Skip to content

vulkan: subgroup size tuning #12087

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 10 commits into from
Mar 17, 2025
Merged

vulkan: subgroup size tuning #12087

Changes from 5 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
53b69a3
vulkan: subgroup size test
daniandtheweb Mar 8, 2025
85e15e6
Vulkan: Add device architecture enum and logic to recognize AMD gener…
0cc4m Mar 8, 2025
7695541
vulkan: use new architecture logic to specify subgroup size
daniandtheweb Mar 8, 2025
43c3e6f
Initial vulkan subgroup size tuning for RDNA3
daniandtheweb Mar 11, 2025
c41619d
vulkan: commonize RDNA subgroup tuning
daniandtheweb Mar 12, 2025
1c17520
vulkan: override subgroup size if required_subgroup_size = 0
daniandtheweb Mar 12, 2025
afb5c2d
vulkan: disable warp 32 for RDNA3
daniandtheweb Mar 12, 2025
29e8104
vulkan: fine tuned RDNA1 subgroup sizes
daniandtheweb Mar 13, 2025
bf7352e
vulkan: adjusted subgroup size map
daniandtheweb Mar 16, 2025
d43537a
vulkan: fixed RDNA2 subgroup map
daniandtheweb Mar 16, 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
149 changes: 139 additions & 10 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 @@ -149,6 +149,66 @@ static void ggml_vk_destroy_buffer(vk_buffer& buf);

static constexpr uint32_t mul_mat_vec_max_cols = 8;

enum vk_device_architecture {
OTHER,
AMD_GCN,
AMD_RDNA1,
AMD_RDNA2,
AMD_RDNA3,
};

static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
vk::PhysicalDeviceProperties props = device.getProperties();

if (props.vendorID == VK_VENDOR_ID_AMD) {
const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();

bool amd_shader_core_properties = false;
bool integer_dot_product = false;
bool subgroup_size_control = false;

for (const auto& properties : ext_props) {
if (strcmp("VK_AMD_shader_core_properties", properties.extensionName) == 0) {
amd_shader_core_properties = true;
} else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0) {
integer_dot_product = true;
} else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
subgroup_size_control = true;
}
}

if (!amd_shader_core_properties || !integer_dot_product || !subgroup_size_control) {
return vk_device_architecture::OTHER;
}

vk::PhysicalDeviceProperties2 props2;
vk::PhysicalDeviceShaderCorePropertiesAMD shader_core_props_amd;
vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props;
vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;

props2.pNext = &shader_core_props_amd;
shader_core_props_amd.pNext = &integer_dot_props;
integer_dot_props.pNext = &subgroup_size_control_props;

device.getProperties2(&props2);

if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 64) {
return vk_device_architecture::AMD_GCN;
}
if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 32) {
// RDNA
if (shader_core_props_amd.wavefrontsPerSimd == 20) {
return vk_device_architecture::AMD_RDNA1;
}
if (integer_dot_props.integerDotProduct4x8BitPackedMixedSignednessAccelerated) {
return vk_device_architecture::AMD_RDNA3;
}
return vk_device_architecture::AMD_RDNA2;
}
}
return vk_device_architecture::OTHER;
}

struct vk_device_struct {
std::mutex mutex;

Expand All @@ -161,6 +221,7 @@ struct vk_device_struct {
bool pipeline_robustness;
vk::Device device;
uint32_t vendor_id;
vk_device_architecture architecture;
vk_queue compute_queue;
vk_queue transfer_queue;
bool single_queue;
Expand Down Expand Up @@ -1445,6 +1506,67 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
return supported;
}

struct GpuPipelineConfig {
// GPU architecture identifier.
// Example: vk_device_architecture::AMD_GCN
vk_device_architecture arch;

// 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;
};

// Common pipeline configuration for RDNA GPUs.
static const std::unordered_map<std::string, uint32_t> rdna_pipelines = {
{"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},
};
static constexpr uint32_t RDNA_DEFAULT_SUBGROUP_SIZE = 32;

