PJRT: add conversion of profiling protobuf output to JSON format.
This commit is contained in:
parent
f7bac1af10
commit
4681ce2f24
@ -1,6 +1,8 @@
|
||||
load("@rules_zig//zig:defs.bzl", "zig_library")
|
||||
load("@zml//bazel:zig.bzl", "zig_cc_binary")
|
||||
load("//bazel:zig_proto_library.bzl", "zig_proto_library")
|
||||
|
||||
|
||||
cc_library(
|
||||
name = "dlfcn",
|
||||
hdrs = ["dlfcn.h"],
|
||||
@ -27,3 +29,27 @@ zig_proto_library(
|
||||
import_name = "//tsl:profiler_options_proto",
|
||||
deps = ["@tsl//tsl/profiler/protobuf:profiler_options_proto"],
|
||||
)
|
||||
|
||||
zig_proto_library(
|
||||
name = "xplane_proto",
|
||||
import_name = "//tsl:xplane_proto",
|
||||
deps = ["@tsl//tsl/profiler/protobuf:xplane_proto"],
|
||||
)
|
||||
|
||||
zig_proto_library(
|
||||
name = "trace_events_proto",
|
||||
import_name = "//tsl:trace_events_proto",
|
||||
deps = ["@tsl//tsl/profiler/protobuf:trace_events_proto"],
|
||||
)
|
||||
|
||||
zig_cc_binary(
|
||||
name = "xspace_to_json",
|
||||
srcs = glob(["convert/*.zig"]),
|
||||
main = "xspace_to_json.zig",
|
||||
visibility = ["//visibility:public"],
|
||||
deps = [
|
||||
":trace_events_proto",
|
||||
":xplane_proto",
|
||||
"//examples/third_party/tigerbeetle:flags",
|
||||
],
|
||||
)
|
||||
|
||||
323
pjrt/convert/trace_container.zig
Normal file
323
pjrt/convert/trace_container.zig
Normal file
@ -0,0 +1,323 @@
|
||||
const std = @import("std");
|
||||
const trace_events_proto = @import("//tsl:trace_events_proto");
|
||||
const xplane_proto = @import("//tsl:xplane_proto");
|
||||
const xplane_schema = @import("xplane_schema.zig");
|
||||
const xplane_visitor = @import("xplane_visitor.zig");
|
||||
|
||||
// Constants used as trace_viewer PID (device_id in trace_events.proto).
|
||||
// PID 0 is unused.
|
||||
// Support up to 500 accelerator devices.
|
||||
const first_device_id = 1;
|
||||
const last_device_id = 500;
|
||||
// Support Upto 200 custom planes as fake devices (i.e., planes with a
|
||||
// "/custom:" prefix). See `<project_name>::custom_plane_prefix` for more
|
||||
// information
|
||||
const first_custom_plane_device_id = last_device_id + 1;
|
||||
const max_custom_plane_devices_per_host = 200;
|
||||
const last_custom_plane_device_id = first_custom_plane_device_id + max_custom_plane_devices_per_host - 1;
|
||||
|
||||
// Host threads are shown as a single fake device.
|
||||
pub const host_threads_device_id = last_custom_plane_device_id + 1;
|
||||
|
||||
pub const xla_async_op_line_name = "Async XLA Ops";
|
||||
|
||||
pub const host_threads_plane_name = "/host:CPU";
|
||||
pub const gpu_plane_prefix = "/device:GPU:";
|
||||
pub const tpu_plane_prefix = "/device:TPU:";
|
||||
pub const custom_plane_prefix = "/device:CUSTOM:";
|
||||
|
||||
pub const TraceContainer = struct {
|
||||
arena: std.heap.ArenaAllocator,
|
||||
events: std.ArrayListUnmanaged(TraceEvent) = .{},
|
||||
devices: std.AutoArrayHashMapUnmanaged(u32, Device) = .{},
|
||||
|
||||
pub const Device = struct {
|
||||
name: []const u8 = &[_]u8{},
|
||||
device_id: u32 = 0,
|
||||
resources: std.AutoArrayHashMapUnmanaged(u32, trace_events_proto.Resource) = .{},
|
||||
};
|
||||
|
||||
pub const TraceEvent = struct {
|
||||
device_id: u32 = 0,
|
||||
resource_id: u32 = 0,
|
||||
name: []const u8 = &[_]u8{},
|
||||
timestamp_ps: u128 = 0,
