On V3D 4.x, we can use ldunifrf to load uniforms to any register, and this
will let us schedule the ldunif wherever we want in the program.
struct v3d_compiler {
const struct v3d_device_info *devinfo;
struct ra_regs *regs;
struct v3d_compiler {
const struct v3d_device_info *devinfo;
struct ra_regs *regs;
+ unsigned int reg_class_any[3];
+ unsigned int reg_class_r5[3];
unsigned int reg_class_phys[3];
unsigned int reg_class_phys_or_acc[3];
};
unsigned int reg_class_phys[3];
unsigned int reg_class_phys_or_acc[3];
};
#define QPU_R(i) { .magic = false, .index = i }
#define ACC_INDEX 0
#define QPU_R(i) { .magic = false, .index = i }
#define ACC_INDEX 0
#define PHYS_INDEX (ACC_INDEX + ACC_COUNT)
#define PHYS_COUNT 64
#define PHYS_INDEX (ACC_INDEX + ACC_COUNT)
#define PHYS_COUNT 64
struct qinst *def = c->defs[temp];
return (def &&
struct qinst *def = c->defs[temp];
return (def &&
- vir_is_raw_mov(def) &&
- def->src[0].file == QFILE_UNIF);
+ (def->qpu.sig.ldunif ||
+ (vir_is_raw_mov(def) &&
+ def->src[0].file == QFILE_UNIF)));
int start_num_temps = c->num_temps;
int start_num_temps = c->num_temps;
- struct qreg uniform_src = c->undef;
- if (is_uniform)
- uniform_src = c->defs[spill_temp]->src[0];
+ int uniform_index = ~0;
+ if (is_uniform) {
+ struct qinst *orig_unif = c->defs[spill_temp];
+ if (orig_unif->qpu.sig.ldunif) {
+ uniform_index = orig_unif->uniform;
+ } else {
+ assert(orig_unif->src[0].file == QFILE_UNIF);
+ uniform_index = orig_unif->src[0].index;
+ }
+ }
vir_for_each_inst_inorder_safe(inst, c) {
for (int i = 0; i < vir_get_nsrc(inst); i++) {
vir_for_each_inst_inorder_safe(inst, c) {
for (int i = 0; i < vir_get_nsrc(inst); i++) {
c->cursor = vir_before_inst(inst);
if (is_uniform) {
c->cursor = vir_before_inst(inst);
if (is_uniform) {
- inst->src[i] = vir_MOV(c, uniform_src);
+ inst->src[i] =
+ vir_MOV(c, vir_uniform(c,
+ c->uniform_contents[uniform_index],
+ c->uniform_data[uniform_index]));
} else {
v3d_emit_spill_tmua(c, spill_offset);
vir_emit_thrsw(c);
} else {
v3d_emit_spill_tmua(c, spill_offset);
vir_emit_thrsw(c);
v3d_ra_select_callback(struct ra_graph *g, BITSET_WORD *regs, void *data)
{
struct v3d_ra_select_callback_data *v3d_ra = data;
v3d_ra_select_callback(struct ra_graph *g, BITSET_WORD *regs, void *data)
{
struct v3d_ra_select_callback_data *v3d_ra = data;
+ int r5 = ACC_INDEX + 5;
+
+ /* Choose r5 for our ldunifs if possible (nobody else can load to that
+ * reg, and it keeps the QPU cond field free from being occupied by
+ * ldunifrf).
+ */
+ if (BITSET_TEST(regs, r5))
+ return r5;
/* Choose an accumulator if possible (I think it's lower power than
* phys regs), but round-robin through them to give post-RA
/* Choose an accumulator if possible (I think it's lower power than
* phys regs), but round-robin through them to give post-RA
return false;
for (int threads = 0; threads < max_thread_index; threads++) {
return false;
for (int threads = 0; threads < max_thread_index; threads++) {
+ compiler->reg_class_any[threads] =
+ ra_alloc_reg_class(compiler->regs);
+ compiler->reg_class_r5[threads] =
+ ra_alloc_reg_class(compiler->regs);
compiler->reg_class_phys_or_acc[threads] =
ra_alloc_reg_class(compiler->regs);
compiler->reg_class_phys[threads] =
compiler->reg_class_phys_or_acc[threads] =
ra_alloc_reg_class(compiler->regs);
compiler->reg_class_phys[threads] =
compiler->reg_class_phys_or_acc[threads], i);
ra_class_add_reg(compiler->regs,
compiler->reg_class_phys[threads], i);
compiler->reg_class_phys_or_acc[threads], i);
ra_class_add_reg(compiler->regs,
compiler->reg_class_phys[threads], i);
+ ra_class_add_reg(compiler->regs,
+ compiler->reg_class_any[threads], i);
- for (int i = ACC_INDEX + 0; i < ACC_INDEX + ACC_COUNT; i++) {
+ for (int i = ACC_INDEX + 0; i < ACC_INDEX + ACC_COUNT - 1; i++) {
ra_class_add_reg(compiler->regs,
compiler->reg_class_phys_or_acc[threads], i);
ra_class_add_reg(compiler->regs,
compiler->reg_class_phys_or_acc[threads], i);
+ ra_class_add_reg(compiler->regs,
+ compiler->reg_class_any[threads], i);
+ /* r5 can only store a single 32-bit value, so not much can
+ * use it.
