Skip to content

Refactor: stop host_build_graph's host side reaching into device facilities #2098

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ChaoZheng109 merged 1 commit into from
Sep 2, 2026
Merged

Refactor: stop host_build_graph's host side reaching into device facilities #2098

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 4 additions & 6 deletions src/a2a3/runtime/host_build_graph/host/runtime_maker.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -630,7 +630,7 @@ int32_t run_host_orchestration(
// actually needs, and compact_live_image moves every address the orchestrator
// wrote onto the real base before the image travels.
if (!orchestrator.init(
host_sm, reinterpret_cast<void *>(HEAP_VIRTUAL_BASE), HEAP_VIRTUAL_CAPACITY, task_capacity, rt->scheduler
host_sm, reinterpret_cast<void *>(HEAP_VIRTUAL_BASE), HEAP_VIRTUAL_CAPACITY, task_capacity
)) {
LOG_ERROR("host-orch: orchestrator init against host SM failed");
return PTO_RUNTIME_ERR_INTERNAL;
Expand Down Expand Up @@ -716,14 +716,12 @@ int32_t run_host_orchestration(
{
const OrchProfilingData prof = orchestrator_get_profiling();
const std::pair<const char *, uint64_t> steps[] = {
{"alloc", prof.alloc_cycle}, {"args", prof.args_cycle}, {"lookup", prof.lookup_cycle},
{"insert", prof.insert_cycle}, {"fanin", prof.fanin_cycle},
{"alloc", prof.alloc_ns}, {"args", prof.args_ns}, {"lookup", prof.lookup_ns},
{"insert", prof.insert_ns}, {"fanin", prof.fanin_ns},
};
for (const auto &step : steps) {
if (step.second == 0) continue;
LOG_TIMING(
"host-orch step=%s cycles=%" PRIu64 " submits=%" PRId64, step.first, step.second, prof.submit_count
);
LOG_TIMING("host-orch step=%s ns=%" PRIu64 " submits=%" PRId64, step.first, step.second, prof.submit_count);
}
}
#endif
Expand Down
10 changes: 4 additions & 6 deletions src/a5/runtime/host_build_graph/host/runtime_maker.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -630,7 +630,7 @@ int32_t run_host_orchestration(
// actually needs, and compact_live_image moves every address the orchestrator
// wrote onto the real base before the image travels.
if (!orchestrator.init(
host_sm, reinterpret_cast<void *>(HEAP_VIRTUAL_BASE), HEAP_VIRTUAL_CAPACITY, task_capacity, rt->scheduler
host_sm, reinterpret_cast<void *>(HEAP_VIRTUAL_BASE), HEAP_VIRTUAL_CAPACITY, task_capacity
)) {
LOG_ERROR("host-orch: orchestrator init against host SM failed");
return PTO_RUNTIME_ERR_INTERNAL;
Expand Down Expand Up @@ -716,14 +716,12 @@ int32_t run_host_orchestration(
{
const OrchProfilingData prof = orchestrator_get_profiling();
const std::pair<const char *, uint64_t> steps[] = {
{"alloc", prof.alloc_cycle}, {"args", prof.args_cycle}, {"lookup", prof.lookup_cycle},
{"insert", prof.insert_cycle}, {"fanin", prof.fanin_cycle},
{"alloc", prof.alloc_ns}, {"args", prof.args_ns}, {"lookup", prof.lookup_ns},
{"insert", prof.insert_ns}, {"fanin", prof.fanin_ns},
};
for (const auto &step : steps) {
if (step.second == 0) continue;
LOG_TIMING(
"host-orch step=%s cycles=%" PRIu64 " submits=%" PRId64, step.first, step.second, prof.submit_count
);
LOG_TIMING("host-orch step=%s ns=%" PRIu64 " submits=%" PRId64, step.first, step.second, prof.submit_count);
}
}
#endif
Expand Down
57 changes: 0 additions & 57 deletions src/common/host_build_graph/host/aicpu_shims.cpp
View file
Open in desktop

This file was deleted.