|
||||
duration_ps: u64 = 0,
|
||||
args: std.StringArrayHashMapUnmanaged([]const u8) = .{},
|
||||
};
|
||||
|
||||
pub fn init(allocator: std.mem.Allocator, pb_buffer: []const u8, max_events: ?usize) !TraceContainer {
|
||||
var self: TraceContainer = .{
|
||||
.arena = std.heap.ArenaAllocator.init(allocator),
|
||||
};
|
||||
const arena = self.arena.allocator();
|
||||
|
||||
const xspace = try xplane_proto.XSpace.decode(pb_buffer, arena);
|
||||
try self.fromXSpace(arena, &xspace, max_events);
|
||||
return self;
|
||||
}
|
||||
|
||||
pub fn deinit(self: *TraceContainer) void {
|
||||
self.arena.deinit();
|
||||
}
|
||||
|
||||
fn picoToMicro(p: anytype) f64 {
|
||||
return @as(f64, @floatFromInt(p)) / 1E6;
|
||||
}
|
||||
|
||||
fn xLineDisplayId(xline: *const xplane_proto.XLine) i64 {
|
||||
return if (xline.display_id != 0) xline.display_id else xline.id;
|
||||
}
|
||||
|
||||
fn xLineDisplayName(xline: *const xplane_proto.XLine) []const u8 {
|
||||
return switch (xline.display_name) {
|
||||
.Empty => xline.name.getSlice(),
|
||||
else => xline.display_name.getSlice(),
|
||||
};
|
||||
}
|
||||
|
||||
fn xstatValueToString(stat: *const xplane_proto.XStat, plane: *const xplane_visitor.XPlaneVisitor, writer: std.io.AnyWriter) !void {
|
||||
if (stat.value) |val| {
|
||||
switch (val) {
|
||||
inline .int64_value, .uint64_value, .double_value => |v| try writer.print("{d}", .{v}),
|
||||
.str_value => |*v| try writer.writeAll(v.getSlice()),
|
||||
.bytes_value => try writer.writeAll("<opaque bytes>"),
|
||||
.ref_value => |v| try writer.writeAll(plane.getStatMetadataName(@intCast(v))),
|
||||
}
|
||||
} else return;
|
||||
}
|
||||
|
||||
fn findPlaneWithName(space: *const xplane_proto.XSpace, name: []const u8) ?*xplane_proto.XPlane {
|
||||
for (space.planes.items) |*v| {
|
||||
if (std.mem.eql(u8, v.name.getSlice(), name)) return v;
|
||||
}
|
||||
return null;
|
||||
}
|
||||
|
||||
fn findPlanesWithPrefix(
|
||||
allocator: std.mem.Allocator,
|
||||
space: *const xplane_proto.XSpace,
|
||||
prefix: []const u8,
|
||||
) ![]*const xplane_proto.XPlane {
|
||||
var res = std.ArrayList(*const xplane_proto.XPlane).init(allocator);
|
||||
for (space.planes.items) |*p| {
|
||||
if (std.mem.startsWith(u8, p.name.getSlice(), prefix)) try res.append(p);
|
||||
}
|
||||
return res.toOwnedSlice();
|
||||
}
|
||||
|
||||
fn buildDeviceAndResources(allocator: std.mem.Allocator, device_id: u32, plane: *const xplane_visitor.XPlaneVisitor, device: *Device) !void {
|
||||
device.name = plane.name();
|
||||
device.device_id = device_id;
|
||||
const sort_by_ordinal = (device_id == host_threads_device_id);
|
||||
var ordinal: u32 = 0;
|
||||
for (plane.plane.lines.items) |*xline| {
|
||||
const resource_id: u32 = @intCast(xLineDisplayId(xline));
|
||||
var resource: trace_events_proto.Resource = .{
|
||||
.resource_id = resource_id,
|
||||
.name = .{ .Const = xLineDisplayName(xline) },
|
||||
};
|
||||
|
||||
if (sort_by_ordinal) {
|
||||
ordinal += 1;
|
||||
resource.sort_index = ordinal;
|
||||
}
|
||||
try device.resources.put(allocator, resource_id, resource);
|
||||
}
|
||||
}
|
||||
|
||||
fn xplaneToTraceEvents(self: *TraceContainer, allocator: std.mem.Allocator, device_id: u32, xplane: *const xplane_visitor.XPlaneVisitor) !void {
|
||||
// Convert devices and resources.
