|
|
|
@@ -1866,7 +1866,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
|
|
|
SAVE4x12
|
|
|
|
|
|
|
|
/* here for the prefetch of next b source block */
|
|
|
|
/* the increment should be proportional to the ratio of GEMM_Q/GEMM_P */
|
|
|
|
/* the increment should be proportional to GEMM_Q/GEMM_P */
|
|
|
|
|
|
|
|
salq $3, K
|
|
|
|
#ifdef WINDOWS_ABI /* GEMM_P == GEMM_Q * 4 */
|
|
|
|
@@ -2184,19 +2184,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
|
|
|
SAVE4x12
|
|
|
|
|
|
|
|
/* here for the prefetch of next b source block */
|
|
|
|
/* the increment should be proportional to the ratio of GEMM_Q/GEMM_P */
|
|
|
|
/* the increment should be proportional to GEMM_Q/GEMM_P */
|
|
|
|
|
|
|
|
salq $3, K
|
|
|
|
#ifdef WINDOWS_ABI /* GEMM_P == GEMM_Q * 4 */
|
|
|
|
prefetcht2 (B)
|
|
|
|
prefetcht2 (B, K, 8)
|
|
|
|
addq $64, B
|
|
|
|
addq $64, B /* increment */
|
|
|
|
#else /* GEMM_P == GEMM_Q * 2 under linux x86_64 */
|
|
|
|
prefetcht2 (B)
|
|
|
|
prefetcht2 (B, K, 8)
|
|
|
|
prefetcht2 64(B)
|
|
|
|
prefetcht2 64(B, K, 8)
|
|
|
|
addq $128, B
|
|
|
|
addq $128, B /* increment */
|
|
|
|
#endif
|
|
|
|
sarq $3, K
|
|
|
|
|
|
|
|
|