112 changes: 53 additions & 59 deletions src/common/host_build_graph/host/orchestrator.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,10 +58,6 @@
#include "host_build_graph/types.h"
#include "tensor.h"

#if SIMPLER_DFX
#include "aicpu/args_dump_aicpu.h"
#endif

// Raises the two edge kinds compute_task_fanin can discover, for the capture
// instantiation. Shared by the ordinary submit path and the outer GRAPH task so
// both describe an edge the same way.
Expand All @@ -80,43 +76,49 @@ struct DepGraphAnnotate {
// Orchestrator Profiling (compile-time toggle)
// =============================================================================
#if SIMPLER_ORCH_PROFILING
#include "aicpu/device_time.h"
#include "aicpu/chip_swimlane_collector_aicpu.h"
// Accumulated cycles per sub-step (only needed for ORCH_PROFILING export)
static uint64_t g_orch_alloc_cycle = 0; // unified task+heap alloc
static uint64_t g_orch_args_cycle = 0; // param copy
static uint64_t g_orch_lookup_cycle = 0; // tensormap lookup + dep building
static uint64_t g_orch_insert_cycle = 0; // tensormap insert
static uint64_t g_orch_fanin_cycle = 0; // fanin list + early-return check
// Accumulated nanoseconds per sub-step (only needed for ORCH_PROFILING export)
static uint64_t g_orch_alloc_ns = 0; // unified task+heap alloc
static uint64_t g_orch_args_ns = 0; // param copy
static uint64_t g_orch_lookup_ns = 0; // tensormap lookup + dep building
static uint64_t g_orch_insert_ns = 0; // tensormap insert
static uint64_t g_orch_fanin_ns = 0; // fanin list + early-return check
static int64_t g_orch_submit_count = 0;
static uint32_t g_orch_submit_idx = 0;
uint64_t g_orch_fanin_wait_cycle = 0;
uint64_t g_orch_fanin_wait_ns = 0;
uint64_t g_orch_args_atomic_count = 0;
// Cycle accumulation is unconditional under SIMPLER_ORCH_PROFILING (that's what
// the flag is for) and feeds the per-sub-step `g_orch_*_cycle` cumulatives
// printed in the cold-path log. Per-event records are a separate channel on a
// separate clock — see ORCH_PHASE_END below.
#define CYCLE_COUNT_START() \
uint64_t _t0 = get_sys_cnt_aicpu(), _t1; \

// The orchestrator runs on the host, so its sub-steps are timed by the host's own
// monotonic clock. Static, so the timing costs a call to clock_gettime and no
// symbol resolution.
static inline uint64_t orch_now_ns() {
struct timespec ts;
clock_gettime(CLOCK_MONOTONIC, &ts);
return static_cast<uint64_t>(ts.tv_sec) * 1000000000ull + static_cast<uint64_t>(ts.tv_nsec);
}

// Accumulation is unconditional under SIMPLER_ORCH_PROFILING (that's what the flag
// is for) and feeds the per-sub-step `g_orch_*_ns` cumulatives printed in the
// cold-path log. Per-event records are a separate channel on a separate clock —
// see ORCH_PHASE_END below.
#define ORCH_STEP_START() \
uint64_t _t0 = orch_now_ns(), _t1; \
(void)_t1
#define CYCLE_COUNT_LAP(acc) \
do { \
_t1 = get_sys_cnt_aicpu(); \
acc += (_t1 - _t0); \
_t0 = _t1; \
#define ORCH_STEP_LAP(acc) \
do { \
_t1 = orch_now_ns(); \
acc += (_t1 - _t0); \
_t0 = _t1; \
} while (0)
#elif SIMPLER_DFX
#include "aicpu/device_time.h"
#include "aicpu/chip_swimlane_collector_aicpu.h"
// submit_idx tags a record with its position in the orchestration's submit order.
static uint32_t g_orch_submit_idx = 0;
// The per-sub-step accumulators exist only in an ORCH_PROFILING build, so at this
// level there is nothing to time.
#define CYCLE_COUNT_START()
#define CYCLE_COUNT_LAP(acc)
#define ORCH_STEP_START()
#define ORCH_STEP_LAP(acc)
#else
#define CYCLE_COUNT_START()
#define CYCLE_COUNT_LAP(acc)
#define ORCH_STEP_START()
#define ORCH_STEP_LAP(acc)
#endif