|
||||
const device_entry = try self.devices.getOrPut(allocator, device_id);
|
||||
if (!device_entry.found_existing) device_entry.value_ptr.* = .{};
|
||||
|
||||
try buildDeviceAndResources(allocator, device_id, xplane, device_entry.value_ptr);
|
||||
|
||||
// Convert events.
|
||||
for (xplane.plane.lines.items) |*xline| {
|
||||
const resource_id: u32 = @intCast(xLineDisplayId(xline));
|
||||
|
||||
if (std.mem.eql(u8, xLineDisplayName(xline), xla_async_op_line_name)) continue;
|
||||
for (xline.events.items) |*xevent| {
|
||||
const event_type = xplane.getEventType(xevent.metadata_id);
|
||||
if (event_type.isInternalEvent()) continue;
|
||||
var event = try self.createEvent(allocator);
|
||||
event.device_id = device_id;
|
||||
event.resource_id = resource_id;
|
||||
|
||||
if (xplane.event_metadata_by_id.get(xevent.metadata_id)) |metadata| {
|
||||
if (metadata.display_name != .Empty) {
|
||||
event.name = metadata.display_name.getSlice();
|
||||
try event.args.put(allocator, "long_name", metadata.name.getSlice());
|
||||
} else {
|
||||
event.name = metadata.name.getSlice();
|
||||
}
|
||||
|
||||
event.timestamp_ps = (@as(u128, @intCast(xline.timestamp_ns)) * 1000) + @as(u128, @intCast(xevent.data.?.offset_ps));
|
||||
event.duration_ps = @intCast(xevent.duration_ps);
|
||||
|
||||
for (metadata.stats.items) |*xstat| {
|
||||
if (xstat.value == null) continue;
|
||||
var stat_buffer = std.ArrayList(u8).init(allocator);
|
||||
try xstatValueToString(xstat, xplane, stat_buffer.writer().any());
|
||||
const stat_str = try stat_buffer.toOwnedSlice();
|
||||
const stat_type = xplane.getStatType(xstat.metadata_id);
|
||||
if (stat_type.isInternalStat()) continue;
|
||||
if (stat_type == .step_name) event.name = stat_str;
|
||||
try event.args.put(allocator, xplane.getStatMetadataName(xstat.metadata_id), stat_str);
|
||||
}
|
||||
}
|
||||
|
||||
for (xevent.stats.items) |*xstat| {
|
||||
if (xstat.value == null) continue;
|
||||
var stat_buffer = std.ArrayList(u8).init(allocator);
|
||||
try xstatValueToString(xstat, xplane, stat_buffer.writer().any());
|
||||
const stat_str = try stat_buffer.toOwnedSlice();
|
||||
const stat_type = xplane.getStatType(xstat.metadata_id);
|
||||
if (stat_type.isInternalStat()) continue;
|
||||
if (stat_type == .step_name) event.name = stat_str;
|
||||
try event.args.put(allocator, xplane.getStatMetadataName(xstat.metadata_id), stat_str);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn fromXSpace(self: *TraceContainer, allocator: std.mem.Allocator, xspace: *const xplane_proto.XSpace, max_events: ?usize) !void {
|
||||
if (findPlaneWithName(xspace, host_threads_plane_name)) |hp| {
|
||||
const xplane = try xplane_visitor.XPlaneVisitor.init(allocator, hp);
|
||||
try self.xplaneToTraceEvents(allocator, host_threads_device_id, &xplane);
|
||||
}
|
||||
|
||||
var device_planes = try findPlanesWithPrefix(allocator, xspace, gpu_plane_prefix);
|
||||
|
||||
// We don't expect GPU and TPU planes and custom devices to be present in the
|
||||
// same XSpace.
|
||||
if (device_planes.len == 0) {
|
||||
device_planes = try findPlanesWithPrefix(allocator, xspace, tpu_plane_prefix);
|
||||
}
|
||||
if (device_planes.len == 0) {
|
||||
device_planes = try findPlanesWithPrefix(allocator, xspace, custom_plane_prefix);
|
||||
}
|
||||
|
||||
for (device_planes) |dp| {
|
||||
const xplane = try xplane_visitor.XPlaneVisitor.init(allocator, dp);
|
||||
const device_id: u32 = first_device_id + @as(u32, @intCast(xplane.plane.id));
|
||||
try self.xplaneToTraceEvents(allocator, device_id, &xplane);
|
||||
}
|
||||
|
||||
// Trace viewer (non-streaming) has scalability issues, we need to drop
|
||||
// events to avoid loading failure for trace viewer.
