int
vir_get_implicit_uniform_src(struct qinst *inst)
{
+ if (!vir_has_implicit_uniform(inst))
+ return -1;
return vir_get_nsrc(inst) - 1;
}
case V3D_QPU_A_STVPMD:
case V3D_QPU_A_STVPMP:
case V3D_QPU_A_VPMWT:
+ case V3D_QPU_A_TMUWT:
return true;
default:
break;
if (inst->dst.file == QFILE_MAGIC)
return v3d_qpu_magic_waddr_is_tmu(inst->dst.index);
- return false;
-}
-
-bool
-vir_depends_on_flags(struct qinst *inst)
-{
- if (inst->qpu.type == V3D_QPU_INSTR_TYPE_BRANCH) {
- return (inst->qpu.branch.cond != V3D_QPU_BRANCH_COND_ALWAYS);
- } else {
- return (inst->qpu.flags.ac != V3D_QPU_COND_NONE &&
- inst->qpu.flags.mc != V3D_QPU_COND_NONE);
+ if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
+ inst->qpu.alu.add.op == V3D_QPU_A_TMUWT) {
+ return true;
}
+
+ return false;
}
bool
}
}
+void
+vir_set_uf(struct qinst *inst, enum v3d_qpu_uf uf)
+{
+ if (vir_is_add(inst)) {
+ inst->qpu.flags.auf = uf;
+ } else {
+ assert(vir_is_mul(inst));
+ inst->qpu.flags.muf = uf;
+ }
+}
+
#if 0
uint8_t
vir_channels_written(struct qinst *inst)
{
assert(inst->dst.file == QFILE_NULL);
+ /* If we're emitting an instruction that's a def, it had better be
+ * writing a register.
+ */
+ if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU) {
+ assert(inst->qpu.alu.add.op == V3D_QPU_A_NOP ||
+ v3d_qpu_add_op_has_dst(inst->qpu.alu.add.op));
+ assert(inst->qpu.alu.mul.op == V3D_QPU_M_NOP ||
+ v3d_qpu_mul_op_has_dst(inst->qpu.alu.mul.op));
+ }
+
inst->dst = vir_get_temp(c);
if (inst->dst.file == QFILE_TEMP)
vir_compile_init(const struct v3d_compiler *compiler,
struct v3d_key *key,
nir_shader *s,
+ void (*debug_output)(const char *msg,
+ void *debug_output_data),
+ void *debug_output_data,
int program_id, int variant_id)
{
struct v3d_compile *c = rzalloc(NULL, struct v3d_compile);
c->program_id = program_id;
c->variant_id = variant_id;
c->threads = 4;
+ c->debug_output = debug_output;
+ c->debug_output_data = debug_output_data;
s = nir_shader_clone(c, s);
c->s = s;
return c;
}
+static int
+type_size_vec4(const struct glsl_type *type)
+{
+ return glsl_count_attribute_slots(type, false);
+}
+
static void
v3d_lower_nir(struct v3d_compile *c)
{
struct nir_lower_tex_options tex_options = {
.lower_txd = true,
- .lower_rect = false, /* XXX */
+ .lower_rect = false, /* XXX: Use this on V3D 3.x */
.lower_txp = ~0,
/* Apply swizzles to all samplers. */
.swizzle_result = ~0,
memcpy(qpu_insts, c->qpu_insts, *final_assembly_size);
+ char *shaderdb;
+ int ret = asprintf(&shaderdb,
+ "%s shader: %d inst, %d threads, %d loops, "
+ "%d uniforms, %d:%d spills:fills",
+ vir_get_stage_name(c),
+ c->qpu_inst_count,
+ c->threads,
+ c->loops,
+ c->num_uniforms,
+ c->spills,
+ c->fills);
+ if (ret >= 0) {
+ c->debug_output(shaderdb, c->debug_output_data);
+ free(shaderdb);
+ }
+
vir_compile_destroy(c);
return qpu_insts;
struct v3d_vs_key *key,
struct v3d_vs_prog_data *prog_data,
nir_shader *s,
+ void (*debug_output)(const char *msg,
+ void *debug_output_data),
+ void *debug_output_data,
int program_id, int variant_id,
uint32_t *final_assembly_size)
{
struct v3d_compile *c = vir_compile_init(compiler, &key->base, s,
+ debug_output, debug_output_data,
program_id, variant_id);
c->vs_key = key;
+ /* 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->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]);
+ 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_lower_io, nir_var_shader_in | nir_var_shader_out,
+ type_size_vec4,
+ (nir_lower_io_options)0);
+
v3d_lower_nir(c);
if (key->clamp_color)
NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
if (key->base.ucp_enables) {
- NIR_PASS_V(c->s, nir_lower_clip_vs, key->base.ucp_enables);
+ NIR_PASS_V(c->s, nir_lower_clip_vs, key->base.ucp_enables,
+ false);
NIR_PASS_V(c->s, nir_lower_io_to_scalar,
nir_var_shader_out);
}
v3d_lower_nir_late(c);
v3d_optimize_nir(c->s);
+ NIR_PASS_V(c->s, nir_lower_bool_to_int32);
NIR_PASS_V(c->s, nir_convert_from_ssa, true);
v3d_nir_to_vir(c);
prog_data->vpm_input_size += c->vattr_sizes[i];
}
- /* Input/output segment size are in 8x32-bit multiples. */
- prog_data->vpm_input_size = align(prog_data->vpm_input_size, 8) / 8;
- prog_data->vpm_output_size = align(c->num_vpm_writes, 8) / 8;
-
prog_data->uses_vid = (s->info.system_values_read &
(1ull << SYSTEM_VALUE_VERTEX_ID));
prog_data->uses_iid = (s->info.system_values_read &
(1ull << SYSTEM_VALUE_INSTANCE_ID));
+ if (prog_data->uses_vid)
+ prog_data->vpm_input_size++;
+ if (prog_data->uses_iid)
+ prog_data->vpm_input_size++;
+
+ /* Input/output segment size are in sectors (8 rows of 32 bits per
+ * channel).
