Browse Source

Ref #380: lowered stack usage for haswell kernels

tags/v0.2.10.rc1^2
wernsaar 11 years ago
parent
commit
23203d52c1
4 changed files with 4 additions and 10 deletions
  1. +1
    -3
      kernel/x86_64/cgemm_kernel_8x2_haswell.S
  2. +1
    -1
      kernel/x86_64/dgemm_kernel_4x4_haswell.S
  3. +1
    -3
      kernel/x86_64/sgemm_kernel_16x4_haswell.S
  4. +1
    -3
      kernel/x86_64/zgemm_kernel_4x2_haswell.S

+ 1
- 3
kernel/x86_64/cgemm_kernel_8x2_haswell.S View File

@@ -93,8 +93,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif
#define L_BUFFER_SIZE 512*8*4
#define LB2_OFFSET 512*8*2
#define L_BUFFER_SIZE 8192
#define Ndiv6 24(%rsp)
#define Nmod6 32(%rsp)
@@ -105,7 +104,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define KK 72(%rsp)
#define KKK 80(%rsp)
#define BUFFER1 128(%rsp)
#define BUFFER2 LB2_OFFSET+128(%rsp)
#if defined(OS_WINDOWS)
#if L_BUFFER_SIZE > 16384


+ 1
- 1
kernel/x86_64/dgemm_kernel_4x4_haswell.S View File

@@ -85,7 +85,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#else
#define STACKSIZE 256
#define L_BUFFER_SIZE 128*8*12+4096
#define L_BUFFER_SIZE 128*8*12+512
#define OLD_A 40 + STACKSIZE(%rsp)
#define OLD_B 48 + STACKSIZE(%rsp)


+ 1
- 3
kernel/x86_64/sgemm_kernel_16x4_haswell.S View File

@@ -90,8 +90,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif
#define L_BUFFER_SIZE 512*8*4
#define LB2_OFFSET 512*8*2
#define L_BUFFER_SIZE 8192
#define Ndiv6 24(%rsp)
#define Nmod6 32(%rsp)
@@ -101,7 +100,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define KK 64(%rsp)
#define KKK 72(%rsp)
#define BUFFER1 128(%rsp)
#define BUFFER2 LB2_OFFSET+128(%rsp)
#if defined(OS_WINDOWS)
#if L_BUFFER_SIZE > 16384


+ 1
- 3
kernel/x86_64/zgemm_kernel_4x2_haswell.S View File

@@ -92,8 +92,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif
#define L_BUFFER_SIZE 512*8*4
#define LB2_OFFSET 512*8*2
#define L_BUFFER_SIZE 8192
#define Ndiv6 24(%rsp)
#define Nmod6 32(%rsp)
@@ -104,7 +103,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define KK 72(%rsp)
#define KKK 80(%rsp)
#define BUFFER1 128(%rsp)
#define BUFFER2 LB2_OFFSET+128(%rsp)
#if defined(OS_WINDOWS)
#if L_BUFFER_SIZE > 16384


Loading…
Cancel
Save