|
||||
if (max_events) |limit| self.capEvents(limit);
|
||||
}
|
||||
|
||||
pub fn createEvent(self: *TraceContainer, allocator: std.mem.Allocator) !*TraceEvent {
|
||||
try self.events.append(allocator, .{});
|
||||
return &self.events.items[self.events.items.len - 1];
|
||||
}
|
||||
|
||||
pub fn capEvents(self: *TraceContainer, max_count: u64) void {
|
||||
const total_count = self.events.items.len;
|
||||
if (total_count <= max_count) {
|
||||
// Nothing to do. Events are not known sorted after return.
|
||||
return;
|
||||
}
|
||||
// sort the events according to start time.
|
||||
// TODO: partial sort would improve performance.
|
||||
std.mem.sort(TraceEvent, self.events.items, {}, struct {
|
||||
pub fn call(_: void, lhs: TraceEvent, rhs: TraceEvent) bool {
|
||||
return lhs.timestamp_ps < rhs.timestamp_ps;
|
||||
}
|
||||
}.call);
|
||||
self.events.shrinkRetainingCapacity(max_count);
|
||||
}
|
||||
|
||||
pub fn toJson(self: *TraceContainer, writer: std.io.AnyWriter) !void {
|
||||
try writer.writeAll(
|
||||
\\{"displayTimeUnit":"ns","metadata":{"highres-ticks":true},"traceEvents":[
|
||||
);
|
||||
|
||||
self.devices.sort(struct {
|
||||
keys: []const u32,
|
||||
pub fn lessThan(ctx: @This(), lhs: usize, rhs: usize) bool {
|
||||
return ctx.keys[lhs] < ctx.keys[rhs];
|
||||
}
|
||||
}{ .keys = self.devices.keys() });
|
||||
|
||||
for (self.devices.keys(), self.devices.values()) |device_id, *device| {
|
||||
if (device.name.len != 0) {
|
||||
try writer.print(
|
||||
\\{{"ph":"M","pid":{d},"name":"process_name","args":{{"name":"{s}"}}}},
|
||||
, .{ device_id, device.name });
|
||||
}
|
||||
try writer.print(
|
||||
\\{{"ph":"M","pid":{d},"name":"process_sort_index","args":{{"sort_index":{d}}}}},
|
||||
, .{
|
||||
device_id,
|
||||
device_id,
|
||||
});
|
||||
|
||||
device.resources.sort(struct {
|
||||
keys: []const u32,
|
||||
pub fn lessThan(ctx: @This(), lhs: usize, rhs: usize) bool {
|
||||
return ctx.keys[lhs] < ctx.keys[rhs];
|
||||
}
|
||||
}{ .keys = device.resources.keys() });
|
||||
|
||||
for (device.resources.keys(), device.resources.values()) |resource_id, resource| {
|
||||
if (resource.name.getSlice().len != 0) {
|
||||
try writer.print(
|
||||
\\{{"ph":"M","pid":{d},"tid":{d},"name":"thread_name","args":{{"name":"{s}"}}}},
|
||||
, .{
|
||||
device_id,
|
||||
resource_id,
|
||||
resource.name.getSlice(),
|
||||
});
|
||||
}
|
||||
const sort_index = if (resource.sort_index != 0) resource.sort_index else resource_id;
|
||||
try writer.print(
|
||||
\\{{"ph":"M","pid":{d},"tid":{d},"name":"thread_sort_index","args":{{"sort_index":{d}}}}},
|
||||
, .{ device_id, resource_id, sort_index });
|
||||
}
|
||||
}
|
||||
|
||||
for (self.events.items) |*event| {
|
||||
const duration_ps = @max(event.duration_ps, 1);
|
||||
try writer.print(
|
||||
\\{{"ph":"X","pid":{d},"tid":{d},"ts":{d:.17},"dur":{d:.17},"name":"{s}"
|
||||
, .{
|
||||
event.device_id,
|
||||
event.resource_id,
|
||||
picoToMicro(event.timestamp_ps),
|
||||
picoToMicro(duration_ps),
|
||||
event.name,
|
||||
});
|
||||
if (event.args.count() != 0) {
|
||||
try writer.writeAll(
|
||||
\\,"args":{
|
||||
);
|
||||
event.args.sort(struct {
|
||||
keys: []const []const u8,
|
||||
|
||||
pub fn lessThan(ctx: @This(), lhs: usize, rhs: usize) bool {
|
||||
return std.mem.order(u8, ctx.keys[lhs], ctx.keys[rhs]).compare(std.math.CompareOperator.lt);
|
||||
}
|
||||
}{ .keys = event.args.keys() });
|
||||
|
||||
for (event.args.keys(), event.args.values(), 0..) |key, value, i| {
|
||||
if (i < event.args.count() - 1) {
|
||||
try writer.print(
|
||||
\\"{s}":"{s}",
|
||||
, .{ key, value });
|
||||
} else {
|
||||
// Last item has closing bracket rather than trailing comma.