// Define configurations for different GPUs.
static std::vector<GpuPipelineConfig> gpu_pipeline_configs = {
{
vk_device_architecture::AMD_RDNA1,
{
rdna_pipelines,
},
RDNA_DEFAULT_SUBGROUP_SIZE
},
{
vk_device_architecture::AMD_RDNA2,
{
rdna_pipelines,
},
RDNA_DEFAULT_SUBGROUP_SIZE
},
{
vk_device_architecture::AMD_RDNA3,
{
rdna_pipelines,
},
RDNA_DEFAULT_SUBGROUP_SIZE
},
};

static uint32_t get_subgroup_size(const std::string &pipeline_name, const vk_device_architecture &arch) {
for (const auto &config : gpu_pipeline_configs) {
if (config.arch == arch) {
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 @@ -1571,6 +1693,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
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) {

if (!require_full_subgroups) {
required_subgroup_size = get_subgroup_size(name, device->architecture);
}

if (!pipeline) {
pipeline = std::make_shared<vk_pipeline_struct>();
pipeline->name = name;
Expand Down Expand Up @@ -2247,7 +2373,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
device->need_compiles = false;
}

static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props);
static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch);

static vk_device ggml_vk_get_device(size_t idx) {
VK_LOG_DEBUG("ggml_vk_get_device(" << idx << ")");
Expand Down Expand Up @@ -2276,6 +2402,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->physical_device = physical_devices[dev_num];
const std::vector<vk::ExtensionProperties> ext_props = device->physical_device.enumerateDeviceExtensionProperties();

device->architecture = get_device_architecture(device->physical_device);

const char* GGML_VK_PREFER_HOST_MEMORY = getenv("GGML_VK_PREFER_HOST_MEMORY");
device->prefer_host_memory = GGML_VK_PREFER_HOST_MEMORY != nullptr;

Expand All @@ -2288,7 +2416,6 @@ static vk_device ggml_vk_get_device(size_t idx) {
bool coopmat2_support = false;
device->coopmat_support = false;

// Check if maintenance4 is supported
for (const auto& properties : ext_props) {
if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
maintenance4_support = true;
Expand Down Expand Up @@ -2401,7 +2528,7 @@ static vk_device ggml_vk_get_device(size_t idx) {

device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;

if (!ggml_vk_khr_cooperative_matrix_support(device->properties, driver_props)) {
if (!ggml_vk_khr_cooperative_matrix_support(device->properties, driver_props, device->architecture)) {
device->coopmat_support = false;
}

Expand Down Expand Up @@ -2779,7 +2906,10 @@ static void ggml_vk_print_gpu_info(size_t idx) {
subgroup_props.pNext = &driver_props;
physical_device.getProperties2(&props2);

const size_t subgroup_size = subgroup_props.subgroupSize;
vk_device_architecture arch = get_device_architecture(physical_device);
uint32_t default_subgroup_size = get_subgroup_size("", arch);
const size_t subgroup_size = (default_subgroup_size != 0) ? default_subgroup_size : subgroup_props.subgroupSize;

const bool uma = props2.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;

bool fp16_storage = false;
Expand All @@ -2805,7 +2935,9 @@ static void ggml_vk_print_gpu_info(size_t idx) {
}
}

if (!ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props)) {
const vk_device_architecture device_architecture = get_device_architecture(physical_device);

if (!ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props, device_architecture)) {
coopmat_support = false;
}

Expand Down Expand Up @@ -8826,18 +8958,15 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve
UNUSED(instance_extensions);
}

static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props) {
static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) {
switch (props.vendorID) {
case VK_VENDOR_ID_INTEL:
// Intel drivers don't support coopmat properly yet
return false;
case VK_VENDOR_ID_AMD:
if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) {
// Workaround for AMD proprietary driver reporting support on all GPUs
const std::string name = props.deviceName;
return name.rfind("AMD Radeon RX 7", 0) == 0 || name.rfind("AMD Radeon(TM) RX 7", 0) == 0 || // RDNA 3 consumer GPUs
name.rfind("AMD Radeon PRO W7", 0) == 0 || name.rfind("AMD Radeon(TM) PRO W7", 0) == 0 || // RDNA 3 workstation GPUs
name.rfind("AMD Radeon 7", 0) == 0 || name.rfind("AMD Radeon(TM) 7", 0) == 0; // RDNA 3 APUs
return arch == vk_device_architecture::AMD_RDNA3;
}
return true;
default:
Expand Down
Loading