->info.fs.early_fragment_tests;
prog_data->uses_center_w = c->uses_center_w;
+ prog_data->uses_implicit_point_line_varyings =
+ c->uses_implicit_point_line_varyings;
+ prog_data->lock_scoreboard_on_first_thrsw =
+ c->lock_scoreboard_on_first_thrsw;
+}
+
+static void
+v3d_cs_set_prog_data(struct v3d_compile *c,
+ struct v3d_compute_prog_data *prog_data)
+{
+ prog_data->shared_size = c->s->info.cs.shared_size;
}
static void
prog_data->threads = c->threads;
prog_data->single_seg = !c->last_thrsw;
prog_data->spill_size = c->spill_size;
+ prog_data->tmu_dirty_rcl = c->tmu_dirty_rcl;
v3d_set_prog_data_uniforms(c, prog_data);
- if (c->s->info.stage == MESA_SHADER_VERTEX) {
+ switch (c->s->info.stage) {
+ case MESA_SHADER_VERTEX:
v3d_vs_set_prog_data(c, (struct v3d_vs_prog_data *)prog_data);
- } else {
- assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
+ break;
+ case MESA_SHADER_GEOMETRY:
+ v3d_gs_set_prog_data(c, (struct v3d_gs_prog_data *)prog_data);
+ break;
+ case MESA_SHADER_FRAGMENT:
v3d_fs_set_prog_data(c, (struct v3d_fs_prog_data *)prog_data);
+ break;
+ case MESA_SHADER_COMPUTE:
+ v3d_cs_set_prog_data(c, (struct v3d_compute_prog_data *)prog_data);
+ break;
+ default:
+ unreachable("unsupported shader stage");
}
}
NIR_PASS_V(c->s, nir_lower_io_to_scalar_early,
nir_var_shader_in | nir_var_shader_out);
uint64_t used_outputs[4] = {0};
- for (int i = 0; i < c->vs_key->num_fs_inputs; i++) {
- int slot = v3d_slot_get_slot(c->vs_key->fs_inputs[i]);
- int comp = v3d_slot_get_component(c->vs_key->fs_inputs[i]);
+ for (int i = 0; i < c->vs_key->num_used_outputs; i++) {
+ int slot = v3d_slot_get_slot(c->vs_key->used_outputs[i]);
+ int comp = v3d_slot_get_component(c->vs_key->used_outputs[i]);
used_outputs[comp] |= 1ull << slot;
}
NIR_PASS_V(c->s, nir_remove_unused_io_vars,
&c->s->outputs, used_outputs, NULL); /* demotes to globals */
NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
v3d_optimize_nir(c->s);
- NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in);
+ NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in, NULL);
+
+ /* This must go before nir_lower_io */
+ if (c->vs_key->per_vertex_point_size)
+ NIR_PASS_V(c->s, nir_lower_point_size, 1.0f, 0.0f);
+
NIR_PASS_V(c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
type_size_vec4,
(nir_lower_io_options)0);
+ /* clean up nir_lower_io's deref_var remains */
+ NIR_PASS_V(c->s, nir_opt_dce);
+}
+
+static void
+v3d_nir_lower_gs_early(struct v3d_compile *c)
+{
+ /* Split our I/O vars and dead code eliminate the unused
+ * components.
+ */
+ NIR_PASS_V(c->s, nir_lower_io_to_scalar_early,
+ nir_var_shader_in | nir_var_shader_out);
+ uint64_t used_outputs[4] = {0};
+ for (int i = 0; i < c->gs_key->num_used_outputs; i++) {
+ int slot = v3d_slot_get_slot(c->gs_key->used_outputs[i]);
+ int comp = v3d_slot_get_component(c->gs_key->used_outputs[i]);
+ used_outputs[comp] |= 1ull << slot;
+ }
+ NIR_PASS_V(c->s, nir_remove_unused_io_vars,
+ &c->s->outputs, used_outputs, NULL); /* demotes to globals */
+ NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
+ v3d_optimize_nir(c->s);
+ NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in, NULL);
+
+ /* This must go before nir_lower_io */
+ if (c->gs_key->per_vertex_point_size)
+ NIR_PASS_V(c->s, nir_lower_point_size, 1.0f, 0.0f);
+
+ NIR_PASS_V(c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
+ type_size_vec4,
+ (nir_lower_io_options)0);
+ /* clean up nir_lower_io's deref_var remains */
+ NIR_PASS_V(c->s, nir_opt_dce);
}
static void
if (c->fs_key->int_color_rb || c->fs_key->uint_color_rb)
v3d_fixup_fs_output_types(c);
+ NIR_PASS_V(c->s, v3d_nir_lower_logic_ops, c);
+
+ if (c->fs_key->line_smoothing) {
+ v3d_nir_lower_line_smooth(c->s);
+ NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
+ /* The lowering pass can introduce new sysval reads */
+ nir_shader_gather_info(c->s, nir_shader_get_entrypoint(c->s));
+ }
+
/* If the shader has no non-TLB side effects, we can promote it to
* enabling early_fragment_tests even if the user didn't.