|
||||
try writer.print(
|
||||
\\"{s}":"{s}"}}
|
||||
, .{ key, value });
|
||||
}
|
||||
}
|
||||
}
|
||||
try writer.writeAll("},");
|
||||
}
|
||||
try writer.writeAll("{}]}");
|
||||
}
|
||||
};
|
||||
297
pjrt/convert/xplane_schema.zig
Normal file
297
pjrt/convert/xplane_schema.zig
Normal file
@ -0,0 +1,297 @@
|
||||
const std = @import("std");
|
||||
|
||||
// `HostEventType` uses the unconventional casing/formatting
|
||||
// so that the string representation of the enum used in the
|
||||
// protobuf encoding directly maps to the zig enum tag name.
|
||||
pub const HostEventType = enum(u16) {
|
||||
unknown = 0,
|
||||
TraceContext,
|
||||
SessionRun,
|
||||
FunctionRun,
|
||||
RunGraph,
|
||||
RunGraphDone,
|
||||
TfOpRun,
|
||||
EagerExecute,
|
||||
@"ExecutorState::Process",
|
||||
ExecutorDoneCallback,
|
||||
MemoryAllocation,
|
||||
MemoryDeallocation,
|
||||
// Performance counter related.
|
||||
RemotePerfCounter,
|
||||
// tf.data captured function events.
|
||||
@"InstantiatedCapturedFunction::Run",
|
||||
@"InstantiatedCapturedFunction::RunWithBorrowedArgs",
|
||||
@"InstantiatedCapturedFunction::RunInstantiated",
|
||||
@"InstantiatedCapturedFunction::RunAsync",
|
||||
// Loop ops.
|
||||
ParallelForOp,
|
||||
ForeverOp,
|
||||
@"WhileOp-EvalCond",
|
||||
@"WhileOp-StartBody",
|
||||
ForOp,
|
||||
// tf.data related.
|
||||
@"IteratorGetNextOp::DoCompute",
|
||||
@"IteratorGetNextAsOptionalOp::DoCompute",
|
||||
Iterator,
|
||||
@"Iterator::Prefetch::Generator",
|
||||
PrefetchProduce,
|
||||
PrefetchConsume,
|
||||
ParallelInterleaveProduce,
|
||||
ParallelInterleaveConsume,
|
||||
ParallelInterleaveInitializeInput,
|
||||
ParallelMapProduce,
|
||||
ParallelMapConsume,
|
||||
MapAndBatchProduce,
|
||||
MapAndBatchConsume,
|
||||
ParseExampleProduce,
|
||||
ParseExampleConsume,
|
||||
ParallelBatchProduce,
|
||||
ParallelBatchConsume,
|
||||
// Batching related.
|
||||
BatchingSessionRun,
|
||||
ProcessBatch,
|
||||
BrainSessionRun,
|
||||
ConcatInputTensors,
|
||||
MergeInputTensors,
|
||||
ScheduleWithoutSplit,
|
||||
ScheduleWithSplit,
|
||||
ScheduleWithEagerSplit,
|
||||
@"ASBSQueue::Schedule",
|
||||
// TFRT related.
|
||||
TfrtModelRun,
|
||||
// Serving related.
|
||||
ServingModelRun,
|
||||
// GPU related.
