Class: Ignis::CUDA::Device

Inherits:
Object
  • Object
show all
Defined in:
lib/nvruby/cuda/device.rb

Overview

Represents a CUDA GPU device.

Uses RuntimeAPI (Fiddle-based) for hot-path attribute queries and DeviceProperties (FFI::Struct) for full device property reads.

Instance Attribute Summary collapse

Class Method Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(index) ⇒ Device

Returns a new instance of Device.

Parameters:

  • index (Integer) —

    Device index



52
53
54
55
56
57
58
 
# File 'lib/nvruby/cuda/device.rb', line 52

def initialize(index)
  @index = index
  @attribute_cache = {}
  @name = nil
  @total_memory = nil
  @props_cache = nil
end

Instance Attribute Details

#index ⇒ Integer (readonly)

Returns Device index.

Returns:

  • (Integer) —

    Device index



49
50
51
 
# File 'lib/nvruby/cuda/device.rb', line 49

def index
  @index
end

Class Method Details

.available? ⇒ Boolean

Check if any CUDA device is available.

Returns:

  • (Boolean) 


254
255
256
257
258
 
# File 'lib/nvruby/cuda/device.rb', line 254

def available?
  count.positive?
rescue CudaRuntimeError
  false
end

.count ⇒ Integer

Get the number of CUDA devices.

Returns:

  • (Integer) 


228
229
230
231
 
# File 'lib/nvruby/cuda/device.rb', line 228

def count
  RuntimeAPI.ensure_loaded!
  RuntimeAPI.get_device_count
end

.current ⇒ Device

Get the current device.

Returns:



241
242
243
244
 
# File 'lib/nvruby/cuda/device.rb', line 241

def current
  RuntimeAPI.ensure_loaded!
  new(RuntimeAPI.get_device)
end

.default ⇒ Device

Get the default device from configuration.

Returns:



248
249
250
 
# File 'lib/nvruby/cuda/device.rb', line 248

def default
  new(Ignis.configuration.default_device)
end

.list ⇒ Array<Device>

List all available devices.

Returns:



235
236
237
 
# File 'lib/nvruby/cuda/device.rb', line 235

def list
  count.times.map { |i| new(i) }
end

Instance Method Details

#clock_rate ⇒ Integer

Returns Clock rate in kHz.

Returns:

  • (Integer) —

    Clock rate in kHz



128
129
130
 
# File 'lib/nvruby/cuda/device.rb', line 128

def clock_rate
  get_attribute(DeviceAttribute::CLOCK_RATE)
end

#compute_capability ⇒ String

Returns Compute capability as “major.minor”.

Returns:

  • (String) —

    Compute capability as “major.minor”



90
91
92
 
# File 'lib/nvruby/cuda/device.rb', line 90

def compute_capability
  "#{compute_capability_major}.#{compute_capability_minor}"
end

#compute_capability_major ⇒ Integer

Returns Major compute capability.

Returns:

  • (Integer) —

    Major compute capability



95
96
97
 
# File 'lib/nvruby/cuda/device.rb', line 95

def compute_capability_major
  get_attribute(DeviceAttribute::COMPUTE_CAPABILITY_MAJOR)
end

#compute_capability_minor ⇒ Integer

Returns Minor compute capability.

Returns:

  • (Integer) —

    Minor compute capability



100
101
102
 
# File 'lib/nvruby/cuda/device.rb', line 100

def compute_capability_minor
  get_attribute(DeviceAttribute::COMPUTE_CAPABILITY_MINOR)
end

#concurrent_kernels? ⇒ Boolean

Returns Whether concurrent kernels are supported.

Returns:

  • (Boolean) —

    Whether concurrent kernels are supported



153
154
155
 
# File 'lib/nvruby/cuda/device.rb', line 153

def concurrent_kernels?
  get_attribute(DeviceAttribute::CONCURRENT_KERNELS) != 0
end

#cooperative_launch? ⇒ Boolean

Returns Whether cooperative launch is supported.

