Skip to content

vulkan: improve im2col #11826

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 12 commits into from
Feb 28, 2025
Merged

vulkan: improve im2col #11826

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
62733f2
vulkan: improve im2col performance
daniandtheweb Feb 9, 2025
35f6369
Force subgroup 32 on RX 5700 and subgroup 64 for im2col
daniandtheweb Feb 14, 2025
293edef
Fixed uint
daniandtheweb Feb 14, 2025
14ea4fa
Helper function to set subgroup size
daniandtheweb Feb 14, 2025
04100e8
Optimize soft_max
daniandtheweb Feb 14, 2025
9036e7a
Apply map to all NAVI1* gpus
daniandtheweb Feb 15, 2025
d151973
Improved gpu configuration map
daniandtheweb Feb 15, 2025
0e5dd68
Cleanup im2col shader
daniandtheweb Feb 16, 2025
27f1301
Fix subgroup override
daniandtheweb Feb 22, 2025
9c9b812
subgroup test improvements
daniandtheweb Feb 22, 2025
4a3988e
Fix wave32 not correctly used
daniandtheweb Feb 22, 2025
c49419f
Removed subgroup changes
daniandtheweb Feb 26, 2025
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
49 changes: 49 additions & 0 deletions ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1423,6 +1423,49 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
return supported;
}

struct GpuPipelineConfig {
// List of all aliases for a given GPU.
// For example, this can include names like "NAVI10", "RX 5700", etc.
std::vector<std::string> device_names;

// Mapping of pipeline names to their specific subgroup sizes.
// Example: {"soft_max_f32", 64}.
std::unordered_map<std::string, uint32_t> pipelines;

// Default subgroup size for this GPU.
// Defaults to 0 if not explicitly provided.
uint32_t default_subgroup_size = 0;
};

// Define configurations for different GPUs.
static std::vector<GpuPipelineConfig> gpu_pipeline_configs = {
{
{"NAVI10", "NAVI14", "RX 5700", "RX 5600", "RX 5500"},
{
{"soft_max_f32", 64}, {"soft_max_f32_wg512", 64},
{"soft_max_f32_f16", 64}, {"soft_max_f32_f16_wg512", 64},
{"im2col_f32", 64}, {"im2col_f32_f16", 64},
},
32
},
};

static uint32_t get_subgroup_size(const std::string &pipeline_name, const std::string &device_name) {
for (const auto &config : gpu_pipeline_configs) {
for (const auto &alias : config.device_names) {
if (device_name.find(alias) != std::string::npos) {
auto pipIt = config.pipelines.find(pipeline_name);
if (pipIt != config.pipelines.end() && pipIt->second != 0) {
return pipIt->second;
}
return config.default_subgroup_size;
}
}
}
// If no matching configuration is found, return 0.
return 0;
}

static void ggml_vk_load_shaders(vk_device& device) {
VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");

Expand Down Expand Up @@ -1543,11 +1586,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
}

vk::PhysicalDeviceProperties2 props2;
device->physical_device.getProperties2(&props2);
std::string device_name = props2.properties.deviceName.data();
Copy link
Collaborator

@0cc4m 0cc4m Feb 21, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think this is needed anymore (also, the device name is available in device->name, and the properties in device->properties)

Edit: I forgot it's used by the get_subgroup_size function. But just use the device field.

Copy link
Contributor Author

@daniandtheweb daniandtheweb Feb 22, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If I don't set the device name like this the shader executes at wave64 speed instead of wave32 and I get no performance improvement at all.

Copy link
Collaborator

@0cc4m 0cc4m Feb 25, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I checked and device->name contains just something like Vulkan0, so that's why. But if you need the actual device name, it should be stored in the device struct on init and only accessed here, something like physica_device_name.

But as mentioned in my other comment, we can probably ignore device names for now.


std::vector<std::future<void>> compiles;
auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint,
uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants,
uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) {

required_subgroup_size = get_subgroup_size(name, device_name);

if (!pipeline) {
pipeline = std::make_shared<vk_pipeline_struct>();
pipeline->name = name;
Expand Down
49 changes: 31 additions & 18 deletions ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,20 @@ void main() {
const uint batch = gl_GlobalInvocationID.z / p.IC;
const uint ic = gl_GlobalInvocationID.z % p.IC;

const uint src_base = ic * p.offset_delta + batch * p.batch_offset;
const uint dst_base = ((batch * p.OH + oh) * p.OW) * p.CHW + ic * (p.KW * p.KH);
const int oh_s1 = int(oh) * p.s1;
const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);

const uint base_linear_idx = gidx * NUM_ITER;

const uint max_ky = ksize / p.OW;

uint current_kx = base_linear_idx / ksize;
const uint rem = base_linear_idx - (current_kx * ksize);
uint current_ky = rem / p.OW;
uint current_ix = rem % p.OW;

A_TYPE values[NUM_ITER];
uint offset_dst[NUM_ITER];
[[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
Expand All @@ -48,36 +62,35 @@ void main() {

[[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {

const uint i = gidx * NUM_ITER + idx;
const uint linear_idx = base_linear_idx + idx;

const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
const uint kx = i / ksize;
const uint kd = kx * ksize;
const uint ky = (i - kd) / p.OW;
const uint ix = i % p.OW;
if (linear_idx >= p.pelements) {
continue;
}

const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
const uint iiw = current_ix * p.s0 + current_kx * p.d0 - p.p0;
const uint iih = oh_s1 + current_ky * p.d1 - p.p1;

offset_dst[idx] =
((batch * p.OH + oh) * p.OW + ix) * p.CHW +
(ic * (p.KW * p.KH) + ky * p.KW + kx);
offset_dst[idx] = dst_base + current_ix * p.CHW + current_ky * p.KW + current_kx;

if (i >= p.pelements) {
continue;
if ((iih < p.IH) && (iiw < p.IW)) {
values[idx] = data_a[src_base + iih * p.IW + iiw];
}

if (iih < p.IH && iiw < p.IW) {
const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
values[idx] = data_a[offset_src + iih * p.IW + iiw];
if (++current_ix == p.OW) {
current_ix = 0;
if (++current_ky == max_ky) {
current_ky = 0;
current_kx++;
}
}
}

[[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {

const uint i = gidx * NUM_ITER + idx;
const uint linear_idx = base_linear_idx + idx;

if (i >= p.pelements) {
if (linear_idx >= p.pelements) {
continue;
}

Expand Down
Loading