AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SWDEV-363074 - Adjust staging copy limits in Windows

Browse Source 
Pinned copy can cause big performance drops, because slow pinning under Windows.
Use up to 128MB for staging transfers. Change staging buffer size to 4MB.
Linux path should still have the old defaults.

Change-Id: I954edceb3ec89e8e670be116aa2d0a9564c8b11c
This commit is contained in:
German
2022-11-17 14:44:24 -05:00
parent 0d9e1de4e3
commit 79d12df147
3 changed files with 13 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/device/pal/palsettings.cpp
+3 -8
View File
@@ -54,7 +54,7 @@ Settings::Settings() {
stagedXferRead_ = true;
stagedXferWrite_ = true;
stagedXferSize_ = GPU_STAGING_BUFFER_SIZE * Ki;
stagedXferSize_ = GPU_STAGING_BUFFER_SIZE * Mi;
// We will enable staged read/write if we use local memory
disablePersistent_ = false;
@@ -78,14 +78,9 @@ Settings::Settings() {
// By default use host blit
blitEngine_ = BlitEngineHost;
constexpr size_t MaxPinnedXferSize = 64;
pinnedXferSize_ = std::min(GPU_PINNED_XFER_SIZE, MaxPinnedXferSize) * Mi;
constexpr size_t PinnedMinXferSize = 4 * Mi;
pinnedXferSize_ = GPU_PINNED_MIN_XFER_SIZE * Mi;
pinnedMinXferSize_ = flagIsDefault(GPU_PINNED_MIN_XFER_SIZE)
? PinnedMinXferSize
: GPU_PINNED_MIN_XFER_SIZE * Ki;
pinnedMinXferSize_ = std::min(pinnedMinXferSize_, pinnedXferSize_);
? 128 * Mi : GPU_PINNED_MIN_XFER_SIZE * Mi;
// Disable FP_FAST_FMA defines by default
reportFMAF_ = false;
rocclr/device/rocm/rocsettings.cpp
+5 -4
View File
@@ -62,14 +62,15 @@ Settings::Settings() {
stagedXferRead_ = true;
stagedXferWrite_ = true;
stagedXferSize_ = GPU_STAGING_BUFFER_SIZE * Ki;
stagedXferSize_ = flagIsDefault(GPU_STAGING_BUFFER_SIZE)
? 1 * Mi : GPU_STAGING_BUFFER_SIZE * Mi;
// Initialize transfer buffer size to 1MB by default
xferBufSize_ = 1024 * Ki;
const static size_t MaxPinnedXferSize = 128;
pinnedXferSize_ = std::min(GPU_PINNED_XFER_SIZE, MaxPinnedXferSize) * Mi;
pinnedMinXferSize_ = std::min(GPU_PINNED_MIN_XFER_SIZE * Ki, pinnedXferSize_);
pinnedXferSize_ = GPU_PINNED_MIN_XFER_SIZE * Mi;
pinnedMinXferSize_ = flagIsDefault(GPU_PINNED_MIN_XFER_SIZE)
? 1 * Mi : GPU_PINNED_MIN_XFER_SIZE * Mi;
sdmaCopyThreshold_ = GPU_FORCE_BLIT_COPY_SIZE * Ki;