Returns:

  • (Boolean) —

    Whether cooperative launch is supported



168
169
170
 
# File 'lib/nvruby/cuda/device.rb', line 168

def cooperative_launch?
  get_attribute(DeviceAttribute::COOPERATIVE_LAUNCH) != 0
end

#ecc_enabled? ⇒ Boolean

Returns Whether ECC memory is enabled.

Returns:

  • (Boolean) —

    Whether ECC memory is enabled



148
149
150
 
# File 'lib/nvruby/cuda/device.rb', line 148

def ecc_enabled?
  get_attribute(DeviceAttribute::ECC_ENABLED) != 0
end

#free_memory ⇒ Integer

Returns Free memory in bytes.

Returns:

  • (Integer) —

    Free memory in bytes



181
182
183
 
# File 'lib/nvruby/cuda/device.rb', line 181

def free_memory
  memory_info[:free_bytes]
end

#inspect ⇒ String

Returns Detailed inspection.

Returns:

  • (String) —

    Detailed inspection



220
221
222
223
 
# File 'lib/nvruby/cuda/device.rb', line 220

def inspect
  "#<Ignis::CUDA::Device:#{object_id} index=#{@index} name=#{name.inspect} " \
    "compute=#{compute_capability} memory=#{total_memory}>"
end

#l2_cache_size ⇒ Integer

Returns L2 cache size in bytes.

Returns:

  • (Integer) —

    L2 cache size in bytes



143
144
145
 
# File 'lib/nvruby/cuda/device.rb', line 143

def l2_cache_size
  get_attribute(DeviceAttribute::L2_CACHE_SIZE)
end

#managed_memory? ⇒ Boolean

Returns Whether managed memory is supported.

Returns:

  • (Boolean) —

    Whether managed memory is supported



163
164
165
 
# File 'lib/nvruby/cuda/device.rb', line 163

def managed_memory?
  get_attribute(DeviceAttribute::MANAGED_MEMORY) != 0
end

#max_grid_size ⇒ Array<Integer>

Returns Max grid size [x, y, z].

Returns:

  • (Array<Integer>) —

    Max grid size [x, y, z]



119
120
121
122
123
124
125
 
# File 'lib/nvruby/cuda/device.rb', line 119

def max_grid_size
  [
    get_attribute(DeviceAttribute::MAX_GRID_DIM_X),
    get_attribute(DeviceAttribute::MAX_GRID_DIM_Y),
    get_attribute(DeviceAttribute::MAX_GRID_DIM_Z)
  ]
end

#max_threads_dim ⇒ Array<Integer>

Returns Max threads per dimension [x, y, z].

Returns:

  • (Array<Integer>) —

    Max threads per dimension [x, y, z]



110
111
112
113
114
115
116
 
# File 'lib/nvruby/cuda/device.rb', line 110

def max_threads_dim
  [
    get_attribute(DeviceAttribute::MAX_BLOCK_DIM_X),
    get_attribute(DeviceAttribute::MAX_BLOCK_DIM_Y),
    get_attribute(DeviceAttribute::MAX_BLOCK_DIM_Z)
  ]
end

#max_threads_per_block ⇒ Integer

Returns Max threads per block.

Returns:

  • (Integer) —

    Max threads per block



105
106
107
 
# File 'lib/nvruby/cuda/device.rb', line 105

def max_threads_per_block
  get_attribute(DeviceAttribute::MAX_THREADS_PER_BLOCK)
end

#memory_bus_width ⇒ Integer

Returns Memory bus width in bits.

Returns:

  • (Integer) —

    Memory bus width in bits



138
139
140
 
# File 'lib/nvruby/cuda/device.rb', line 138

def memory_bus_width
  get_attribute(DeviceAttribute::GLOBAL_MEMORY_BUS_WIDTH)
end

#memory_clock_rate ⇒ Integer

Returns Memory clock rate in kHz.

