Skip to content

batch infrastructure #6656

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

Open
nihui wants to merge 15 commits into
base: master
Choose a base branch
from
Open

batch infrastructure #6656

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
378795f
batch infrastructure
nihui Apr 7, 2026
2aad18c
apply code-format changes
nihui Apr 7, 2026
e9a2dee
cpu batch works
nihui Apr 8, 2026
8227c6d
Merge branch 'b' of https://github.com/nihui/ncnn into b
nihui Apr 8, 2026
ea4a577
enable gpu batch test
nihui Apr 8, 2026
dcbdee1
gpu batch
nihui Apr 8, 2026
b17ef4e
f
nihui Apr 8, 2026
a54144a
use offset
nihui Apr 8, 2026
a2c84cd
apply code-format changes
nihui Apr 8, 2026
b921fe2
batch-aware record upload download
nihui Apr 8, 2026
c7f7667
w
nihui Apr 8, 2026
c25a57e
cast and packing batch
nihui Apr 8, 2026
4e076cd
f
nihui Apr 8, 2026
240a2d9
s
nihui Apr 8, 2026
d637571
ss
nihui Apr 8, 2026
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
138 changes: 86 additions & 52 deletions src/command.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -359,10 +359,12 @@ void VkCompute::record_upload(const Mat& src, VkMat& dst, const Option& opt)
{
// NCNN_LOGE("record_upload buffer");

const int B = src.n;

// cpu cast to fp16 (discrete gpu)
Mat src_fp16;
if (src.elemsize == src.elempack * 4u)
{
// cpu cast to fp16 (discrete gpu)
if (vkdev->info.type() == 0 && (opt.use_bf16_storage || opt.use_bf16_packed))
{
ncnn::cast_float32_to_bfloat16(src, src_fp16, opt);
Expand All @@ -389,26 +391,32 @@ void VkCompute::record_upload(const Mat& src, VkMat& dst, const Option& opt)
src_fp16 = src_fp16_pack4;
}

// upload
// upload staging buffer
VkMat dst_staging;
dst_staging.create_like(src_fp16, opt.staging_vkallocator);
if (B > 1)
dst_staging.create_like_batch(src_fp16.batch(0), B, opt.staging_vkallocator);
else
dst_staging.create_like(src_fp16, opt.staging_vkallocator);
if (dst_staging.empty())
return;

// stash staging
d->upload_staging_buffers.push_back(dst_staging);

// NCNN_LOGE("upload_staging_buffer %p -> %p +%d ~%d", src_fp16.data, dst_staging.buffer(), dst_staging.buffer_offset(), dst_staging.buffer_capacity());

// memcpy src to device
memcpy(dst_staging.mapped_ptr(), src_fp16.data, src_fp16.total() * src_fp16.elemsize);
for (int b = 0; b < B; b++)
{
const Mat src_b = src_fp16.batch(b);
VkMat staging_b = dst_staging.batch(b);
memcpy(staging_b.mapped_ptr(), src_b.data, src_b.total() * src_b.elemsize);
}
dst_staging.allocator->flush(dst_staging.data);

// mark device host-write @ null
dst_staging.data->access_flags = VK_ACCESS_HOST_WRITE_BIT;
dst_staging.data->stage_flags = VK_PIPELINE_STAGE_HOST_BIT;

// resolve dst_elempack
// resolve dst_elempack (from single sample dimensions)
int dims = src_fp16.dims;
int elemcount = 0;
if (dims == 1) elemcount = src_fp16.elempack * src_fp16.w;
Expand All @@ -435,6 +443,8 @@ void VkCompute::record_download(const VkMat& src, Mat& dst, const Option& opt)
{
// NCNN_LOGE("record_download buffer");

const int B = src.n;

// resolve dst_elempack
int dims = src.dims;
int elemcount = 0;
Expand Down Expand Up @@ -480,7 +490,7 @@ void VkCompute::record_download(const VkMat& src, Mat& dst, const Option& opt)
barriers[0].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
barriers[0].buffer = dst_staging.buffer();
barriers[0].offset = dst_staging.buffer_offset();
barriers[0].size = dst_staging.buffer_capacity();
barriers[0].size = B > 1 ? dst_staging.nstep * B * dst_staging.elemsize : dst_staging.buffer_capacity();

VkPipelineStageFlags src_stage = dst_staging.data->stage_flags;
VkPipelineStageFlags dst_stage = VK_PIPELINE_STAGE_HOST_BIT;
Expand Down Expand Up @@ -509,7 +519,10 @@ void VkCompute::record_download(const VkMat& src, Mat& dst, const Option& opt)