*/
if (!(c->s->info.num_images ||
- c->s->info.num_ssbos ||
- c->s->info.num_abos)) {
+ c->s->info.num_ssbos)) {
c->s->info.fs.early_fragment_tests = true;
}
}
+static void
+v3d_nir_lower_gs_late(struct v3d_compile *c)
+{
+ if (c->key->ucp_enables) {
+ NIR_PASS_V(c->s, nir_lower_clip_gs, c->key->ucp_enables,
+ false, NULL);
+ }
+
+ /* Note: GS output scalarizing must happen after nir_lower_clip_gs. */
+ NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
+}
+
static void
v3d_nir_lower_vs_late(struct v3d_compile *c)
{
if (c->key->ucp_enables) {
NIR_PASS_V(c->s, nir_lower_clip_vs, c->key->ucp_enables,
- false);
+ false, false, NULL);
NIR_PASS_V(c->s, nir_lower_io_to_scalar,
nir_var_shader_out);
}
v3d_nir_lower_fs_late(struct v3d_compile *c)
{
if (c->fs_key->light_twoside)
- NIR_PASS_V(c->s, nir_lower_two_sided_color);
+ NIR_PASS_V(c->s, nir_lower_two_sided_color, true);
if (c->fs_key->clamp_color)
NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
if (c->fs_key->alpha_test) {
NIR_PASS_V(c->s, nir_lower_alpha_test,
c->fs_key->alpha_test_func,
- false);
+ false, NULL);
}
if (c->key->ucp_enables)
- NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables);
+ NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables,
+ false);
/* Note: FS input scalarizing must happen after
* nir_lower_two_sided_color, which only handles a vec4 at a time.
return max_temps;
}
+enum v3d_dependency_class {
+ V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0
+};
+
+static bool
+v3d_intrinsic_dependency_cb(nir_intrinsic_instr *intr,
+ nir_schedule_dependency *dep,
+ void *user_data)
+{
+ struct v3d_compile *c = user_data;
+
+ switch (intr->intrinsic) {
+ case nir_intrinsic_store_output:
+ /* Writing to location 0 overwrites the value passed in for
+ * gl_PrimitiveID on geometry shaders
+ */
+ if (c->s->info.stage != MESA_SHADER_GEOMETRY ||
+ nir_intrinsic_base(intr) != 0)
+ break;
+
+ nir_const_value *const_value =
+ nir_src_as_const_value(intr->src[1]);
+
+ if (const_value == NULL)
+ break;
+
+ uint64_t offset =
+ nir_const_value_as_uint(*const_value,
+ nir_src_bit_size(intr->src[1]));
+ if (offset != 0)
+ break;
+
+ dep->klass = V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0;
+ dep->type = NIR_SCHEDULE_WRITE_DEPENDENCY;
+ return true;
+
+ case nir_intrinsic_load_primitive_id:
+ if (c->s->info.stage != MESA_SHADER_GEOMETRY)
+ break;
+
+ dep->klass = V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0;
+ dep->type = NIR_SCHEDULE_READ_DEPENDENCY;
+ return true;
+
+ default:
+ break;
+ }
+
+ return false;
+}
+
uint64_t *v3d_compile(const struct v3d_compiler *compiler,
struct v3d_key *key,
struct v3d_prog_data **out_prog_data,
c->vs_key = (struct v3d_vs_key *)key;
prog_data = rzalloc_size(NULL, sizeof(struct v3d_vs_prog_data));
break;
+ case MESA_SHADER_GEOMETRY:
+ c->gs_key = (struct v3d_gs_key *)key;
+ prog_data = rzalloc_size(NULL, sizeof(struct v3d_gs_prog_data));
+ break;
case MESA_SHADER_FRAGMENT:
c->fs_key = (struct v3d_fs_key *)key;
prog_data = rzalloc_size(NULL, sizeof(struct v3d_fs_prog_data));
break;
+ case MESA_SHADER_COMPUTE:
+ prog_data = rzalloc_size(NULL,
+ sizeof(struct v3d_compute_prog_data));
+ break;
default:
unreachable("unsupported shader stage");
}
- if (c->s->info.stage == MESA_SHADER_VERTEX) {
+
+ switch (c->s->info.stage) {