Returns:

  • (Integer) —

    Memory clock rate in kHz



133
134
135
 
# File 'lib/nvruby/cuda/device.rb', line 133

def memory_clock_rate
  get_attribute(DeviceAttribute::MEMORY_CLOCK_RATE)
end

#memory_info ⇒ Hash

Get available and total memory via RuntimeAPI (Fiddle hot path).

Returns:

  • (Hash) —

    total_bytes:



174
175
176
177
178
 
# File 'lib/nvruby/cuda/device.rb', line 174

def memory_info
  RuntimeAPI.ensure_loaded!
  RuntimeAPI.set_device(@index)
  RuntimeAPI.mem_get_info
end

#multiprocessor_count ⇒ Integer

Returns Number of streaming multiprocessors.

Returns:

  • (Integer) —

    Number of streaming multiprocessors



80
81
82
 
# File 'lib/nvruby/cuda/device.rb', line 80

def multiprocessor_count
  get_attribute(DeviceAttribute::MULTIPROCESSOR_COUNT)
end

#name ⇒ String

Returns Device name.

Returns:

  • (String) —

    Device name



61
62
63
64
65
 
# File 'lib/nvruby/cuda/device.rb', line 61

def name
  return @name if @name

  @name = properties[:name]
end

#properties ⇒ Hash

Get full device properties via DeviceProperties (FFI::Struct). Cached after first call.

Returns:

  • (Hash) —

    device property summary



188
189
190
 
# File 'lib/nvruby/cuda/device.rb', line 188

def properties
  @props_cache ||= DeviceProperties.summary(@index)
end

#reset! ⇒ void

This method returns an undefined value.

Reset this device.



208
209
210
211
 
# File 'lib/nvruby/cuda/device.rb', line 208

def reset!
  set_current!
  RuntimeAPI.device_reset
end

#set_current! ⇒ void

This method returns an undefined value.

Set this device as the current device.



194
195
196
197
 
# File 'lib/nvruby/cuda/device.rb', line 194

def set_current!
  RuntimeAPI.ensure_loaded!
  RuntimeAPI.set_device(@index)
end

#shared_memory_per_block ⇒ Integer

Returns Shared memory per block in bytes.

Returns:

  • (Integer) —

    Shared memory per block in bytes



75
76
77
 
# File 'lib/nvruby/cuda/device.rb', line 75

def shared_memory_per_block
  get_attribute(DeviceAttribute::MAX_SHARED_MEMORY_PER_BLOCK)
end

#synchronize ⇒ void

This method returns an undefined value.

Synchronize this device.



201
202
203
204
 
# File 'lib/nvruby/cuda/device.rb', line 201

def synchronize
  set_current!
  RuntimeAPI.device_synchronize
end

#to_s ⇒ String

Returns Human-readable device description.

Returns:

  • (String) —

    Human-readable device description



214
215
216
217
 
# File 'lib/nvruby/cuda/device.rb', line 214

def to_s
  mem_mb = total_memory / (1024 * 1024)
  "Device[#{@index}]: #{name} (CC #{compute_capability}, #{mem_mb} MB)"
end

#total_memory ⇒ Integer

Returns Total global memory in bytes.

Returns:

  • (Integer) —

    Total global memory in bytes



68
69
70
71
72
 
# File 'lib/nvruby/cuda/device.rb', line 68

def total_memory
  return @total_memory if @total_memory

  @total_memory = memory_info[:total_bytes]
end

#unified_addressing? ⇒ Boolean

Returns Whether unified addressing is supported.

Returns:

  • (Boolean) —

    Whether unified addressing is supported



158
159
160
 
# File 'lib/nvruby/cuda/device.rb', line 158

def unified_addressing?
  get_attribute(DeviceAttribute::UNIFIED_ADDRESSING) != 0
end

#warp_size ⇒ Integer

Returns Warp size (typically 32).

Returns:

  • (Integer) —

    Warp size (typically 32)



85
86
87
 
# File 'lib/nvruby/cuda/device.rb', line 85

def warp_size
  get_attribute(DeviceAttribute::WARP_SIZE)
end