CUDA programming

Function declaration:

• 　　__global__ void KernelFunc()　　Executed:device　　Callable:host
• 　　__device__ float DeviceFunc()　　......:device　　　　......:device
• 　　__host__ float HostFunc()　　　　......:host　　　　　......:host

　　__global__:　

　　　　　　The return value must be void

　　__device__:

　　　　　　It used to be inline by default, now there are some changes.

　　Global and Device functions:

1. 1. Use recursion as little as possible
   2. don't use static variables
   3. use malloc sparingly
   4. Be careful with function calls implemented through pointers

• Vector data type:
  • 　　char[1-4],uchar[1-4]
  • 　　short[1-4],ushort[1-4]
  • 　　int[1-4],uint[1-4]
  • 　　long[1-4],ulong[1-4]
  • 　　longlong[1-4], ulonglong[1-4]
  • 　　floa [1-4]
  • 　　double1,double2
• vector data type
  • 　　Applicable to both host and device code, constructed by the function make_<type name>

int2 i2 = make_int2(1, 2);

float4 f4 = make_float4(1.0f, 2.0f, 3.0f, 4.0f);

• • Access via .x, .y, .z, and .w

int2 i2 = make_int2(1, 2);

int x = i2.x;

int y = i2.y;

• math function
  • 　　Partial function list
    • sqrt,rsqrt
    • exp,log
    • sin, cos, tan, sincos
    • asin,acos,atan2
    • trunc,ceil,floor
  • Intrinsic function built-in function
    • Only for Device
    • faster but less accurate
    • Prefixed with __, for example:
      • __exp,__log,__sin,__pow,......
• thread synchronization
  • Fast threads can be synchronized
  • Call __syncthreads to create a barrier
  • Each thread waits at the call site for all threads in the block to execute to this place, and then all threads continue to execute subsequent instructions

Mds[i] = Md[j]

__syncthreads()

func(Mds[i], Mds[i+1])

• thread scheduling
  • Wrap a set of threads within a block　　
    • G80/GT200 - 32 threads
    • run on the same SM
    • The basic unit of thread scheduling
    • threadIdx value is continuous
    • An implementation detail - theoretically
      • WrapSize
• memory model
  • Register Registers
    • Dedicated to each thread
    • Fast, on-chip, readable and writable
    • What will be the result of increasing Kernal's register usage?
  • Register Register
    • per SM
      • up to 768threads
      • 8K registers
  • Local Memory
    • stored in global memory
      • scope per thread
    • Used to store arrays of automatic variables
      • Access via constant index
  • Shared Memory
    • each block
    • Fast, on-chip, readable and writable
    • Full speed random access
    • Each SM includes 8 blocks, 16KB shared memory
  • Global Memory
    • Long delay (100 cycles)
    • off-chip, readable and writable
    • Random access affects performance
    • Host host class read and write
    • GT200
      • Bandwidth: 150GB/s
      • Capacity: 4GB
    • G80 - 86.4GB/s
  • Constant Memory
    • Short latency, high bandwidth, read-only when all threads access the same location
    • Stored in gloabl memory but cached
    • Host host can read and write
    • Capacity: 64KB

1. Globl and constant variables
   • Host can be accessed through the following functions:
     • cudaGetSymbolAddress()
     • cudaGetSymbolSize()
     • cudaMemcpyToSymbol()
     • cudaMemcpyFromSymbol()
   • constants variables must be declared outside the function