CUDA syntax

Source code is in .cu files, which contain mixture of host (CPU) and device (GPU) code.

Declaring functions

`__global__`	declares kernel, which is called on host and executed on device
`__device__`	declares device function, which is called and executed on device
`__host__`	declares host function, which is called and executed on host
`__noinline__`	to avoid inlining
`__forceinline__`	to force inlining

Declaring variables

`__device__`	declares device variable in global memory, accessible from all threads, with lifetime of application
`__constant__`	declares device variable in constant memory, accessible from all threads, with lifetime of application
`__shared__`	declares device varibale in block's shared memory, accessible from all threads within a block, with lifetime of block
`__restrict__`	standard C definition that pointers are not aliased

Types

Most routines return an error code of type cudaError_t.

Vector types

char1, uchar1, short1, ushort1, int1, uint1, long1, ulong1, float1
char2, uchar2, short2, ushort2, int2, uint2, long2, ulong2, float2
char3, uchar3, short3, ushort3, int3, uint3, long3, ulong3, float3
char4, uchar4, short4, ushort4, int4, uint4, long4, ulong4, float4

longlong1, ulonglong1, double1
longlong2, ulonglong2, double2

dim3

Components are accessible as variable.x, variable.y, variable.z, variable.w.
Constructor is make_<type>( x, ... ), for example:

float2 xx = make_float2( 1., 2. );

dim3 can take 1, 2, or 3 argumetns:

dim3 blocks1D( 5       );
dim3 blocks2D( 5, 5    );
dim3 blocks3D( 5, 5, 5 );

Pre-defined variables

`dim3 gridDim`	dimensions of grid
`dim3 blockDim`	dimensions of block
`uint3 blockIdx`	block index within grid
`uint3 threadIdx`	thread index within block
`int warpSize`	number of threads in warp

Kernel invocation

__global__ void kernel( ... ) { ... }

dim3 blocks( nx, ny, nz );           // cuda 1.x has 1D and 2D grids, cuda 2.x adds 3D grids
dim3 threadsPerBlock( mx, my, mz );  // cuda 1.x has 1D, 2D, and 3D blocks

kernel<<< blocks, threadsPerBlock >>>( ... );

Thread management

`__threadfence_block();`	wait until memory accesses are visible to block
`__threadfence();`	wait until memory accesses are visible to block and device
`__threadfence_system();`	wait until memory accesses are visible to block and device and host (2.x)
`__syncthreads();`	wait until all threads reach sync

Memory management

__device__ float* pointer;
cudaMalloc( (void**) &pointer, size );
cudaFree( pointer );

__constant__ float dev_data[n];
float host_data[n];
cudaMemcpyToSymbol  ( dev_data,  host_data, sizeof(host_data) );  // dev_data  = host_data
cudaMemcpyFromSymbol( host_data, dev_data,  sizeof(host_data) );  // host_data = dev_data

// direction is one of cudaMemcpyHostToDevice or cudaMemcpyDeviceToHost
cudaMemcpy     ( dst_pointer, src_pointer, size, direction );
cudaMemcpyAsync( dst_pointer, src_pointer, size, direction, stream );

// using column-wise notation
// (the CUDA docs describe it for images; a “row” there equals a matrix column)
// _bytes indicates arguments that must be specified in bytes
cudaMemcpy2D     ( A_dst, lda_bytes, B_src, ldb_bytes, m_bytes, n, direction );
cudaMemcpy2DAsync( A_dst, lda_bytes, B_src, ldb_bytes, m_bytes, n, direction, stream );

// cublas makes copies easier for matrices, e.g., less use of sizeof
// copy x => y
cublasSetVector     ( n, elemSize, x_src_host, incx, y_dst_dev,  incy );
cublasGetVector     ( n, elemSize, x_src_dev,  incx, y_dst_host, incy );
cublasSetVectorAsync( n, elemSize, x_src_host, incx, y_dst_dev,  incy, stream );
cublasGetVectorAsync( n, elemSize, x_src_dev,  incx, y_dst_host, incy, stream );

// copy A => B
cublasSetMatrix     ( rows, cols, elemSize, A_src_host, lda, B_dst_dev,  ldb );
cublasGetMatrix     ( rows, cols, elemSize, A_src_dev,  lda, B_dst_host, ldb );
cublasSetMatrixAsync( rows, cols, elemSize, A_src_host, lda, B_dst_dev,  ldb, stream );
cublasGetMatrixAsync( rows, cols, elemSize, A_src_dev,  lda, B_dst_host, ldb, stream );

Also, malloc and free work inside a kernel (2.x), but memory allocated in a kernel must be deallocated in a kernel (not the host). It can be freed in a different kernel, though.

Atomic functions

old = atomicAdd ( &addr, value );  // old = *addr;  *addr += value
old = atomicSub ( &addr, value );  // old = *addr;  *addr –= value
old = atomicExch( &addr, value );  // old = *addr;  *addr  = value

old = atomicMin ( &addr, value );  // old = *addr;  *addr = min( old, value )
old = atomicMax ( &addr, value );  // old = *addr;  *addr = max( old, value )

// increment up to value, then reset to 0  
// decrement down to 0, then reset to value
old = atomicInc ( &addr, value );  // old = *addr;  *addr = ((old >= value) ? 0 : old+1 )
old = atomicDec ( &addr, value );  // old = *addr;  *addr = ((old == 0) or (old > val) ? val : old–1 )

old = atomicAnd ( &addr, value );  // old = *addr;  *addr &= value
old = atomicOr  ( &addr, value );  // old = *addr;  *addr |= value
old = atomicXor ( &addr, value );  // old = *addr;  *addr ^= value

// compare-and-store
old = atomicCAS ( &addr, compare, value );  // old = *addr;  *addr = ((old == compare) ? value : old)

Warp vote

int __all   ( predicate );
int __any   ( predicate );
int __ballot( predicate );  // nth thread sets nth bit to predicate

Timer

wall clock cycle counter

clock_t clock();

Texture

can also return float2 or float4, depending on texRef.