// create dst
Mat dst_fp16;
dst_fp16.create_like(dst_staging, opt.blob_allocator);
if (B > 1)
dst_fp16.create_like_batch(dst_staging.batch(0), B, opt.blob_allocator);
else
dst_fp16.create_like(dst_staging, opt.blob_allocator);
if (dst_fp16.empty())
return;

Expand All @@ -530,48 +543,42 @@ void VkCompute::record_download(const VkMat& src, Mat& dst, const Option& opt)
// cast to fp32 (discrete gpu)
if (dst_fp16.elemsize == dst_fp16.elempack * 2u)
{
int post_cast_type = 0; // 0=none, 1=bf16, 2=fp16
if (vkdev->info.type() == 0 && (opt.use_bf16_storage || opt.use_bf16_packed))
{
int dims = dst_fp16.dims;
if (dims == 1)
dst.create(dst_fp16.w, (size_t)(dst_fp16.elempack * 4u), dst_fp16.elempack, opt.blob_allocator);
if (dims == 2)
dst.create(dst_fp16.w, dst_fp16.h, (size_t)(dst_fp16.elempack * 4u), dst_fp16.elempack, opt.blob_allocator);
if (dims == 3)
dst.create(dst_fp16.w, dst_fp16.h, dst_fp16.c, (size_t)(dst_fp16.elempack * 4u), dst_fp16.elempack, opt.blob_allocator);
if (dims == 4)
dst.create(dst_fp16.w, dst_fp16.h, dst_fp16.d, dst_fp16.c, (size_t)(dst_fp16.elempack * 4u), dst_fp16.elempack, opt.blob_allocator);

d->download_post_mats.push_back(dst);

VkComputePrivate::record r;
r.type = VkComputePrivate::record::TYPE_post_cast_bfloat16_to_float32;
r.command_buffer = 0;
r.post_cast_bfloat16_to_float32.download_post_mat_bf16_offset = d->download_post_mats_fp16.size() - 1;
r.post_cast_bfloat16_to_float32.download_post_mat_offset = d->download_post_mats.size() - 1;
r.post_cast_bfloat16_to_float32.num_threads = opt.num_threads;
d->delayed_records.push_back(r);
}
post_cast_type = 1;
else if (vkdev->info.type() == 0 && (opt.use_fp16_storage || opt.use_fp16_packed))
post_cast_type = 2;

if (post_cast_type > 0)
{
int dims = dst_fp16.dims;
if (dims == 1)
dst.create(dst_fp16.w, (size_t)(dst_fp16.elempack * 4u), dst_fp16.elempack, opt.blob_allocator);
if (dims == 2)
dst.create(dst_fp16.w, dst_fp16.h, (size_t)(dst_fp16.elempack * 4u), dst_fp16.elempack, opt.blob_allocator);
if (dims == 3)
dst.create(dst_fp16.w, dst_fp16.h, dst_fp16.c, (size_t)(dst_fp16.elempack * 4u), dst_fp16.elempack, opt.blob_allocator);
if (dims == 4)
dst.create(dst_fp16.w, dst_fp16.h, dst_fp16.d, dst_fp16.c, (size_t)(dst_fp16.elempack * 4u), dst_fp16.elempack, opt.blob_allocator);
size_t fp32_elemsize = (size_t)(dst_fp16.elempack * 4u);
if (dst_fp16.dims == 1)
dst.create_batch(dst_fp16.w, B, fp32_elemsize, dst_fp16.elempack, opt.blob_allocator);
else if (dst_fp16.dims == 2)
dst.create_batch(dst_fp16.w, dst_fp16.h, B, fp32_elemsize, dst_fp16.elempack, opt.blob_allocator);
else if (dst_fp16.dims == 3)
dst.create_batch(dst_fp16.w, dst_fp16.h, dst_fp16.c, B, fp32_elemsize, dst_fp16.elempack, opt.blob_allocator);
else if (dst_fp16.dims == 4)
dst.create_batch(dst_fp16.w, dst_fp16.h, dst_fp16.d, dst_fp16.c, B, fp32_elemsize, dst_fp16.elempack, opt.blob_allocator);

d->download_post_mats.push_back(dst);

