projects / platform / upstream / openblas.git / commitdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: a5b1649)
use emms instead, add WIN guards
authorMatti Picus <matti.picus@gmail.com>
Mon, 12 Oct 2020 15:15:01 +0000 (18:15 +0300)
committerMatti Picus <matti.picus@gmail.com>
Mon, 12 Oct 2020 15:15:01 +0000 (18:15 +0300)
27 files changed:
kernel/x86_64/amax.S patch | blob | history
kernel/x86_64/asum.S patch | blob | history
kernel/x86_64/dot.S patch | blob | history
kernel/x86_64/iamax.S patch | blob | history
kernel/x86_64/izamax.S patch | blob | history
kernel/x86_64/nrm2.S patch | blob | history
kernel/x86_64/qconjg.S patch | blob | history
kernel/x86_64/qdot.S patch | blob | history
kernel/x86_64/qgemm_kernel_2x2.S patch | blob | history
kernel/x86_64/qgemv_n.S patch | blob | history
kernel/x86_64/qgemv_t.S patch | blob | history
kernel/x86_64/qtrsm_kernel_LN_2x2.S patch | blob | history
kernel/x86_64/qtrsm_kernel_LT_2x2.S patch | blob | history
kernel/x86_64/qtrsm_kernel_RT_2x2.S patch | blob | history
kernel/x86_64/sum.S patch | blob | history
kernel/x86_64/xdot.S patch | blob | history
kernel/x86_64/xgemm3m_kernel_2x2.S patch | blob | history
kernel/x86_64/xgemm_kernel_1x1.S patch | blob | history
kernel/x86_64/xgemv_n.S patch | blob | history
kernel/x86_64/xgemv_t.S patch | blob | history
kernel/x86_64/xtrsm_kernel_LT_1x1.S patch | blob | history
kernel/x86_64/zamax.S patch | blob | history
kernel/x86_64/zasum.S patch | blob | history
kernel/x86_64/zdot.S patch | blob | history
kernel/x86_64/znrm2.S patch | blob | history
kernel/x86_64/zscal.S patch | blob | history
kernel/x86_64/zsum.S patch | blob | history

diff --git a/kernel/x86_64/amax.S b/kernel/x86_64/amax.S
index 257147d..1498bb2 100644 (file)
--- a/kernel/x86_64/amax.S
+++ b/kernel/x86_64/amax.S
@@ -55,7 +55,9 @@
        PROLOGUE
        PROFCODE
        
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        salq    $BASE_SHIFT, INCX
 
diff --git a/kernel/x86_64/asum.S b/kernel/x86_64/asum.S
index 24f57dd..a2cbfd4 100644 (file)
--- a/kernel/x86_64/asum.S
+++ b/kernel/x86_64/asum.S
@@ -50,7 +50,10 @@
        PROLOGUE
        PROFCODE
        
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
+
        fldz
        testq   M, M
        jle     .L999
diff --git a/kernel/x86_64/dot.S b/kernel/x86_64/dot.S
index 2319885..a11d25e 100644 (file)
--- a/kernel/x86_64/dot.S
+++ b/kernel/x86_64/dot.S
@@ -49,7 +49,10 @@
 
        PROLOGUE
        PROFCODE
-       fninit
+
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        salq    $BASE_SHIFT, INCX
        salq    $BASE_SHIFT, INCY
diff --git a/kernel/x86_64/iamax.S b/kernel/x86_64/iamax.S
index 0c666d6..00999e2 100644 (file)
--- a/kernel/x86_64/iamax.S
+++ b/kernel/x86_64/iamax.S
@@ -59,7 +59,10 @@
 
        PROLOGUE
        PROFCODE
-       fninit
+
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        salq    $BASE_SHIFT, INCX
 
diff --git a/kernel/x86_64/izamax.S b/kernel/x86_64/izamax.S
index e450c2c..b24b2e6 100644 (file)
--- a/kernel/x86_64/izamax.S
+++ b/kernel/x86_64/izamax.S
@@ -59,7 +59,10 @@
 
