Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
11 changes: 11 additions & 0 deletions clean/FORMATS.md
Original file line number Diff line number Diff line change
Expand Up @@ -467,6 +467,17 @@ see "Context teardown and re-creation"). Other keys go to the network runtime's
`k_input=16|f32` (model K's input pass, below); the create flag `WG_CREATE_HDR` reaches it as
`WgRtConfig.hdr`; timing log parts then named exposure, input, network, output).

The D3D12 executor makes the network's allocations twice: once to measure the arena they need, in the default arena
(`WG_D3D12_ARENA_MB`, 2560 MiB), then in an arena of exactly that size (executor_d3d12.cpp `prepare_device`). A network
that does not fit the default arena is measured again in a 4095 MiB arena, the largest (logged `the network does not fit
the default arena (...); measuring again in a 4095 MiB arena`, then `measured in the 4095 MiB arena: the network needs N
MiB; the D3D12 executor runs it in an arena of that size` or `measuring again in a 4095 MiB arena failed (...)`); not
when `WG_D3D12_ARENA_MB` is set. Model L at 3840x2160 native needs 2738 MiB and runs on the D3D12 executor this way
(clean/l/records/issue182_create_destroy_record.json; issue #182). If the second measure fails too, the context goes to
the HIP executor (`executor dx12 (default on RDNA 4) could not take this context (...); the HIP executor runs it`), and
after it is destroyed the process keeps HIP's instance and the pooled imported arena until it ends, as "Context
teardown and re-creation" describes (the same record's control: +1419 MiB after every destroy, constant).

### Shared resources (created at create on the game's device, imported into HIP)

Two slots (used alternately by the frames whose network is enqueued). All ten buffers (twelve with `motion_hi`) live in ONE committed buffer,
Expand Down
208 changes: 208 additions & 0 deletions clean/l/records/issue182_create_destroy_record.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,208 @@
{
"what": "Issue #182: process GPU memory after an upscaler is created and destroyed (5 cycles, no resize), Model L and Model M at 3840x2160 output, native and Quality, with the D3D12 executor's arena re-measure",
"harness": "clean/fsr/tests/test_dll_d3d12.cpp resize_real (WG_TEST_RESIZE_STEPS=0, WG_TEST_RESIZE_CYCLES=5, settle 800 ms): DXGI QueryVideoMemoryInfo, local segment, this process only; the runtime-reported figure is the front end's FFX_API_QUERY_DESC_TYPE_UPSCALE_GPU_MEMORY_USAGE",
"setup": {
"gpu": "AMD Radeon RX 9070 XT (gfx1201), 16191 MiB dedicated",
"base": "this branch merged with main (merge cabcd4c)",
"runtime": "this branch, clean/runner/build.py --model-l --no-model-k, D3D12 kernels build_dx12.py --model-l (69 kernels)",
"front_end": "clean/fsr with WG_MODEL_L=ON, WG_MODEL_K=OFF (test_dll_d3d12 and its front-end copy)",
"runtime_dll_sha256": "9204b29fa4ba92111235a61473aa0b2f412195aa739a914c957b607c2296ddb0",
"front_end_dll_sha256": "b89ceea05b3dd1de875f366266a0bf88bddc7bf3d307bc24f7b06b252529eb55",
"models": "synthetic files from the repository's generators (Model M: synthetic_model.synthetic_tensors(3) with its Swin weights scaled so every stage takes the clamp-free kernels, as a real model's do; Model L: l_frame_check's synthetic model, seed 89); memory depends on the network's shapes, not on the values",
"gpu_runs": "one at a time, each through tools/gpu_slot.py --cap-s 170"
},
"runs": [
{
"run": "l4k-native",
"model": "L",
"output": "3840x2160",
"mode": "native (render 3840x2160)",
"what": "fix",
"cycles": 5,
"process_local_mib_before_any_context": 204,
"process_growth_mib_per_cycle_peak": [
2742,
2742,
2742,
2742,
2742
],
"runtime_reported_mib": [
2739
],
"growth_mib_after_each_destroy": [
2,
2,
2,
2,
2
],
"executor_per_cycle": [
"dx12",
"dx12",
"dx12",
"dx12",
"dx12"
],
"remeasured_in_4095_mib_arena": 5,
"d3d12_arena_need_mib": [
2738
],
"hip_fallback_lines": 0,
"pooled_arena_kept_lines": 0,
"test_passed": true
},
{
"run": "l4k-q",
"model": "L",
"output": "3840x2160",
"mode": "Quality (render 2560x1440)",
"what": "fix",
"cycles": 5,
"process_local_mib_before_any_context": 135,
"process_growth_mib_per_cycle_peak": [
1444,
1444,
1444,
1444,
1444
],
"runtime_reported_mib": [
1441
],
"growth_mib_after_each_destroy": [
2,
2,
2,
2,
2
],
"executor_per_cycle": [
"dx12",
"dx12",
"dx12",
"dx12",
"dx12"
],
"remeasured_in_4095_mib_arena": 0,
"d3d12_arena_need_mib": [],
"hip_fallback_lines": 0,
"pooled_arena_kept_lines": 0,
"test_passed": true
},
{
"run": "m4k-native",
"model": "M",
"output": "3840x2160",
"mode": "native (render 3840x2160)",
"what": "fix",
"cycles": 5,
"process_local_mib_before_any_context": 204,
"process_growth_mib_per_cycle_peak": [
1593,
1593,
1593,
1593,
1593
],
"runtime_reported_mib": [
1590
],
"growth_mib_after_each_destroy": [
2,
2,
2,
2,
2
],
"executor_per_cycle": [
"dx12",
"dx12",
"dx12",
"dx12",
"dx12"
],
"remeasured_in_4095_mib_arena": 0,
"d3d12_arena_need_mib": [],
"hip_fallback_lines": 0,
"pooled_arena_kept_lines": 0,
"test_passed": true
},
{
"run": "m4k-q",
"model": "M",
"output": "3840x2160",
"mode": "Quality (render 2560x1440)",
"what": "fix",
"cycles": 5,
"process_local_mib_before_any_context": 135,
"process_growth_mib_per_cycle_peak": [
936,
936,
936,
936,
936
],
"runtime_reported_mib": [
933
],
"growth_mib_after_each_destroy": [
2,
2,
2,
2,
2
],
"executor_per_cycle": [
"dx12",
"dx12",
"dx12",
"dx12",
"dx12"
],
"remeasured_in_4095_mib_arena": 0,
"d3d12_arena_need_mib": [],
"hip_fallback_lines": 0,
"pooled_arena_kept_lines": 0,
"test_passed": true
},
{
"run": "l4k-native-arena2560",
"model": "L",
"output": "3840x2160",
"mode": "native (render 3840x2160)",
"what": "control: same build, WG_D3D12_ARENA_MB=2560 (no re-measure: the behaviour before this change)",
"cycles": 5,
"process_local_mib_before_any_context": 204,
"process_growth_mib_per_cycle_peak": [
4230,
4230,
4230,
4230,
4230
],
"runtime_reported_mib": [
3433
],
"growth_mib_after_each_destroy": [
1419,
1419,
1419,
1419,
1419
],
"executor_per_cycle": [
"hip",
"hip",
"hip",
"hip",
"hip"
],
"remeasured_in_4095_mib_arena": 0,
"d3d12_arena_need_mib": [],
"hip_fallback_lines": 5,
"pooled_arena_kept_lines": 5,
"test_passed": true
}
]
}
42 changes: 38 additions & 4 deletions clean/runtime/interop/executor_d3d12.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@
// into an RG32F texture of ours, which the network then reads as its render-size motion texture.
#include <algorithm>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <deque>
#include <string>
Expand All @@ -39,6 +40,10 @@ namespace wgi {