#if SIMPLER_DFX
Expand Down Expand Up @@ -207,9 +209,7 @@ void OrchestratorState::report_fatal(int32_t error_code, const char *func, const
va_end(args);
}

bool OrchestratorState::init(
void *sm_base, void *gm_heap, uint64_t heap_size, uint64_t max_tasks, SchedulerState *scheduler_arg
) {
bool OrchestratorState::init(void *sm_base, void *gm_heap, uint64_t heap_size, uint64_t max_tasks) {
// Reset in place rather than by move-assignment: fatal_code is a std::atomic,
// which is neither copy- nor move-assignable, and a re-init has to clear every
// field the previous pass left behind (the pool cursors below rely on it).
Expand All @@ -219,7 +219,6 @@ bool OrchestratorState::init(
always_assert(max_tasks > 0);

orch->sm_header = reinterpret_cast<SharedMemoryHeader *>(sm_base);
orch->scheduler = scheduler_arg;

orch->task_allocator.init(static_cast<int32_t>(max_tasks), gm_heap, heap_size, &orch->fatal_code);

Expand Down Expand Up @@ -1613,7 +1612,7 @@ static TaskOutputTensors submit_task_common(
OrchestratorState *orch, const CoreTaskArgs &args, ActiveMask active_mask, TaskAttrs task_attrs,
int32_t aic_kernel_id, int32_t aiv0_kernel_id, int32_t aiv1_kernel_id
) {
CYCLE_COUNT_START();
ORCH_STEP_START();
ORCH_PHASE_START();
TaskOutputTensors result;
DispatchPredicate resolved_predicate{};
Expand All @@ -1623,7 +1622,6 @@ static TaskOutputTensors submit_task_common(
if (!prepare_task(orch, args, layout.total_output_size, active_mask, task_attrs, &prepared)) {
return result;
}
SchedulerState *sched = orch->scheduler;
TaskId task_id = prepared.task_id;
TaskDescriptor &task = *prepared.task;
TaskPayload &payload = *prepared.payload;
Expand Down Expand Up @@ -1655,7 +1653,7 @@ static TaskOutputTensors submit_task_common(
int32_t *fanin_slots = payload.fanin_data();
payload.fanin_count = 0;

CYCLE_COUNT_LAP(g_orch_alloc_cycle);
ORCH_STEP_LAP(g_orch_alloc_ns);

#if SIMPLER_DFX
if (layout.total_output_size > 0) {
Expand Down Expand Up @@ -1707,7 +1705,7 @@ static TaskOutputTensors submit_task_common(
}
}

CYCLE_COUNT_LAP(g_orch_lookup_cycle);
ORCH_STEP_LAP(g_orch_lookup_ns);

// === STEP 4: Register outputs/inouts in TensorMap (must be separate from lookup) ===
// Reserve pool capacity for this task's inserts before registering, so an
Expand All @@ -1719,7 +1717,7 @@ static TaskOutputTensors submit_task_common(
}
register_task_outputs(dep_inputs, task_id, orch->tensor_map, orch->in_manual_scope());

CYCLE_COUNT_LAP(g_orch_insert_cycle);
ORCH_STEP_LAP(g_orch_insert_ns);

// === STEP 5: Batch-write to GM (single cache line burst) ===
// Deferred from allocation phase to avoid scattered GM writes that get
Expand All @@ -1741,7 +1739,7 @@ static TaskOutputTensors submit_task_common(
// Predicate validation runs before task allocation. Copy the resolved, bounded
// operand address into the device payload only after the rest of the payload exists.
payload.predicate = resolved_predicate;
CYCLE_COUNT_LAP(g_orch_args_cycle);
ORCH_STEP_LAP(g_orch_args_ns);

// === STEP 6: close the fanin region (device boot classifies) ===
// Polling + host-orch: append_fanin_or_fail already wrote each producer's local
Expand All @@ -1761,9 +1759,7 @@ static TaskOutputTensors submit_task_common(
debug_assert(orch->fanin_pool_cursor == static_cast<int32_t>(payload.fanin_data() - orch->fanin_pool));
orch->fanin_pool_cursor += CHIP_ALIGN_UP(payload.fanin_count, ARG_POOL_ALIGN / (int32_t)sizeof(int32_t));

(void)sched;

CYCLE_COUNT_LAP(g_orch_fanin_cycle);
ORCH_STEP_LAP(g_orch_fanin_ns);
ORCH_PHASE_END(HostPhaseKind::OrchSubmitTask, task_id.raw);

