undesired BRAM replication

I am using a single work item kernel to do matrix multiplication, and my BRAM usage explored (estimated 100+% BRAM usage while only 16% for DSP).

==============================================================================================

#define MAT_A_ROWS 128

#define MAT_A_COLS 64

#define MAT_B_ROWS MAT_A_COLS

#define MAT_B_COLS 128

#define BLOCK_SIZE 16

__kernel

__attribute__((task))

void matrix_mult(__global float *restrict matA, __global float *restrict matB, __global float *restrict matC, ) {

__local float __attribute__((num_banks(BLOCK_SIZE), bandwidth(1))) A[BLOCK_SIZE][BLOCK_SIZE];

__local float __attribute__((num_banks(BLOCK_SIZE), bandwidth(BLOCK_SIZE))) B[BLOCK_SIZE][BLOCK_SIZE];

__local C[BLOCK_SIZE][BLOCK_SIZE];

for (int k = 0; k < MAT_A_COLS / BLOCK_SIZE; k++) {

for(int i = 0; i < BLOCK_SIZE; i++) {

for(int j = 0; j < BLOCK_SIZE; j++) {

A[j] = mata[.....];

}

}

for(int i = 0; i < block_size; i++) {

for(int j = 0; j < block_size; j++) {

b[j] = matB[.....];

}

}

for(int i = 0; i < BLOCK_SIZE; i++){

for(int j = 0; j < BLOCK_SIZE; j++) {

float running_sum = 0;

for(int k = 0; k < BLOCK_SIZE; k ++) {

running_sum += A[k] + b[k][j];

}

c[j] += running_sum;

}

}

}

......

//write to C to matC

}

======================================================

According to the report, there 8 threads being pipelined from loop "for(int k = 0; k < MAT_A_COLS / BLOCK_SIZE; k++) ", thus my A and B are replicated 7 times. Is it possible to prevent the memory being replicated?

Any advice would be greatly appreciated!

Lancer Chiang