|
||||
KernelLaunch,
|
||||
KernelExecute,
|
||||
// TPU related
|
||||
EnqueueRequestLocked,
|
||||
RunProgramRequest,
|
||||
HostCallbackRequest,
|
||||
TransferH2DRequest,
|
||||
TransferPreprocessedH2DRequest,
|
||||
TransferD2HRequest,
|
||||
OnDeviceSendRequest,
|
||||
OnDeviceRecvRequest,
|
||||
OnDeviceSendRecvLocalRequest,
|
||||
CustomWait,
|
||||
OnDeviceSendRequestMulti,
|
||||
OnDeviceRecvRequestMulti,
|
||||
PjrtAsyncWait,
|
||||
DoEnqueueProgram,
|
||||
DoEnqueueContinuationProgram,
|
||||
WriteHbm,
|
||||
ReadHbm,
|
||||
TpuExecuteOp,
|
||||
CompleteCallbacks,
|
||||
@"tpu::System::TransferToDevice=>IssueEvent",
|
||||
@"tpu::System::TransferToDevice=>IssueEvent=>Done",
|
||||
@"tpu::System::TransferFromDevice=>IssueEvent",
|
||||
@"tpu::System::TransferFromDevice=>IssueEvent=>Done",
|
||||
@"tpu::System::Execute",
|
||||
@"TPUPartitionedCallOp-InitializeVarOnTPU",
|
||||
@"TPUPartitionedCallOp-ExecuteRemote",
|
||||
@"TPUPartitionedCallOp-ExecuteLocal",
|
||||
Linearize,
|
||||
Delinearize,
|
||||
@"TransferBufferFromDevice-FastPath",
|
||||
|
||||
pub fn fromString(event_name: []const u8) HostEventType {
|
||||
return std.meta.stringToEnum(HostEventType, event_name) orelse .unknown;
|
||||
}
|
||||
|
||||
pub fn isInternalEvent(event_type: HostEventType) bool {
|
||||
// TODO(b/162102421): Introduce a prefix for internal event names.
|
||||
return switch (event_type) {
|
||||
.MemoryAllocation,
|
||||
.MemoryDeallocation,
|
||||
.PrefetchProduce,
|
||||
.PrefetchConsume,
|
||||
.ParallelInterleaveProduce,
|
||||
.ParallelInterleaveConsume,
|
||||
.ParallelInterleaveInitializeInput,
|
||||
.ParallelMapProduce,
|
||||
.ParallelMapConsume,
|
||||
.MapAndBatchProduce,
|
||||
.MapAndBatchConsume,
|
||||
.ParseExampleProduce,
|
||||
.ParseExampleConsume,
|
||||
=> true,
|
||||
else => false,
|
||||
};
|
||||
}
|
||||
};
|
||||
|
||||
// `StatType` uses the unconventional casing/formatting
|
||||
// so that the string representation of the enum used in the
|
||||
// protobuf encoding directly maps to the zig enum tag name.
|
||||
pub const StatType = enum(u16) {
|
||||
unknown = 0,
|
||||
// TraceMe arguments.
|
||||
id,
|
||||
device_ordinal,
|
||||
chip_ordinal,
|
||||
node_ordinal,
|
||||
model_id,
|
||||
queue_id,
|
||||
queue_addr,
|
||||
request_id,
|
||||
run_id,
|
||||
replica_id,
|
||||
graph_type,
|
||||
step_num,
|
||||
iter_num,
|
||||
index_on_host,
|
||||
allocator_name,
|
||||
bytes_reserved,
|
||||
bytes_allocated,
|
||||
bytes_available,
|
||||
fragmentation,
|
||||
peak_bytes_in_use,
|
||||
requested_bytes,
|
||||
allocation_bytes,
|
||||
addr,
|
||||
region_type,
|
||||
data_type,
|
||||
shape,
|
||||
layout,
|
||||
kpi_name,
|
||||
kpi_value,
|
||||
element_id,
|
||||
parent_id,
|
||||
core_type,
|
||||
// XPlane semantics related.
|
||||
_pt,
|
||||
_ct,
|
||||
_p,
|
||||
_c,
|
||||
_r,
|
||||
_a,
|
||||
// Device trace arguments.
|
||||
device_id,
|
||||
device_type_string,
|
||||
context_id,
|
||||
correlation_id,
|
||||
// TODO(b/176137043): These "details" should differentiate between activity
|
||||
// and API event sources.
|
||||
memcpy_details,
|
||||
memalloc_details,
|
||||
MemFree_details,
|
||||
Memset_details,
|
||||
MemoryResidency_details,
|
||||
nvtx_range,
|
||||
kernel_details,
|
||||
stream,
|
||||
// Stats added when processing traces.