// integer index
float tex1Dfetch( texRef, ix );

// float index
float tex1D( texRef, x       );
float tex2D( texRef, x, y    );
float tex3D( texRef, x, y, z );

float tex1DLayered( texRef, x    );
float tex2DLayered( texRef, x, y );

Low-level Driver API

#include <cuda.h>

CUdevice dev;
CUdevprop properties;
char name[n];
int major, minor;
size_t bytes;

cuInit( 0 );  // takes flags for future use
cuDeviceGetCount         ( &cnt );
cuDeviceGet              ( &dev, index );
cuDeviceGetName          ( name, sizeof(name), dev );
cuDeviceComputeCapability( &major, &minor,     dev );
cuDeviceTotalMem         ( &bytes,             dev );
cuDeviceGetProperties    ( &properties,        dev );  // max threads, etc.

cuBLAS

Matrices are column-major. Indices are 1-based; this affects result of i<t>amax and i<t>amin.

#include <cublas_v2.h>

cublasHandle_t handle;
cudaStream_t   stream;

cublasCreate( &handle );
cublasDestroy( handle );
cublasGetVersion( handle, &version );
cublasSetStream( handle,  stream );
cublasGetStream( handle, &stream );
cublasSetPointerMode( handle,  mode );
cublasGetPointerMode( handle, &mode );

Constants

argument	constants	description (Fortran letter)
`trans`	`CUBLAS_OP_N`	non-transposed ('N')
	`CUBLAS_OP_T`	transposed ('T')
	`CUBLAS_OP_C`	conjugate transposed ('C')

`uplo`	`CUBLAS_FILL_MODE_LOWER`	lower part filled ('L')
	`CUBLAS_FILL_MODE_UPPER`	upper part filled ('U')

`side`	`CUBLAS_SIDE_LEFT`	matrix on left ('L')
	`CUBLAS_SIDE_RIGHT`	matrix on right ('R')

`mode`	`CUBLAS_POINTER_MODE_HOST`	alpha and beta scalars passed on host
	`CUBLAS_POINTER_MODE_DEVICE`	alpha and beta scalars passed on device

BLAS functions have cublas prefix and first letter of usual BLAS function name is capitalized. Arguments are the same as standard BLAS, with these exceptions:

All functions add handle as first argument.
All functions return cublasStatus_t error code.
Constants alpha and beta are passed by pointer. All other scalars (n, incx, etc.) are bassed by value.
Functions that return a value, such as ddot, add result as last argument, and save value to result.
Constants are given in table above, instead of using characters.

Examples:

cublasDdot ( handle, n, x, incx, y, incy, &result );  // result = ddot( n, x, incx, y, incy );
cublasDaxpy( handle, n, &alpha, x, incx, y, incy );   // daxpy( n, alpha, x, incx, y, incy );

Compiler

nvcc, often found in /usr/local/cuda/bin

Defines __CUDACC__

Flags common with cc

Short flag	Long flag	Output or Description
`-c`	`--compile`	.o object file
`-E`	`--preprocess`	on standard output
`-M`	`--generate-dependencies`	on standard output
`-o file`	`--output-file file`
`-I directory`	`--include-path directory`	header search path
`-L directory`	`--library-path directory`	library search path
`-l lib`	`--library lib`	link with library
`-lib`		generate library
`-shared`		generate shared library
`-pg`	`--profile`	for gprof
`-g level`	`--debug level`
`-G`	`--device-debug`
`-O level`	`--optimize level`

Undocumented (but in sample makefiles)
`-m32`		compile 32-bit i386 host CPU code
`-m64`		compile 64-bit x86_64 host CPU code

Flags specific to nvcc

`-v`	list compilation commands as they are executed
`-dryrun`	list compilation commands, without executing
`-keep`	saves intermediate files (e.g., pre-processed) for debugging
`-clean`	removes output files (with same exact compiler options)
`-arch=<compute_xy>`	generate PTX for capability x.y
`-code=<sm_xy>`	generate binary for capability x.y, by default same as `-arch`
`-gencode arch=...,code=...`	same as `-arch` and `-code`, but may be repeated

Argumenents for `-arch` and `-code`

It makes most sense (to me) to give -arch a virtual architecture and -code a real architecture, though both flags accept both virtual and real architectures (at times).

	Virtual architecture	Real architecture	Features
Tesla	`compute_10`	`sm_10`	Basic features
	`compute_11`	`sm_11`	+ atomic memory ops on global memory
	`compute_12`	`sm_12`	+ atomic memory ops on shared memory + vote instructions
	`compute_13`	`sm_13`	+ double precision
Fermi	`compute_20`	`sm_20`	+ Fermi

Some hardware constraints

	1.x	2.x
max x- or y-dimension of block	512	1024
max z-dimension of block	64	64
max threads per block	512	1024
warp size	32	32
max blocks per MP	8	8
max warps per MP	32	48
max threads per MP	1024	1536
max 32-bit registers per MP	16k	32k
max shared memory per MP	16 KB	48 KB
shared memory banks	16	32
local memory per thread	16 KB	512 KB
const memory	64 KB	64 KB
const cache	8 KB	8 KB
texture cache	8 KB	8 KB