struct schedule_node_child;
struct schedule_node {
- struct simple_node link;
+ struct list_head link;
struct queued_qpu_inst *inst;
struct schedule_node_child *children;
uint32_t child_count;
uint32_t child_array_size;
uint32_t parent_count;
+
+ /* Longest cycles + instruction_latency() of any parent of this node. */
+ uint32_t unblocked_time;
+
+ /**
+ * Minimum number of cycles from scheduling this instruction until the
+ * end of the program, based on the slowest dependency chain through
+ * the children.
+ */
uint32_t delay;
+
+ /**
+ * cycles between this instruction being scheduled and when its result
+ * can be consumed.
+ */
+ uint32_t latency;
+
+ /**
+ * Which uniform from uniform_data[] this instruction read, or -1 if
+ * not reading a uniform.
+ */
+ int uniform;
};
struct schedule_node_child {
struct schedule_node *last_rb[32];
struct schedule_node *last_sf;
struct schedule_node *last_vpm_read;
- struct schedule_node *last_unif_read;
struct schedule_node *last_tmu_write;
struct schedule_node *last_tlb;
struct schedule_node *last_vpm;
enum direction dir;
+ /* Estimated cycle when the current instruction would start. */
+ uint32_t time;
};
static void
break;
case QPU_R_UNIF:
- add_write_dep(state, &state->last_unif_read, n);
- break;
-
case QPU_R_NOP:
case QPU_R_ELEM_QPU:
case QPU_R_XY_PIXEL_COORD:
}
}
+static bool
+reads_uniform(uint64_t inst)
+{
+ if (QPU_GET_FIELD(inst, QPU_SIG) == QPU_SIG_LOAD_IMM)
+ return false;
+
+ return (QPU_GET_FIELD(inst, QPU_RADDR_A) == QPU_R_UNIF ||
+ (QPU_GET_FIELD(inst, QPU_RADDR_B) == QPU_R_UNIF &&
+ QPU_GET_FIELD(inst, QPU_SIG) != QPU_SIG_SMALL_IMM) ||
+ is_tmu_write(QPU_GET_FIELD(inst, QPU_WADDR_ADD)) ||
+ is_tmu_write(QPU_GET_FIELD(inst, QPU_WADDR_MUL)));
+}
+
static void
process_mux_deps(struct schedule_state *state, struct schedule_node *n,
uint32_t mux)
}
-static bool
-is_direct_tmu_read(uint64_t inst)
-{
- /* If it's a direct read, we happen to structure the code such that
- * there's an explicit uniform read in the instruction (for kernel
- * texture reloc processing).
- */
- return (QPU_GET_FIELD(inst, QPU_RADDR_A) == QPU_R_UNIF ||
- QPU_GET_FIELD(inst, QPU_RADDR_B) == QPU_R_UNIF);
-}
-
static void
process_waddr_deps(struct schedule_state *state, struct schedule_node *n,
uint32_t waddr, bool is_add)
}
} else if (is_tmu_write(waddr)) {
add_write_dep(state, &state->last_tmu_write, n);
-
- /* There is an implicit uniform read in texture ops in
- * hardware, unless this is a direct-addressed uniform read,
- * so we need to keep it in the same order as the other
- * uniforms.
- */
- if (!is_direct_tmu_read(n->inst->inst))
- add_write_dep(state, &state->last_unif_read, n);
- } else if (qpu_waddr_is_tlb(waddr)) {
+ } else if (qpu_waddr_is_tlb(waddr) ||
+ waddr == QPU_W_MS_FLAGS) {
add_write_dep(state, &state->last_tlb, n);
} else {
switch (waddr) {
break;
case QPU_W_VPM:
- case QPU_W_VPMVCD_SETUP:
add_write_dep(state, &state->last_vpm, n);
break;
+ case QPU_W_VPMVCD_SETUP:
+ if (is_a)
+ add_write_dep(state, &state->last_vpm_read, n);
+ else
+ add_write_dep(state, &state->last_vpm, n);
+ break;
+
case QPU_W_SFU_RECIP:
case QPU_W_SFU_RECIPSQRT:
case QPU_W_SFU_EXP:
add_write_dep(state, &state->last_tlb, n);
break;
+ case QPU_W_MS_FLAGS:
+ add_write_dep(state, &state->last_tlb, n);
+ break;
+
case QPU_W_NOP:
break;
uint32_t mul_b = QPU_GET_FIELD(inst, QPU_MUL_B);
uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
- process_raddr_deps(state, n, raddr_a, true);
- process_raddr_deps(state, n, raddr_b, false);
+ if (sig != QPU_SIG_LOAD_IMM) {