|
||||
group_id,
|
||||
flow,
|
||||
step_name,
|
||||
tf_op,
|
||||
hlo_op,
|
||||
deduplicated_name,
|
||||
hlo_category,
|
||||
hlo_module,
|
||||
program_id,
|
||||
equation,
|
||||
is_eager,
|
||||
is_func,
|
||||
tf_function_call,
|
||||
tracing_count,
|
||||
flops,
|
||||
model_flops,
|
||||
bytes_accessed,
|
||||
memory_access_breakdown,
|
||||
source,
|
||||
model_name,
|
||||
model_version,
|
||||
bytes_transferred,
|
||||
queue,
|
||||
dcn_collective_info,
|
||||
// Performance counter related.
|
||||
@"Raw Value",
|
||||
@"Scaled Value",
|
||||
@"Thread Id",
|
||||
matrix_unit_utilization_percent,
|
||||
// XLA metadata map related.
|
||||
@"Hlo Proto",
|
||||
// Device capability related.
|
||||
clock_rate,
|
||||
// For GPU, this is the number of SMs.
|
||||
core_count,
|
||||
memory_bandwidth,
|
||||
memory_size,
|
||||
compute_cap_major,
|
||||
compute_cap_minor,
|
||||
peak_teraflops_per_second,
|
||||
peak_hbm_bw_gigabytes_per_second,
|
||||
peak_sram_rd_bw_gigabytes_per_second,
|
||||
peak_sram_wr_bw_gigabytes_per_second,
|
||||
device_vendor,
|
||||
// Batching related.
|
||||
batch_size_after_padding,
|
||||
padding_amount,
|
||||
batching_input_task_size,
|
||||
// GPU occupancy metrics
|
||||
theoretical_occupancy_pct,
|
||||
occupancy_min_grid_size,
|
||||
occupancy_suggested_block_size,
|
||||
// Aggregated Stats
|
||||
self_duration_ps,
|
||||
min_duration_ps,
|
||||
total_profile_duration_ps,
|
||||
max_iteration_num,
|
||||
device_type,
|
||||
uses_megacore,
|
||||
symbol_id,
|
||||
tf_op_name,
|
||||
dma_stall_duration_ps,
|
||||
key,
|
||||
payload_size_bytes,
|
||||
duration_us,
|
||||
buffer_size,
|
||||
transfers,
|
||||
// Dcn message Stats
|
||||
dcn_label,
|
||||
dcn_source_slice_id,
|
||||
dcn_source_per_slice_device_id,
|
||||
dcn_destination_slice_id,
|
||||
dcn_destination_per_slice_device_id,
|
||||
dcn_chunk,
|
||||
dcn_loop_index,
|
||||
@"EdgeTPU Model information",
|
||||
@"EdgeTPU Model Profile information",
|
||||
@"EdgeTPU MLIR",
|
||||
dropped_traces,
|
||||
cuda_graph_id,
|
||||
// Many events have `.cuda_graph_id`, such as graph sub events when tracing is in
|
||||
// node level. Yet `.cuda_graph_exec_id` is used only for CudaGraphExecution events
|
||||
// on the GPU device when tracing is in graph level.
|
||||
cuda_graph_exec_id,
|
||||
cuda_graph_orig_id,
|
||||
step_idle_time_ps,
|
||||
gpu_device_name,
|
||||
source_stack,
|
||||
device_offset_ps,
|
||||
device_duration_ps,
|
||||
|
||||
pub fn fromString(stat_name: []const u8) StatType {
|
||||
return std.meta.stringToEnum(StatType, stat_name) orelse .unknown;
|
||||
}
|
||||
|
||||
pub fn isInternalStat(stat_type: StatType) bool {
|
||||
return switch (stat_type) {
|
||||
.kernel_details,
|
||||
._pt,
|
||||
._p,
|
||||
._ct,
|
||||
._c,
|
||||
._r,
|
||||
.flops,
|
||||
.bytes_accessed,
|
||||
.program_id,
|
||||
.symbol_id,
|
||||
=> true,
|
||||
else => false,
|
||||
};
|
||||
}
|
||||
};
|
||||
50
pjrt/convert/xplane_visitor.zig
Normal file
50
pjrt/convert/xplane_visitor.zig
Normal file
@ -0,0 +1,50 @@
|
||||
const std = @import("std");
|
||||
const xplane_proto = @import("//tsl:xplane_proto");
|
||||
const xplane_schema = @import("xplane_schema.zig");
|
||||
|
||||
pub const XPlaneVisitor = struct {
|
||||
plane: *const xplane_proto.XPlane,
|
||||
event_metadata_by_id: std.AutoHashMapUnmanaged(i64, *const xplane_proto.XEventMetadata) = .{},