+ */
+ ra_class_add_reg(compiler->regs,
+ compiler->reg_class_r5[threads],
+ ACC_INDEX + 5);
+ ra_class_add_reg(compiler->regs,
+ compiler->reg_class_any[threads],
+ ACC_INDEX + 5);
}
ra_set_finalize(compiler->regs, NULL);
}
ra_set_finalize(compiler->regs, NULL);
#define CLASS_BIT_PHYS (1 << 0)
#define CLASS_BIT_ACC (1 << 1)
#define CLASS_BIT_PHYS (1 << 0)
#define CLASS_BIT_ACC (1 << 1)
+#define CLASS_BIT_R5 (1 << 4)
+#define CLASS_BITS_ANY (CLASS_BIT_PHYS | \
+ CLASS_BIT_ACC | \
+ CLASS_BIT_R5)
/**
* Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
/**
* Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
* start with any temp being able to be in any file, then instructions
* incrementally remove bits that the temp definitely can't be in.
*/
* start with any temp being able to be in any file, then instructions
* incrementally remove bits that the temp definitely can't be in.
*/
- memset(class_bits,
- CLASS_BIT_PHYS | CLASS_BIT_ACC,
- sizeof(class_bits));
+ memset(class_bits, CLASS_BITS_ANY, sizeof(class_bits));
int ip = 0;
vir_for_each_inst_inorder(inst, c) {
int ip = 0;
vir_for_each_inst_inorder(inst, c) {
+ if (inst->dst.file == QFILE_TEMP) {
+ /* Only a ldunif gets to write to R5, which only has a
+ * single 32-bit channel of storage.
+ */
+ if (!inst->qpu.sig.ldunif) {
+ class_bits[inst->dst.index] &= ~CLASS_BIT_R5;
+ } else {
+ /* Until V3D 4.x, we could only load a uniform
+ * to r5, so we'll need to spill if uniform
+ * loads interfere with each other.
+ */
+ if (c->devinfo->ver < 40) {
+ class_bits[inst->dst.index] &=
+ CLASS_BIT_R5;
+ }
+ }
+ }
+
if (inst->qpu.sig.thrsw) {
/* All accumulators are invalidated across a thread
* switch.
if (inst->qpu.sig.thrsw) {
/* All accumulators are invalidated across a thread
* switch.
if (class_bits[i] == CLASS_BIT_PHYS) {
ra_set_node_class(g, temp_to_node[i],
c->compiler->reg_class_phys[thread_index]);
if (class_bits[i] == CLASS_BIT_PHYS) {
ra_set_node_class(g, temp_to_node[i],
c->compiler->reg_class_phys[thread_index]);
- } else {
- assert(class_bits[i] == (CLASS_BIT_PHYS |
- CLASS_BIT_ACC));
+ } else if (class_bits[i] == (CLASS_BIT_R5)) {
+ ra_set_node_class(g, temp_to_node[i],
+ c->compiler->reg_class_r5[thread_index]);
+ } else if (class_bits[i] == (CLASS_BIT_PHYS | CLASS_BIT_ACC)) {
ra_set_node_class(g, temp_to_node[i],
c->compiler->reg_class_phys_or_acc[thread_index]);
ra_set_node_class(g, temp_to_node[i],
c->compiler->reg_class_phys_or_acc[thread_index]);
+ } else {
+ assert(class_bits[i] == CLASS_BITS_ANY);
+ ra_set_node_class(g, temp_to_node[i],
+ c->compiler->reg_class_any[thread_index]);