Cuda Matrix Multiplication

# Include "cuda_runtime.h"
# Include "device_launch_parameters.h"

# Include <stdio. h>
# Include <stdlib. h>
# Include "cublas_v2.h"

# Define block_size 16

/***************/

Using the built-in function API of cublas, cublassgemm

Cudaerror_t multiwithcublase (float * C, float * a, float * B, unsigned int ah, unsigned int aw, unsigned int BH, unsigned int BW );

{

..................

Cublashandle_t handle;
Cublasstatus_t ret;
Ret = cublascreate (& handle );

Const float alpha = 1.0f;
Const float Beta = 0.0f;

Ret = cublassgemm (handle, cublas_op_n, cublas_op_n, ah, BW, aw, & Alpha, gpu_a, ah, gpu_ B, BH, & beta, gpu_c, AH );

..................

}

/***************/

Auxiliary with shared memory

_ Global _ void multikernel (float * C, float * a, float * B, unsigned int aw, unsigned int BW)

{
Int xblock = blockidx. X;
Int yblock = blockidx. Y;

Int xthread = threadidx. X;
Int ythread = threadidx. Y;

Unsigned int awidth = aw;
Unsigned int bwidth = bw;

Float cvalue = 0;

For (INT I = 0; I <awidth/block_size; ++ I)
{
_ Shared _ int asub [block_size] [block_size];
_ Shared _ int bsub [block_size] [block_size];

Asub [ythread] [xthread] = A [(yblock * blockdim. Y + ythread) * awidth + I * blockdim. x + xthread];
Bsub [ythread] [xthread] = B [(I * blockdim. Y + ythread) * bwidth + xblock * blockdim. x + xthread];

_ Syncthreads ();

For (int e = 0; e <block_size; ++ E)
{
Cvalue + = asub [ythread] [e] * bsub [E] [xthread];
}
_ Syncthreads ();
}

Int cIndex = (yblock * blockdim. Y + ythread) * bwidth + xblock * blockdim. x + xthread;
C [cIndex] = cvalue;
}

 

/***************/

Use shared memory to assist and enable Loop

_ Global _ void multikernel_noloop (float * C, float * a, float * B, unsigned int aw, unsigned int BW)
{
Int xblock = blockidx. X;
Int yblock = blockidx. Y;

Int xthread = threadidx. X;
Int ythread = threadidx. Y;

Unsigned int awidth = aw;
Unsigned int bwidth = bw;

Float cvalue = 0;

For (INT I = 0; I <awidth/block_size; ++ I)
{
_ Shared _ int asub [block_size] [block_size];
_ Shared _ int bsub [block_size] [block_size];

Asub [ythread] [xthread] = A [(yblock * blockdim. Y + ythread) * awidth + I * blockdim. x + xthread];
Bsub [ythread] [xthread] = B [(I * blockdim. Y + ythread) * bwidth + xblock * blockdim. x + xthread];

_ Syncthreads ();

Cvalue + = asub [ythread] [0] * bsub [0] [xthread] + asub [ythread] [1] * bsub [1] [xthread] + \
Asub [ythread] [2] * bsub [2] [xthread] + asub [ythread] [3] * bsub [3] [xthread] + \
Asub [ythread] [4] * bsub [4] [xthread] + asub [ythread] [5] * bsub [5] [xthread] + \
Asub [ythread] [6] * bsub [6] [xthread] + asub [ythread] [7] * bsub [7] [xthread] + \
Asub [ythread] [8] * bsub [8] [xthread] + asub [ythread] [9] * bsub [9] [xthread] + \
Asub [ythread] [10] * bsub [10] [xthread] + asub [ythread] [11] * bsub [11] [xthread] + \
Asub [ythread] [12] * bsub [12] [xthread] + asub [ythread] [13] * bsub [13] [xthread] + \
Asub [ythread] [14] * bsub [14] [xthread] + asub [ythread] [15] * bsub [15] [xthread];

_ Syncthreads ();
}

Int cIndex = (yblock * blockdim. Y + ythread) * bwidth + xblock * blockdim. x + xthread;
C [cIndex] = cvalue;
}

Matrix size: 320*320:

512*512:

PS machines are too bad.