|
||||
stat_metadata_by_id: std.AutoHashMapUnmanaged(i64, *const xplane_proto.XStatMetadata) = .{},
|
||||
|
||||
pub fn init(
|
||||
allocator: std.mem.Allocator,
|
||||
plane: *const xplane_proto.XPlane,
|
||||
) !XPlaneVisitor {
|
||||
var res: XPlaneVisitor = .{ .plane = plane };
|
||||
|
||||
// build event metadata map
|
||||
for (plane.event_metadata.items) |*event_metadata| {
|
||||
try res.event_metadata_by_id.put(allocator, event_metadata.key, &event_metadata.value.?);
|
||||
}
|
||||
|
||||
// build stat metadata map
|
||||
for (plane.stat_metadata.items) |*stat_metadata| {
|
||||
try res.stat_metadata_by_id.put(allocator, stat_metadata.key, &stat_metadata.value.?);
|
||||
}
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
pub fn getEventType(self: *const XPlaneVisitor, event_metadata_id: i64) xplane_schema.HostEventType {
|
||||
if (self.event_metadata_by_id.get(event_metadata_id)) |v| {
|
||||
return xplane_schema.HostEventType.fromString(v.name.getSlice());
|
||||
} else return .unknown;
|
||||
}
|
||||
|
||||
pub fn name(self: *const XPlaneVisitor) []const u8 {
|
||||
return self.plane.name.getSlice();
|
||||
}
|
||||
|
||||
pub fn getStatMetadataName(self: *const XPlaneVisitor, stat_metadata_id: i64) []const u8 {
|
||||
if (self.stat_metadata_by_id.get(stat_metadata_id)) |v| {
|
||||
return v.name.getSlice();
|
||||
} else return &[_]u8{};
|
||||
}
|
||||
|
||||
pub fn getStatType(self: *const XPlaneVisitor, stat_metadata_id: i64) xplane_schema.StatType {
|
||||
if (self.stat_metadata_by_id.get(stat_metadata_id)) |v| {
|
||||
return xplane_schema.StatType.fromString(v.name.getSlice());
|
||||
} else return .unknown;
|
||||
}
|
||||
};
|
||||
45
pjrt/xspace_to_json.zig
Normal file
45
pjrt/xspace_to_json.zig
Normal file
@ -0,0 +1,45 @@
|
||||
const flags = @import("tigerbeetle/flags");
|
||||
const std = @import("std");
|
||||
|
||||
const TraceContainer = @import("convert/trace_container.zig").TraceContainer;
|
||||
|
||||
const CliArgs = struct {
|
||||
pub const help =
|
||||
\\ llama --path=path_to_profiling_data
|
||||
;
|
||||
path: []const u8,
|
||||
max_events: ?usize = null,
|
||||
};
|
||||
|
||||
pub fn main() !void {
|
||||
var gpa = std.heap.GeneralPurposeAllocator(.{ .thread_safe = true }){};
|
||||
defer _ = gpa.deinit();
|
||||
const allocator = gpa.allocator();
|
||||
|
||||
var args = std.process.args();
|
||||
const cli_args = flags.parse(&args, CliArgs);
|
||||
|
||||
var fd = try std.fs.openFileAbsolute(cli_args.path, .{});
|
||||
defer fd.close();
|
||||
|
||||
const pb_buffer = try fd.readToEndAlloc(allocator, (try fd.stat()).size);
|
||||
defer allocator.free(pb_buffer);
|
||||
if (pb_buffer.len == 0) return error.EmptyBuffer;
|
||||
|
||||
var converter = try TraceContainer.init(allocator, pb_buffer, cli_args.max_events);
|
||||
defer converter.deinit();
|
||||
|
||||
var path_buffer: [1028]u8 = undefined;
|
||||
|
||||
const output_path = try std.fmt.bufPrint(&path_buffer, "{s}/{s}.json", .{
|
||||
std.fs.path.dirname(cli_args.path) orelse "",
|
||||
std.fs.path.stem(cli_args.path),
|
||||
});
|
||||
|
||||
var output_file = try std.fs.createFileAbsolute(output_path, .{});
|
||||
defer output_file.close();
|
||||
|
||||
try converter.toJson(output_file.writer().any());
|
||||
|
||||
std.debug.print("Wrote JSON to {s}\n", .{output_path});
|
||||
}
|
||||
Loading…
Reference in New Issue
Block a user