Skip to content

feat: reduce gpu sync size. #246

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
merged 1 commit into from
Aug 12, 2025
Merged

feat: reduce gpu sync size. #246

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
68 changes: 49 additions & 19 deletions src/nn/nn-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,13 +1,14 @@
#include "nn-vulkan.hpp"

#define DEBUG_VULKAN_BUFFERS false
#define DEBUG_VULKAN_TRACE false

#if DEBUG_VULKAN_TRACE
#define VULKAN_TRACE(...) printf("VULKAN_TRACE: "); printf(__VA_ARGS__); printf("\n");
#else
#define VULKAN_TRACE(...)
#endif

#define DEBUG_VULKAN_BUFFERS false

static bool hasPortabilityExtension() {
#ifdef __APPLE__
const std::vector<vk::ExtensionProperties> extensionProperties = vk::enumerateInstanceExtensionProperties();
Expand Down Expand Up @@ -122,7 +123,7 @@ void NnVulkanStagingCopy::executeCopyCommand() {
context->device.freeCommandBuffers(context->commandPool, 1, &commandBuffer);
}

NnVulkanBuffer::NnVulkanBuffer(NnVulkanContext *context, const char *name, const vk::DeviceSize bufferSize, vk::BufferUsageFlags usageFlags, bool fastAccess) {
NnVulkanBuffer::NnVulkanBuffer(NnVulkanContext *context, const char *name, const NnSize bufferSize, vk::BufferUsageFlags usageFlags, bool fastAccess) {
this->context = context;
this->name = name;
this->bufferSize = bufferSize;
Expand All @@ -131,7 +132,7 @@ NnVulkanBuffer::NnVulkanBuffer(NnVulkanContext *context, const char *name, const

isHostVisible = false;

VULKAN_TRACE("Creating buffer of size %zu (fastAccess=%d)", (NnSize)bufferSize, fastAccess);
VULKAN_TRACE("Creating buffer of size %zu (fastAccess=%d)", bufferSize, fastAccess);

uint32_t memoryTypeIndex = MEMORY_TYPE_INDEX_NOT_FOUND;
if (fastAccess) {
Expand Down Expand Up @@ -164,43 +165,67 @@ NnVulkanBuffer::NnVulkanBuffer(NnVulkanContext *context, const char *name, const
deviceMemory = b.second;
if (isHostVisible)
hostPointer = context->device.mapMemory(deviceMemory, 0, bufferSize);
VULKAN_TRACE("Created buffer of size %zu (fastAccess=%d)", (NnSize)bufferSize, fastAccess);
VULKAN_TRACE("Created buffer of size %zu (fastAccess=%d)", bufferSize, fastAccess);
}

NnVulkanBuffer::~NnVulkanBuffer() {
if (hostPointer != nullptr)
context->device.unmapMemory(deviceMemory);
context->device.freeMemory(deviceMemory);
context->device.destroyBuffer(deviceBuffer);
VULKAN_TRACE("Destroyed buffer of size %zu", (NnSize)bufferSize);
VULKAN_TRACE("Destroyed buffer of size %zu", bufferSize);
}

void NnVulkanBuffer::write(const NnByte *data) {
write(data, bufferSize);
}

void NnVulkanBuffer::write(const NnByte *data, const NnSize size) {
assert(size <= bufferSize);

if (isHostVisible && hostPointer != nullptr) {
std::memcpy(hostPointer, data, bufferSize);
context->device.flushMappedMemoryRanges({ { deviceMemory, 0, bufferSize } });
VULKAN_TRACE("Wrote %zu bytes to host visible buffer", (NnSize)bufferSize);
std::memcpy(hostPointer, data, size);
context->device.flushMappedMemoryRanges({ { deviceMemory, 0, (vk::DeviceSize)size } });
VULKAN_TRACE("Wrote %zu bytes to host visible buffer", size);
} else {
NnVulkanStagingCopy copy(context, deviceBuffer, bufferSize, COPY_TO_DEVICE);
NnVulkanStagingCopy copy(context, deviceBuffer, size, COPY_TO_DEVICE);
copy.copy((NnByte *)data);
copy.executeCopyCommand();
VULKAN_TRACE("Wrote %zu bytes to buffer", (NnSize)bufferSize);
VULKAN_TRACE("Wrote %zu bytes to buffer", size);
}
}

