From 23203d52c1e741994d6c2aabf4714cdf38d75d37 Mon Sep 17 00:00:00 2001 From: wernsaar Date: Thu, 19 Jun 2014 14:31:52 +0200 Subject: [PATCH] Ref #380: lowered stack usage for haswell kernels --- kernel/x86_64/cgemm_kernel_8x2_haswell.S | 4 +--- kernel/x86_64/dgemm_kernel_4x4_haswell.S | 2 +- kernel/x86_64/sgemm_kernel_16x4_haswell.S | 4 +--- kernel/x86_64/zgemm_kernel_4x2_haswell.S | 4 +--- 4 files changed, 4 insertions(+), 10 deletions(-) diff --git a/kernel/x86_64/cgemm_kernel_8x2_haswell.S b/kernel/x86_64/cgemm_kernel_8x2_haswell.S index 38c864ce5..285a9f09d 100644 --- a/kernel/x86_64/cgemm_kernel_8x2_haswell.S +++ b/kernel/x86_64/cgemm_kernel_8x2_haswell.S @@ -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 diff --git a/kernel/x86_64/dgemm_kernel_4x4_haswell.S b/kernel/x86_64/dgemm_kernel_4x4_haswell.S index 4964d17e6..a49a51ee9 100644 --- a/kernel/x86_64/dgemm_kernel_4x4_haswell.S +++ b/kernel/x86_64/dgemm_kernel_4x4_haswell.S @@ -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) diff --git a/kernel/x86_64/sgemm_kernel_16x4_haswell.S b/kernel/x86_64/sgemm_kernel_16x4_haswell.S index 6c3cda022..1f9f88657 100644 --- a/kernel/x86_64/sgemm_kernel_16x4_haswell.S +++ b/kernel/x86_64/sgemm_kernel_16x4_haswell.S @@ -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 diff --git a/kernel/x86_64/zgemm_kernel_4x2_haswell.S b/kernel/x86_64/zgemm_kernel_4x2_haswell.S index 515939df6..289fbbfe9 100644 --- a/kernel/x86_64/zgemm_kernel_4x2_haswell.S +++ b/kernel/x86_64/zgemm_kernel_4x2_haswell.S @@ -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