+ */
+ prog_data->vpm_input_size = align(prog_data->vpm_input_size, 8) / 8;
+ prog_data->vpm_output_size = align(c->num_vpm_writes, 8) / 8;
+
+ /* Set us up for shared input/output segments. This is apparently
+ * necessary for our VCM setup to avoid varying corruption.
+ */
+ prog_data->separate_segments = false;
+ prog_data->vpm_output_size = MAX2(prog_data->vpm_output_size,
+ prog_data->vpm_input_size);
+ prog_data->vpm_input_size = 0;
+
+ /* Compute VCM cache size. We set up our program to take up less than
+ * half of the VPM, so that any set of bin and render programs won't
+ * run out of space. We need space for at least one input segment,
+ * and then allocate the rest to output segments (one for the current
+ * program, the rest to VCM). The valid range of the VCM cache size
+ * field is 1-4 16-vertex batches, but GFXH-1744 limits us to 2-4
+ * batches.
+ */
+ assert(c->devinfo->vpm_size);
+ int sector_size = 16 * sizeof(uint32_t) * 8;
+ int vpm_size_in_sectors = c->devinfo->vpm_size / sector_size;
+ int half_vpm = vpm_size_in_sectors / 2;
+ int vpm_output_sectors = half_vpm - prog_data->vpm_input_size;
+ int vpm_output_batches = vpm_output_sectors / prog_data->vpm_output_size;
+ assert(vpm_output_batches >= 2);
+ prog_data->vcm_cache_size = CLAMP(vpm_output_batches - 1, 2, 4);
+
return v3d_return_qpu_insts(c, final_assembly_size);
}
for (int i = 0; i < V3D_MAX_FS_INPUTS; i++) {
if (BITSET_TEST(c->flat_shade_flags, i))
prog_data->flat_shade_flags[i / 24] |= 1 << (i % 24);
+
+ if (BITSET_TEST(c->noperspective_flags, i))
+ prog_data->noperspective_flags[i / 24] |= 1 << (i % 24);
+
+ if (BITSET_TEST(c->centroid_flags, i))
+ prog_data->centroid_flags[i / 24] |= 1 << (i % 24);
}
}
struct v3d_fs_key *key,
struct v3d_fs_prog_data *prog_data,
nir_shader *s,
+ void (*debug_output)(const char *msg,
+ void *debug_output_data),
+ void *debug_output_data,
int program_id, int variant_id,
uint32_t *final_assembly_size)
{
struct v3d_compile *c = vir_compile_init(compiler, &key->base, s,
+ debug_output, debug_output_data,
program_id, variant_id);
c->fs_key = key;
v3d_lower_nir_late(c);
v3d_optimize_nir(c->s);
+ NIR_PASS_V(c->s, nir_lower_bool_to_int32);
NIR_PASS_V(c->s, nir_convert_from_ssa, true);
v3d_nir_to_vir(c);
v3d_set_fs_prog_data_inputs(c, prog_data);
prog_data->writes_z = (c->s->info.outputs_written &
(1 << FRAG_RESULT_DEPTH));
- prog_data->discard = c->s->info.fs.uses_discard;
+ prog_data->discard = (c->s->info.fs.uses_discard ||
+ c->fs_key->sample_alpha_to_coverage);
+ prog_data->uses_center_w = c->uses_center_w;
return v3d_return_qpu_insts(c, final_assembly_size);
}
return vir_reg(QFILE_UNIF, uniform);
}
+static bool
+vir_can_set_flags(struct v3d_compile *c, struct qinst *inst)
+{
+ if (c->devinfo->ver >= 40 && (v3d_qpu_reads_vpm(&inst->qpu) ||
+ v3d_qpu_uses_sfu(&inst->qpu))) {
+ return false;
+ }
+
+ if (inst->qpu.type != V3D_QPU_INSTR_TYPE_ALU ||
+ (inst->qpu.alu.add.op == V3D_QPU_A_NOP &&
+ inst->qpu.alu.mul.op == V3D_QPU_M_NOP)) {
+ return false;
+ }
+
+ return true;
+}
+
void
vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf)
{
if (src.file != QFILE_TEMP ||
!c->defs[src.index] ||
- last_inst != c->defs[src.index]) {
+ last_inst != c->defs[src.index] ||
+ !vir_can_set_flags(c, last_inst)) {
/* XXX: Make the MOV be the appropriate type */
last_inst = vir_MOV_dest(c, vir_reg(QFILE_NULL, 0), src);
}
OPTPASS(vir_opt_copy_propagate);
OPTPASS(vir_opt_dead_code);
+ OPTPASS(vir_opt_small_immediates);
if (!progress)
break;
projects / mesa.git / blobdiff