void NnVulkanBuffer::read(NnByte *data) {
read(data, bufferSize);
}

void NnVulkanBuffer::read(NnByte *data, const NnSize size) {
assert(size <= bufferSize);

if (isHostVisible && hostPointer != nullptr) {
context->device.invalidateMappedMemoryRanges({ {deviceMemory, 0, bufferSize} });
std::memcpy(data, hostPointer, bufferSize);
context->device.invalidateMappedMemoryRanges({ {deviceMemory, 0, (vk::DeviceSize)size} });
std::memcpy(data, hostPointer, size);

VULKAN_TRACE("Read %zu bytes from host visible buffer", (NnSize)bufferSize);
VULKAN_TRACE("Read %zu bytes from host visible buffer", size);
} else {
NnVulkanStagingCopy copy(context, deviceBuffer, bufferSize, COPY_FROM_DEVICE);
copy.executeCopyCommand();
copy.copy(data);

VULKAN_TRACE("Read %zu bytes from buffer", (NnSize)bufferSize);
VULKAN_TRACE("Read %zu bytes from buffer", size);
}
}

NnSize NnVulkanBuffer::calcSliceSize(const NnSize nominator, const NnSize denominator) {
assert(bufferSize % denominator == 0);

NnSize size = (bufferSize / denominator) * nominator;
if (context->nonCoherentAtomSize != 0) {
// TODO: this alignment is not needed for coherent memory
size += context->nonCoherentAtomSize - (size % context->nonCoherentAtomSize);
size = std::min(size, bufferSize);
}
return size;
}

static NnByte *findFirstOpConfig(NnNodeConfig *nodeConfig, NnOpCode opCode) {
Expand Down Expand Up @@ -332,11 +357,12 @@ NnVulkanDevice::NnVulkanDevice(NnUint gpuIndex, NnNetConfig *netConfig, NnNodeCo
printf("🌋 Device: %s\n", (char*)deviceProps.deviceName);
printf("🌋 DeviceApiVersion: %d.%d.%d\n", VK_VERSION_MAJOR(deviceProps.apiVersion), VK_VERSION_MINOR(deviceProps.apiVersion), VK_VERSION_PATCH(deviceProps.apiVersion));
printf("🌋 MaxComputeSharedMemory: %d kB\n", deviceProps.limits.maxComputeSharedMemorySize / 1024);
printf("🌋 NonCoherentAtomSize: %lu bytes\n", (NnSize)deviceProps.limits.nonCoherentAtomSize);

vk::PhysicalDeviceMemoryProperties memoryProperties = context.physicalDevice.getMemoryProperties();
for (unsigned int h = 0; h < memoryProperties.memoryHeapCount; h++) {
if (memoryProperties.memoryHeaps[h].flags & vk::MemoryHeapFlagBits::eDeviceLocal)
printf("🌋 Heap[%u]: %lu MB\n", h, ((unsigned long)memoryProperties.memoryHeaps[h].size) / (1024 * 1024));
printf("🌋 Heap[%u]: %lu MB\n", h, ((NnSize)memoryProperties.memoryHeaps[h].size) / (1024 * 1024));
}

vk::PhysicalDeviceFeatures deviceFeatures = context.physicalDevice.getFeatures();
Expand Down Expand Up @@ -375,6 +401,7 @@ NnVulkanDevice::NnVulkanDevice(NnUint gpuIndex, NnNetConfig *netConfig, NnNodeCo
vk::CommandPoolCreateInfo commandPoolCreateInfo(vk::CommandPoolCreateFlags(vk::CommandPoolCreateFlagBits::eTransient | vk::CommandPoolCreateFlagBits::eResetCommandBuffer), context.queueFamilyIndex);
context.commandPool = context.device.createCommandPool(commandPoolCreateInfo);
context.queue = context.device.getQueue(context.queueFamilyIndex, 0);
context.nonCoherentAtomSize = deviceProps.limits.nonCoherentAtomSize;

