Ref #380: lowered stack usage for haswell kernels

diff --git a/kernel/x86_64/cgemm_kernel_8x2_haswell.S b/kernel/x86_64/cgemm_kernel_8x2_haswell.S
index 38c864c..285a9f0 100644 (file)
--- 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.
 \r
 #endif\r
 \r
-#define L_BUFFER_SIZE 512*8*4\r
-#define LB2_OFFSET    512*8*2\r
+#define L_BUFFER_SIZE 8192\r
 \r
 #define Ndiv6   24(%rsp)\r
 #define Nmod6   32(%rsp)\r
@@ -105,7 +104,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define KK       72(%rsp)\r
 #define KKK      80(%rsp)\r
 #define BUFFER1                   128(%rsp)\r
-#define BUFFER2        LB2_OFFSET+128(%rsp)\r
 \r
 #if defined(OS_WINDOWS)\r
 #if   L_BUFFER_SIZE > 16384\r

diff --git a/kernel/x86_64/dgemm_kernel_4x4_haswell.S b/kernel/x86_64/dgemm_kernel_4x4_haswell.S
index 4964d17..a49a51e 100644 (file)
--- 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\r
 \r
 #define STACKSIZE 256\r
-#define L_BUFFER_SIZE 128*8*12+4096\r
+#define L_BUFFER_SIZE 128*8*12+512\r
 \r
 #define OLD_A          40 + STACKSIZE(%rsp)\r
 #define OLD_B          48 + STACKSIZE(%rsp)\r

diff --git a/kernel/x86_64/sgemm_kernel_16x4_haswell.S b/kernel/x86_64/sgemm_kernel_16x4_haswell.S
index 6c3cda0..1f9f886 100644 (file)
--- 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.
 \r
 #endif\r
 \r
-#define L_BUFFER_SIZE 512*8*4\r
-#define LB2_OFFSET    512*8*2\r
+#define L_BUFFER_SIZE 8192\r
 \r
 #define Ndiv6   24(%rsp)\r
 #define Nmod6   32(%rsp)\r
@@ -101,7 +100,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define KK      64(%rsp)\r
 #define KKK     72(%rsp)\r
 #define BUFFER1                   128(%rsp)\r
-#define BUFFER2        LB2_OFFSET+128(%rsp)\r
 \r
 #if defined(OS_WINDOWS)\r
 #if   L_BUFFER_SIZE > 16384\r

diff --git a/kernel/x86_64/zgemm_kernel_4x2_haswell.S b/kernel/x86_64/zgemm_kernel_4x2_haswell.S
index 515939d..289fbbf 100644 (file)
--- 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.
 \r
 #endif\r
 \r
-#define L_BUFFER_SIZE 512*8*4\r
-#define LB2_OFFSET    512*8*2\r
+#define L_BUFFER_SIZE 8192\r
 \r
 #define Ndiv6   24(%rsp)\r
 #define Nmod6   32(%rsp)\r
@@ -104,7 +103,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define KK       72(%rsp)\r
 #define KKK      80(%rsp)\r
 #define BUFFER1                   128(%rsp)\r
-#define BUFFER2        LB2_OFFSET+128(%rsp)\r
 \r
 #if defined(OS_WINDOWS)\r
 #if   L_BUFFER_SIZE > 16384\r