VkComputePrivate::record r;
r.type = VkComputePrivate::record::TYPE_post_cast_float16_to_float32;
r.command_buffer = 0;
r.post_cast_float16_to_float32.download_post_mat_fp16_offset = d->download_post_mats_fp16.size() - 1;
r.post_cast_float16_to_float32.download_post_mat_offset = d->download_post_mats.size() - 1;
r.post_cast_float16_to_float32.num_threads = opt.num_threads;
if (post_cast_type == 1)
{
r.type = VkComputePrivate::record::TYPE_post_cast_bfloat16_to_float32;
r.post_cast_bfloat16_to_float32.download_post_mat_bf16_offset = d->download_post_mats_fp16.size() - 1;
r.post_cast_bfloat16_to_float32.download_post_mat_offset = d->download_post_mats.size() - 1;
r.post_cast_bfloat16_to_float32.num_threads = opt.num_threads;
}
else
{
r.type = VkComputePrivate::record::TYPE_post_cast_float16_to_float32;
r.post_cast_float16_to_float32.download_post_mat_fp16_offset = d->download_post_mats_fp16.size() - 1;
r.post_cast_float16_to_float32.download_post_mat_offset = d->download_post_mats.size() - 1;
r.post_cast_float16_to_float32.num_threads = opt.num_threads;
}
d->delayed_records.push_back(r);
}
else
Expand All @@ -589,14 +596,24 @@ void VkCompute::record_clone(const Mat& src, VkMat& dst, const Option& opt)
{
// NCNN_LOGE("record_clone host to buffer");

const int B = src.n;

// host to staging
VkMat dst_staging;
dst_staging.create_like(src, opt.staging_vkallocator);
if (B > 1)
dst_staging.create_like_batch(src.batch(0), B, opt.staging_vkallocator);
else
dst_staging.create_like(src, opt.staging_vkallocator);
if (dst_staging.empty())
return;

// memcpy src to device
memcpy(dst_staging.mapped_ptr(), src.data, src.total() * src.elemsize);
for (int b = 0; b < B; b++)
{
const Mat src_b = src.batch(b);
VkMat staging_b = dst_staging.batch(b);
memcpy(staging_b.mapped_ptr(), src_b.data, src_b.total() * src_b.elemsize);
}
dst_staging.allocator->flush(dst_staging.data);

// mark device host-write @ null
Expand Down Expand Up @@ -631,6 +648,8 @@ void VkCompute::record_clone(const VkMat& src, Mat& dst, const Option& opt)
{
// NCNN_LOGE("record_clone buffer to host");

const int B = src.n;

if (!src.allocator->mappable)
{
// device to staging
Expand All @@ -646,7 +665,10 @@ void VkCompute::record_clone(const VkMat& src, Mat& dst, const Option& opt)
}

// create dst
dst.create_like(src, opt.blob_allocator);
if (B > 1)
dst.create_like_batch(src.batch(0), B, opt.blob_allocator);
else
dst.create_like(src, opt.blob_allocator);
if (dst.empty())
return;

Expand All @@ -662,7 +684,7 @@ void VkCompute::record_clone(const VkMat& src, Mat& dst, const Option& opt)
barriers[0].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
barriers[0].buffer = src.buffer();
barriers[0].offset = src.buffer_offset();
barriers[0].size = src.buffer_capacity();
barriers[0].size = B > 1 ? src.nstep * B * src.elemsize : src.buffer_capacity();

VkPipelineStageFlags src_stage = src.data->stage_flags;
VkPipelineStageFlags dst_stage = VK_PIPELINE_STAGE_HOST_BIT;
Expand Down Expand Up @@ -722,8 +744,13 @@ void VkCompute::record_clone(const VkMat& src, VkMat& dst, const Option& opt)
{
// NCNN_LOGE("record_clone buffer to buffer");

const int B = src.n;

// create dst
dst.create_like(src, opt.blob_vkallocator);
if (B > 1)
dst.create_like_batch(src.batch(0), B, opt.blob_vkallocator);
else
dst.create_like(src, opt.blob_vkallocator);
if (dst.empty())
return;

Expand All @@ -739,7 +766,7 @@ void VkCompute::record_clone(const VkMat& src, VkMat& dst, const Option& opt)
barriers[0].dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
barriers[0].buffer = src.buffer();
barriers[0].offset = src.buffer_offset();
barriers[0].size = src.buffer_capacity();
barriers[0].size = B > 1 ? src.nstep * B * src.elemsize : src.buffer_capacity();

VkPipelineStageFlags src_stage = src.data->stage_flags;
VkPipelineStageFlags dst_stage = VK_PIPELINE_STAGE_TRANSFER_BIT;
Expand Down Expand Up @@ -774,12 +801,14 @@ void VkCompute::record_clone(const VkMat& src, VkMat& dst, const Option& opt)
dst.data->stage_flags = VK_PIPELINE_STAGE_TRANSFER_BIT;
}