#if SIMPLER_DFX
Expand Down Expand Up @@ -2717,7 +2713,6 @@ TaskOutputTensors OrchestratorState::submit_task(const MixedKernels &mixed_kerne
orch_mark_fatal(orch, SIMPLER_ERROR_INVALID_ARGS);
return TaskOutputTensors{};
}
always_assert(orch->scheduler != nullptr);
// === Validate submit inputs ===
ActiveMask active_mask = mixed_kernels.to_active_mask();
if (!static_cast<bool>(active_mask)) {
Expand Down Expand Up @@ -2805,7 +2800,6 @@ TaskOutputTensors OrchestratorState::submit_dummy_task(const CoreTaskArgs &args)
orch_mark_fatal(orch, SIMPLER_ERROR_INVALID_ARGS);
return TaskOutputTensors{};
}
always_assert(orch->scheduler != nullptr);

// Dummy tasks never dispatch to an AICore, so sync_start / has_predicate do
// not apply; only the early-dispatch hint and timing tag carry over.
Expand Down Expand Up @@ -2851,7 +2845,7 @@ TaskOutputTensors OrchestratorState::alloc_tensors(const CoreTaskArgs &args) {
}
}

CYCLE_COUNT_START();
ORCH_STEP_START();
ORCH_PHASE_START();

if (args.has_error) {
Expand Down Expand Up @@ -2883,7 +2877,7 @@ TaskOutputTensors OrchestratorState::alloc_tensors(const CoreTaskArgs &args) {
TaskDescriptor &task = *prepared.task;
TaskPayload &payload = *prepared.payload;

CYCLE_COUNT_LAP(g_orch_alloc_cycle);
ORCH_STEP_LAP(g_orch_alloc_ns);

#if SIMPLER_DFX
if (layout.total_output_size > 0) {
Expand All @@ -2903,7 +2897,7 @@ TaskOutputTensors OrchestratorState::alloc_tensors(const CoreTaskArgs &args) {
outputs.set_task_id(prepared.task_id);
payload.init(args, outputs, prepared.alloc_result, layout);
payload.fanin_count = 0; // hidden-alloc tasks have no producer dependencies
CYCLE_COUNT_LAP(g_orch_args_cycle);
ORCH_STEP_LAP(g_orch_args_ns);

if (prepared.slot_state != nullptr) {
// Hidden alloc tasks complete inline in the orchestrator before any
Expand Down Expand Up @@ -2935,7 +2929,7 @@ TaskOutputTensors OrchestratorState::alloc_tensors(const CoreTaskArgs &args) {
}
orch->inline_completed_tasks++;

CYCLE_COUNT_LAP(g_orch_fanin_cycle);
ORCH_STEP_LAP(g_orch_fanin_ns);
ORCH_PHASE_END(HostPhaseKind::OrchAllocTensors, prepared.task_id.raw);

#if SIMPLER_DFX
Expand Down Expand Up @@ -2969,22 +2963,22 @@ void OrchestratorState::mark_done() {
#if SIMPLER_ORCH_PROFILING
OrchProfilingData orchestrator_get_profiling() {
OrchProfilingData d;
d.alloc_cycle = g_orch_alloc_cycle;
d.args_cycle = g_orch_args_cycle;
d.lookup_cycle = g_orch_lookup_cycle;
d.insert_cycle = g_orch_insert_cycle;
d.fanin_cycle = g_orch_fanin_cycle;
d.alloc_ns = g_orch_alloc_ns;
d.args_ns = g_orch_args_ns;
d.lookup_ns = g_orch_lookup_ns;
d.insert_ns = g_orch_insert_ns;
d.fanin_ns = g_orch_fanin_ns;
d.submit_count = g_orch_submit_count;
d.fanin_wait_cycle = g_orch_fanin_wait_cycle;
d.fanin_wait_ns = g_orch_fanin_wait_ns;
d.args_atomic_count = g_orch_args_atomic_count;

// Reset
g_orch_alloc_cycle = g_orch_args_cycle = 0;
g_orch_lookup_cycle = g_orch_insert_cycle = 0;
g_orch_fanin_cycle = 0;
g_orch_alloc_ns = g_orch_args_ns = 0;
g_orch_lookup_ns = g_orch_insert_ns = 0;
g_orch_fanin_ns = 0;
g_orch_submit_count = 0;
g_orch_submit_idx = 0;
g_orch_fanin_wait_cycle = 0;
g_orch_fanin_wait_ns = 0;
g_orch_args_atomic_count = 0;
return d;
}
Expand Down
Loading
Loading