+ case MESA_SHADER_VERTEX:
v3d_nir_lower_vs_early(c);
- } else {
- assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
+ break;
+ case MESA_SHADER_GEOMETRY:
+ v3d_nir_lower_gs_early(c);
+ break;
+ case MESA_SHADER_FRAGMENT:
v3d_nir_lower_fs_early(c);
+ break;
+ default:
+ break;
}
v3d_lower_nir(c);
- if (c->s->info.stage == MESA_SHADER_VERTEX) {
+ switch (c->s->info.stage) {
+ case MESA_SHADER_VERTEX:
v3d_nir_lower_vs_late(c);
- } else {
- assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
+ break;
+ case MESA_SHADER_GEOMETRY:
+ v3d_nir_lower_gs_late(c);
+ break;
+ case MESA_SHADER_FRAGMENT:
v3d_nir_lower_fs_late(c);
+ break;
+ default:
+ break;
}
NIR_PASS_V(c->s, v3d_nir_lower_io, c);
NIR_PASS_V(c->s, v3d_nir_lower_txf_ms, c);
NIR_PASS_V(c->s, v3d_nir_lower_image_load_store);
- NIR_PASS_V(c->s, nir_lower_idiv);
+ NIR_PASS_V(c->s, nir_lower_idiv, nir_lower_idiv_fast);
v3d_optimize_nir(c->s);
+
+ /* Do late algebraic optimization to turn add(a, neg(b)) back into
+ * subs, then the mandatory cleanup after algebraic. Note that it may
+ * produce fnegs, and if so then we need to keep running to squash
+ * fneg(fneg(a)).
+ */
+ bool more_late_algebraic = true;
+ while (more_late_algebraic) {
+ more_late_algebraic = false;
+ NIR_PASS(more_late_algebraic, c->s, nir_opt_algebraic_late);
+ NIR_PASS_V(c->s, nir_opt_constant_folding);
+ NIR_PASS_V(c->s, nir_copy_prop);
+ NIR_PASS_V(c->s, nir_opt_dce);
+ NIR_PASS_V(c->s, nir_opt_cse);
+ }
+
NIR_PASS_V(c->s, nir_lower_bool_to_int32);
NIR_PASS_V(c->s, nir_convert_from_ssa, true);
+ struct nir_schedule_options schedule_options = {
+ /* Schedule for about half our register space, to enable more
+ * shaders to hit 4 threads.
+ */
+ .threshold = 24,
+
+ /* Vertex shaders share the same memory for inputs and outputs,
+ * fragement and geometry shaders do not.
+ */
+ .stages_with_shared_io_memory =
+ (((1 << MESA_ALL_SHADER_STAGES) - 1) &
+ ~((1 << MESA_SHADER_FRAGMENT) |
+ (1 << MESA_SHADER_GEOMETRY))),
+
+ .intrinsic_cb = v3d_intrinsic_dependency_cb,
+ .intrinsic_cb_data = c,
+ };
+ NIR_PASS_V(c->s, nir_schedule, &schedule_options);
+
v3d_nir_to_vir(c);
v3d_set_prog_data(c, prog_data);
char *shaderdb;
int ret = asprintf(&shaderdb,
"%s shader: %d inst, %d threads, %d loops, "
- "%d uniforms, %d max-temps, %d:%d spills:fills",
+ "%d uniforms, %d max-temps, %d:%d spills:fills, "
+ "%d sfu-stalls, %d inst-and-stalls",
vir_get_stage_name(c),
c->qpu_inst_count,
c->threads,
c->num_uniforms,
vir_get_max_temps(c),
c->spills,
- c->fills);
+ c->fills,
+ c->qpu_inst_stalled_count,
+ c->qpu_inst_count + c->qpu_inst_stalled_count);
if (ret >= 0) {
+ if (V3D_DEBUG & V3D_DEBUG_SHADERDB)
+ fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
+
c->debug_output(shaderdb, c->debug_output_data);
free(shaderdb);
}
c->cursor.link = NULL;
vir_for_each_block(block, c) {
- while (!list_empty(&block->instructions)) {
+ while (!list_is_empty(&block->instructions)) {
struct qinst *qinst =
list_first_entry(&block->instructions,
struct qinst, link);
vir_get_stage_name(struct v3d_compile *c)
{
if (c->vs_key && c->vs_key->is_coord)
- return "MESA_SHADER_COORD";
+ return "MESA_SHADER_VERTEX_BIN";
+ else if (c->gs_key && c->gs_key->is_coord)
+ return "MESA_SHADER_GEOMETRY_BIN";
else
return gl_shader_stage_name(c->s->info.stage);
}
projects / mesa.git / blobdiff