// record device to staging
// record copy
{
VkDeviceSize copy_size = B > 1 ? src.nstep * B * src.elemsize : std::min(src.buffer_capacity(), dst.buffer_capacity());

VkBufferCopy* regions = new VkBufferCopy[1];
regions[0].srcOffset = src.buffer_offset();
regions[0].dstOffset = dst.buffer_offset();
regions[0].size = std::min(src.buffer_capacity(), dst.buffer_capacity());
regions[0].size = copy_size;

if (vkdev->info.support_VK_KHR_push_descriptor())
{
Expand Down Expand Up @@ -1985,7 +2014,12 @@ int VkCompute::submit_and_wait()
// NCNN_LOGE("post_download %p +%d ~%d -> %p", src.buffer(), src.buffer_offset(), src.buffer_capacity(), dst.data);

src.allocator->invalidate(src.data);
memcpy(dst.data, src.mapped_ptr(), dst.total() * dst.elemsize);
for (int b = 0; b < dst.n; b++)
{
Mat dst_b = dst.batch(b);
size_t src_batch_offset = src.nstep * b * src.elemsize;
memcpy(dst_b.data, (const unsigned char*)src.mapped_ptr() + src_batch_offset, dst_b.total() * dst_b.elemsize);
}
break;
}
case VkComputePrivate::record::TYPE_post_cast_float16_to_float32:
Expand Down
3 changes: 3 additions & 0 deletions src/layer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,8 @@ Layer::Layer()
support_any_packing = false;
support_vulkan_any_packing = false;

support_batch = false;

featmask = 0;

#if NCNN_VULKAN
Expand Down Expand Up @@ -240,6 +242,7 @@ class Layer_final : public Layer
support_fp16_storage = layer_cpu->support_fp16_storage;
support_int8_storage = layer_cpu->support_int8_storage;
support_any_packing = layer_cpu->support_any_packing;
support_batch = layer_cpu->support_batch;

support_vulkan = false;
support_tensor_storage = false;
Expand Down
3 changes: 2 additions & 1 deletion src/layer.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,7 +75,8 @@ class NCNN_EXPORT Layer
// vulkan accept input blob with any elempack
bool support_vulkan_any_packing;

bool support_reserved_1;
// support batched input (n > 1), replaces support_reserved_1
bool support_batch;
bool support_reserved_2;
bool support_reserved_3;
bool support_reserved_4;
Expand Down
28 changes: 12 additions & 16 deletions src/layer/arm/cast_arm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,13 +40,14 @@ int Cast_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
int dims = bottom_blob.dims;
size_t elemsize = bottom_blob.elemsize;
int elempack = bottom_blob.elempack;
int batch = bottom_blob.n;

size_t out_elemsize = elemsize;
if (type_to == 1)
{
if (type_from == 3)
{
Cast::forward(bottom_blob, top_blob, opt);
return Cast::forward(bottom_blob, top_blob, opt);
}

// float32
Expand All @@ -69,21 +70,13 @@ int Cast_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt)
}

if (dims == 1)
{
top_blob.create(w, out_elemsize, elempack, opt.blob_allocator);
}
top_blob.create_batch(w, batch, out_elemsize, elempack, opt.blob_allocator);
else if (dims == 2)
{
top_blob.create(w, h, out_elemsize, elempack, opt.blob_allocator);
}
top_blob.create_batch(w, h, batch, out_elemsize, elempack, opt.blob_allocator);
else if (dims == 3)
{
top_blob.create(w, h, channels, out_elemsize, elempack, opt.blob_allocator);
}
top_blob.create_batch(w, h, channels, batch, out_elemsize, elempack, opt.blob_allocator);
else if (dims == 4)
{
top_blob.create(w, h, d, channels, out_elemsize, elempack, opt.blob_allocator);
}
top_blob.create_batch(w, h, d, channels, batch, out_elemsize, elempack, opt.blob_allocator);
if (top_blob.empty())
return -100;

Expand All @@ -101,11 +94,14 @@ int Cast_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt)

if (type_from == 3 && type_to == 1)
{
const int total_bc = batch * channels;
#pragma omp parallel for num_threads(opt.num_threads)
for (int q = 0; q < channels; q++)
for (int bc = 0; bc < total_bc; bc++)
{
const signed char* ptr = bottom_blob.channel(q);
float* outptr = top_blob.channel(q);
int b = bc / channels;
int q = bc % channels;
const signed char* ptr = bottom_blob.batch(b).channel(q);
float* outptr = top_blob.batch(b).channel(q);

for (int i = 0; i < size; i++)
{
Expand Down
Loading
Loading