namespace {

// The largest native D3D12 arena (WG_D3D12_ARENA_MB's upper bound, 4095 MiB): the arena that measures a network
// larger than the default one.
constexpr uint64_t kMeasureMaxBytes = uint64_t(4095) << 20;

template <class T>
void release_ref(T*& p) {
if (p) p->Release();
Expand Down Expand Up @@ -94,15 +99,44 @@ class D3d12Executor : public Executor {
const std::string build = cfg_.build_dir ? cfg_.build_dir : ".";
// Pass 1 measures the arena the network needs (the default arena is large); pass 2 makes the same allocations in
// an arena of exactly that size. The allocation sequence of runtime_create is deterministic.
// A network larger than the default arena (Model L at 4K native: about 3.4 GiB) is measured again in the largest
// arena (kMeasureMaxBytes, WG_D3D12_ARENA_MB's upper bound) unless WG_D3D12_ARENA_MB is set. Without this the
// context went to the HIP executor, whose process keeps HIP's instance and the pooled imported arena after the
// context is destroyed (clean/FORMATS.md "Context teardown and re-creation"; issue #182).
uint64_t need = 0;
{
const char* arena_env = std::getenv("WG_D3D12_ARENA_MB");
const bool may_grow = !arena_env || !*arena_env;
std::string e;
std::unique_ptr<wgrt::Device> dev = wgrt::make_d3d12_device_on(device, queue_, build, &e);
if (!dev) return *err = "D3D12 device: " + e, false;
wgrt::Runtime* rt = wgrt::runtime_create(c, std::move(dev), &e, true, true);
if (!rt) return *err = "D3D12 runtime: " + e, false;
wgrt::Runtime* rt = nullptr;
bool grown = false;
const std::string max_mib = std::to_string(kMeasureMaxBytes >> 20);
for (uint64_t arena : {uint64_t(0), kMeasureMaxBytes}) {
e.clear();
std::unique_ptr<wgrt::Device> dev = wgrt::make_d3d12_device_on(device, queue_, build, &e, arena);
if (!dev) {
if (grown && log_)
log_(true, "measuring again in a " + max_mib + " MiB arena failed (" + e + ")");
return *err = "D3D12 device: " + e, false;
}
rt = wgrt::runtime_create(c, std::move(dev), &e, true, true);
if (rt || !may_grow || arena || e.find("native D3D12 arena exhausted") == std::string::npos) break;
grown = true;
if (log_)
log_(false, "the network does not fit the default arena (" + e + "); measuring again in a " +
max_mib + " MiB arena");
}
if (!rt) {
if (grown && log_)
log_(true, "measuring again in a " + max_mib + " MiB arena failed (" + e + ")");
return *err = "D3D12 runtime: " + e, false;
}
need = wgrt::d3d12_ext(wgrt::runtime_device(rt))->arena_high_water();
wgrt::runtime_destroy(rt);
if (grown && log_)
log_(false, "measured in the " + max_mib + " MiB arena: the network needs " +
std::to_string((need + (1u << 20) - 1) >> 20) + " MiB; the D3D12 executor runs it in an arena "
"of that size");
}
std::string e;
std::unique_ptr<wgrt::Device> dev = wgrt::make_d3d12_device_on(device, queue_, build, &e, need + (1u << 20));
Expand Down
Loading