+ process_raddr_deps(state, n, raddr_a, true);
+ if (sig != QPU_SIG_SMALL_IMM)
+ process_raddr_deps(state, n, raddr_b, false);
+ }
+
if (add_op != QPU_A_NOP) {
process_mux_deps(state, n, add_a);
process_mux_deps(state, n, add_b);
}
process_cond_deps(state, n, QPU_GET_FIELD(inst, QPU_COND_ADD));
- process_cond_deps(state, n, QPU_GET_FIELD(inst, QPU_COND_ADD));
+ process_cond_deps(state, n, QPU_GET_FIELD(inst, QPU_COND_MUL));
if (inst & QPU_SF)
add_write_dep(state, &state->last_sf, n);
}
static void
-calculate_forward_deps(struct vc4_compile *c, struct simple_node *schedule_list)
+calculate_forward_deps(struct vc4_compile *c, struct list_head *schedule_list)
{
- struct simple_node *node;
struct schedule_state state;
memset(&state, 0, sizeof(state));
state.dir = F;
- foreach(node, schedule_list)
- calculate_deps(&state, (struct schedule_node *)node);
+ list_for_each_entry(struct schedule_node, node, schedule_list, link)
+ calculate_deps(&state, node);
}
static void
-calculate_reverse_deps(struct vc4_compile *c, struct simple_node *schedule_list)
+calculate_reverse_deps(struct vc4_compile *c, struct list_head *schedule_list)
{
- struct simple_node *node;
+ struct list_head *node;
struct schedule_state state;
memset(&state, 0, sizeof(state));
{
uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
+ uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
uint32_t src_muxes[] = {
QPU_GET_FIELD(inst, QPU_ADD_A),
QPU_GET_FIELD(inst, QPU_ADD_B),
raddr_a < 32 &&
scoreboard->last_waddr_a == raddr_a) ||
(src_muxes[i] == QPU_MUX_B &&
+ sig != QPU_SIG_SMALL_IMM &&
raddr_b < 32 &&
scoreboard->last_waddr_b == raddr_b)) {
return true;
static struct schedule_node *
choose_instruction_to_schedule(struct choose_scoreboard *scoreboard,
- struct simple_node *schedule_list,
- uint64_t prev_inst)
+ struct list_head *schedule_list,
+ struct schedule_node *prev_inst)
{
struct schedule_node *chosen = NULL;
- struct simple_node *node;
int chosen_prio = 0;
- foreach(node, schedule_list) {
- struct schedule_node *n = (struct schedule_node *)node;
+ list_for_each_entry(struct schedule_node, n, schedule_list, link) {
uint64_t inst = n->inst->inst;
/* "An instruction must not read from a location in physical
/* If we're trying to pair with another instruction, check
* that they're compatible.
*/
- if (prev_inst != 0) {
- inst = qpu_merge_inst(prev_inst, inst);
+ if (prev_inst) {
+ if (prev_inst->uniform != -1 && n->uniform != -1)
+ continue;
+
+ inst = qpu_merge_inst(prev_inst->inst->inst, inst);
if (!inst)
continue;
}
} else if (prio < chosen_prio) {
continue;
}
+
+ if (n->delay > chosen->delay) {
+ chosen = n;
+ chosen_prio = prio;
+ } else if (n->delay < chosen->delay) {
+ continue;
+ }
}
return chosen;
}
static void
-dump_state(struct simple_node *schedule_list)
+dump_state(struct list_head *schedule_list)
{
- struct simple_node *node;
-
- uint32_t i = 0;
- foreach(node, schedule_list) {
- struct schedule_node *n = (struct schedule_node *)node;
-
- fprintf(stderr, "%3d: ", i++);
+ list_for_each_entry(struct schedule_node, n, schedule_list, link) {
+ fprintf(stderr, " t=%4d: ", n->unblocked_time);
vc4_qpu_disasm(&n->inst->inst, 1);
fprintf(stderr, "\n");
if (!child)
continue;
- fprintf(stderr, " - ");
+ fprintf(stderr, " - ");
vc4_qpu_disasm(&child->inst->inst, 1);
fprintf(stderr, " (%d parents, %c)\n",
child->parent_count,
}
}
+static uint32_t waddr_latency(uint32_t waddr, uint64_t after)
+{
+ if (waddr < 32)
+ return 2;
+
+ /* Apply some huge latency between texture fetch requests and getting
+ * their results back.