VULKAN_TRACE("Context created");
data = new NnVulkanDeviceData(&context, netConfig, nodeConfig);
Expand Down Expand Up @@ -790,15 +817,17 @@ void NnVulkanDeviceSegment::forward(NnUint opIndex, NnUint nThreads, NnUint thre
for (NnUint i = 0; i < netConfig->nPreSyncs; i++) {
NnPreSyncConfig *preSyncConfig = &netConfig->preSyncs[i];
NnByte *pipeData = netExecution->pipes[preSyncConfig->pipeIndex];
data->pipes[preSyncConfig->pipeIndex]->write(pipeData);
NnVulkanBuffer *buffer = data->pipes[preSyncConfig->pipeIndex].get();
buffer->write(pipeData, buffer->calcSliceSize(batchSize, netConfig->nBatches));
}
}

for (NnUint opIndex = 0; opIndex < segmentConfig->nOps; opIndex++) {
NnOpConfig *opConfig = &segmentConfig->ops[opIndex];
if (opConfig->input.source == SRC_PIPE) {
NnByte *pipeData = netExecution->pipes[opConfig->input.pointerIndex];
data->pipes[opConfig->input.pointerIndex]->write(pipeData);
NnVulkanBuffer *buffer = data->pipes[opConfig->input.pointerIndex].get();
buffer->write(pipeData, buffer->calcSliceSize(batchSize, netConfig->nBatches));
}
}
}
Expand Down Expand Up @@ -856,7 +885,8 @@ void NnVulkanDeviceSegment::forward(NnUint opIndex, NnUint nThreads, NnUint thre
NnOpConfig *opConfig = &segmentConfig->ops[opIndex];
if (opConfig->output.source == SRC_PIPE) {
NnByte *pipeData = netExecution->pipes[opConfig->output.pointerIndex];
data->pipes[opConfig->output.pointerIndex]->read(pipeData);
NnVulkanBuffer *buffer = data->pipes[opConfig->output.pointerIndex].get();
buffer->read(pipeData, buffer->calcSliceSize(batchSize, netConfig->nBatches));
}
}
}
Expand Down
10 changes: 6 additions & 4 deletions src/nn/nn-vulkan.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,15 +6,14 @@
#include "nn-executor.hpp"
#include "nn-cpu-ops.hpp"

#define DEBUG_VULKAN_TRACE false

typedef struct {
vk::Instance instance;
vk::PhysicalDevice physicalDevice;
vk::Device device;
uint32_t queueFamilyIndex;
vk::CommandPool commandPool;
vk::Queue queue;
NnSize nonCoherentAtomSize;
} NnVulkanContext;

enum NnStagingVulkanCopyDirection {
Expand Down Expand Up @@ -47,13 +46,16 @@ class NnVulkanBuffer {
void *hostPointer;
public:
const char *name;
vk::DeviceSize bufferSize;
NnSize bufferSize;
vk::Buffer deviceBuffer;
vk::BufferUsageFlags usageFlags;
NnVulkanBuffer(NnVulkanContext *context, const char *name, const vk::DeviceSize bufferSize, vk::BufferUsageFlags usageFlags, bool fastAccess);
NnVulkanBuffer(NnVulkanContext *context, const char *name, const NnSize bufferSize, vk::BufferUsageFlags usageFlags, bool fastAccess);
~NnVulkanBuffer();
void write(const NnByte *data);
void write(const NnByte *data, const NnSize size);
void read(NnByte *data);
void read(NnByte *data, const NnSize size);
NnSize calcSliceSize(const NnSize nominator, const NnSize denominator);
};

typedef struct {
Expand Down
Loading