+ */
+ if (waddr == QPU_W_TMU0_S) {
+ if (QPU_GET_FIELD(after, QPU_SIG) == QPU_SIG_LOAD_TMU0)
+ return 100;
+ }
+ if (waddr == QPU_W_TMU1_S) {
+ if (QPU_GET_FIELD(after, QPU_SIG) == QPU_SIG_LOAD_TMU1)
+ return 100;
+ }
+
+ switch(waddr) {
+ case QPU_W_SFU_RECIP:
+ case QPU_W_SFU_RECIPSQRT:
+ case QPU_W_SFU_EXP:
+ case QPU_W_SFU_LOG:
+ return 3;
+ default:
+ return 1;
+ }
+}
+
+static uint32_t
+instruction_latency(struct schedule_node *before, struct schedule_node *after)
+{
+ uint64_t before_inst = before->inst->inst;
+ uint64_t after_inst = after->inst->inst;
+
+ return MAX2(waddr_latency(QPU_GET_FIELD(before_inst, QPU_WADDR_ADD),
+ after_inst),
+ waddr_latency(QPU_GET_FIELD(before_inst, QPU_WADDR_MUL),
+ after_inst));
+}
+
/** Recursive computation of the delay member of a node. */
static void
compute_delay(struct schedule_node *n)
if (!n->children[i].node->delay)
compute_delay(n->children[i].node);
n->delay = MAX2(n->delay,
- n->children[i].node->delay + 1);
+ n->children[i].node->delay +
+ instruction_latency(n, n->children[i].node));
}
}
}
static void
-mark_instruction_scheduled(struct simple_node *schedule_list,
+mark_instruction_scheduled(struct list_head *schedule_list,
+ uint32_t time,
struct schedule_node *node,
bool war_only)
{
if (war_only && !node->children[i].write_after_read)
continue;
+ /* If the requirement is only that the node not appear before
+ * the last read of its destination, then it can be scheduled
+ * immediately after (or paired with!) the thing reading the
+ * destination.
+ */
+ uint32_t latency = 0;
+ if (!war_only) {
+ latency = instruction_latency(node,
+ node->children[i].node);
+ }
+
+ child->unblocked_time = MAX2(child->unblocked_time,
+ time + latency);
child->parent_count--;
if (child->parent_count == 0)
- insert_at_head(schedule_list, &child->link);
+ list_add(&child->link, schedule_list);
node->children[i].node = NULL;
}
}
-static void
-schedule_instructions(struct vc4_compile *c, struct simple_node *schedule_list)
+static uint32_t
+schedule_instructions(struct vc4_compile *c, struct list_head *schedule_list,
+ enum quniform_contents *orig_uniform_contents,
+ uint32_t *orig_uniform_data,
+ uint32_t *next_uniform)
{
- struct simple_node *node, *t;
struct choose_scoreboard scoreboard;
+ uint32_t time = 0;
memset(&scoreboard, 0, sizeof(scoreboard));
scoreboard.last_waddr_a = ~0;
}
/* Remove non-DAG heads from the list. */
- foreach_s(node, t, schedule_list) {
- struct schedule_node *n = (struct schedule_node *)node;
-
+ list_for_each_entry_safe(struct schedule_node, n, schedule_list, link) {
if (n->parent_count != 0)
- remove_from_list(&n->link);
+ list_del(&n->link);
}
- while (!is_empty_list(schedule_list)) {
+ while (!list_empty(schedule_list)) {
struct schedule_node *chosen =
choose_instruction_to_schedule(&scoreboard,
schedule_list,
- 0);
+ NULL);
struct schedule_node *merge = NULL;
/* If there are no valid instructions to schedule, drop a NOP
uint64_t inst = chosen ? chosen->inst->inst : qpu_NOP();
if (debug) {
- fprintf(stderr, "current list:\n");
+ fprintf(stderr, "t=%4d: current list:\n",
+ time);
dump_state(schedule_list);
- fprintf(stderr, "chose: ");
+ fprintf(stderr, "t=%4d: chose: ", time);
vc4_qpu_disasm(&inst, 1);
fprintf(stderr, "\n");
}
* find an instruction to pair with it.
*/
if (chosen) {
- remove_from_list(&chosen->link);
- mark_instruction_scheduled(schedule_list, chosen, true);
+ time = MAX2(chosen->unblocked_time, time);
+ list_del(&chosen->link);
+ mark_instruction_scheduled(schedule_list, time,
+ chosen, true);
+ if (chosen->uniform != -1) {
+ c->uniform_data[*next_uniform] =
+ orig_uniform_data[chosen->uniform];
+ c->uniform_contents[*next_uniform] =
+ orig_uniform_contents[chosen->uniform];
+ (*next_uniform)++;
+ }
merge = choose_instruction_to_schedule(&scoreboard,
schedule_list,
- inst);
+ chosen);
if (merge) {
- remove_from_list(&merge->link);
+ time = MAX2(merge->unblocked_time, time);
+ list_del(&merge->link);
inst = qpu_merge_inst(inst, merge->inst->inst);
assert(inst != 0);
+ if (merge->uniform != -1) {
+ c->uniform_data[*next_uniform] =
+ orig_uniform_data[merge->uniform];
+ c->uniform_contents[*next_uniform] =
+ orig_uniform_contents[merge->uniform];
+ (*next_uniform)++;
+ }
if (debug) {
- fprintf(stderr, "merging: ");
+ fprintf(stderr, "t=%4d: merging: ",
+ time);
vc4_qpu_disasm(&merge->inst->inst, 1);
fprintf(stderr, "\n");
- fprintf(stderr, "resulting in: ");
+ fprintf(stderr, " resulting in: ");
vc4_qpu_disasm(&inst, 1);
fprintf(stderr, "\n");
}
* be scheduled. Update the children's unblocked time for this
* DAG edge as we do so.
*/
- mark_instruction_scheduled(schedule_list, chosen, false);
- mark_instruction_scheduled(schedule_list, merge, false);
+ mark_instruction_scheduled(schedule_list, time, chosen, false);
+ mark_instruction_scheduled(schedule_list, time, merge, false);
scoreboard.tick++;
+ time++;
}
+
+ return time;
}
-void
-qpu_schedule_instructions(struct vc4_compile *c)
+static uint32_t
+qpu_schedule_instructions_block(struct vc4_compile *c, struct qblock *block,
+ enum quniform_contents *orig_uniform_contents,
+ uint32_t *orig_uniform_data,
+ uint32_t *next_uniform)
{
void *mem_ctx = ralloc_context(NULL);
- struct simple_node schedule_list;
- struct simple_node *node;
+ struct list_head schedule_list;
- make_empty_list(&schedule_list);
-
- if (debug) {
- fprintf(stderr, "Pre-schedule instructions\n");
- foreach(node, &c->qpu_inst_list) {
- struct queued_qpu_inst *q =
- (struct queued_qpu_inst *)node;
- vc4_qpu_disasm(&q->inst, 1);
- fprintf(stderr, "\n");
- }
- fprintf(stderr, "\n");
- }
+ list_inithead(&schedule_list);
/* Wrap each instruction in a scheduler structure. */
- while (!is_empty_list(&c->qpu_inst_list)) {
+ uint32_t next_sched_uniform = *next_uniform;
+ while (!list_empty(&block->qpu_inst_list)) {
struct queued_qpu_inst *inst =
- (struct queued_qpu_inst *)c->qpu_inst_list.next;
+ (struct queued_qpu_inst *)block->qpu_inst_list.next;
struct schedule_node *n = rzalloc(mem_ctx, struct schedule_node);
n->inst = inst;
- remove_from_list(&inst->link);
- insert_at_tail(&schedule_list, &n->link);
+
+ if (reads_uniform(inst->inst)) {
+ n->uniform = next_sched_uniform++;
+ } else {
+ n->uniform = -1;
+ }
+ list_del(&inst->link);
+ list_addtail(&n->link, &schedule_list);
}
calculate_forward_deps(c, &schedule_list);
calculate_reverse_deps(c, &schedule_list);
- foreach(node, &schedule_list) {
- struct schedule_node *n = (struct schedule_node *)node;
+ list_for_each_entry(struct schedule_node, n, &schedule_list, link) {
compute_delay(n);
}
- schedule_instructions(c, &schedule_list);
+ uint32_t cycles = schedule_instructions(c, &schedule_list,
+ orig_uniform_contents,
+ orig_uniform_data,
+ next_uniform);
+
+ ralloc_free(mem_ctx);
+
+ return cycles;
+}
+
+uint32_t
+qpu_schedule_instructions(struct vc4_compile *c)
+{
+ /* We reorder the uniforms as we schedule instructions, so save the
+ * old data off and replace it.
+ */
+ uint32_t *uniform_data = c->uniform_data;
+ enum quniform_contents *uniform_contents = c->uniform_contents;
+ c->uniform_contents = ralloc_array(c, enum quniform_contents,
+ c->num_uniforms);
+ c->uniform_data = ralloc_array(c, uint32_t, c->num_uniforms);
+ c->uniform_array_size = c->num_uniforms;
+ uint32_t next_uniform = 0;
+
+ if (debug) {
+ fprintf(stderr, "Pre-schedule instructions\n");
+ qir_for_each_block(block, c) {
+ fprintf(stderr, "BLOCK %d\n", block->index);
+ list_for_each_entry(struct queued_qpu_inst, q,
+ &block->qpu_inst_list, link) {
+ vc4_qpu_disasm(&q->inst, 1);
+ fprintf(stderr, "\n");
+ }
+ }
+ fprintf(stderr, "\n");
+ }
+
+ uint32_t cycles = 0;
+ qir_for_each_block(block, c) {
+ cycles += qpu_schedule_instructions_block(c, block,
+ uniform_contents,
+ uniform_data,
+ &next_uniform);
+ }
+
+ assert(next_uniform == c->num_uniforms);
if (debug) {
fprintf(stderr, "Post-schedule instructions\n");
fprintf(stderr, "\n");
}
- ralloc_free(mem_ctx);
+ return cycles;
}