        PROLOGUE
        PROFCODE
-       fninit
+
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        salq    $ZBASE_SHIFT, INCX
 
diff --git a/kernel/x86_64/nrm2.S b/kernel/x86_64/nrm2.S
index 548e3b7..b79ac2a 100644 (file)
--- a/kernel/x86_64/nrm2.S
+++ b/kernel/x86_64/nrm2.S
@@ -50,7 +50,10 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
+
        fldz
        testq   M, M
        jle     .L999
diff --git a/kernel/x86_64/qconjg.S b/kernel/x86_64/qconjg.S
index bab5418..823a15a 100644 (file)
--- a/kernel/x86_64/qconjg.S
+++ b/kernel/x86_64/qconjg.S
@@ -41,7 +41,10 @@
 
        PROLOGUE
        PROFCODE
-       fninit
+
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        fldz
        FLD     1 * SIZE(ARG1)
diff --git a/kernel/x86_64/qdot.S b/kernel/x86_64/qdot.S
index e7d3136..2243b6b 100644 (file)
--- a/kernel/x86_64/qdot.S
+++ b/kernel/x86_64/qdot.S
@@ -58,7 +58,9 @@
 
        PROLOGUE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        pushl   %edi
        pushl   %esi
diff --git a/kernel/x86_64/qgemm_kernel_2x2.S b/kernel/x86_64/qgemm_kernel_2x2.S
index 7b5e770..c11f3a9 100644 (file)
--- a/kernel/x86_64/qgemm_kernel_2x2.S
+++ b/kernel/x86_64/qgemm_kernel_2x2.S
@@ -74,7 +74,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/qgemv_n.S b/kernel/x86_64/qgemv_n.S
index 1b65b03..c9d345c 100644 (file)
--- a/kernel/x86_64/qgemv_n.S
+++ b/kernel/x86_64/qgemv_n.S
@@ -76,7 +76,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/qgemv_t.S b/kernel/x86_64/qgemv_t.S
index 00188c2..32372ff 100644 (file)
--- a/kernel/x86_64/qgemv_t.S
+++ b/kernel/x86_64/qgemv_t.S
@@ -75,7 +75,10 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
+
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
        movq    %rbp,  8(%rsp)
diff --git a/kernel/x86_64/qtrsm_kernel_LN_2x2.S b/kernel/x86_64/qtrsm_kernel_LN_2x2.S
index 030eff8..0a545fa 100644 (file)
--- a/kernel/x86_64/qtrsm_kernel_LN_2x2.S
+++ b/kernel/x86_64/qtrsm_kernel_LN_2x2.S
@@ -74,7 +74,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/qtrsm_kernel_LT_2x2.S b/kernel/x86_64/qtrsm_kernel_LT_2x2.S
index d86972c..16063fb 100644 (file)
--- a/kernel/x86_64/qtrsm_kernel_LT_2x2.S
+++ b/kernel/x86_64/qtrsm_kernel_LT_2x2.S
@@ -74,7 +74,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/qtrsm_kernel_RT_2x2.S b/kernel/x86_64/qtrsm_kernel_RT_2x2.S
index 2826a62..4c94ac0 100644 (file)
--- a/kernel/x86_64/qtrsm_kernel_RT_2x2.S
+++ b/kernel/x86_64/qtrsm_kernel_RT_2x2.S
@@ -74,8 +74,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
-
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/sum.S b/kernel/x86_64/sum.S
index 3d5fa7c..9f2cdc1 100644 (file)
--- a/kernel/x86_64/sum.S
+++ b/kernel/x86_64/sum.S
@@ -50,7 +50,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        fldz
        testq   M, M
diff --git a/kernel/x86_64/xdot.S b/kernel/x86_64/xdot.S
index ec89b79..c4b4734 100644 (file)
--- a/kernel/x86_64/xdot.S
+++ b/kernel/x86_64/xdot.S
@@ -59,7 +59,9 @@
 
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
 
 #define N      %ebx
diff --git a/kernel/x86_64/xgemm3m_kernel_2x2.S b/kernel/x86_64/xgemm3m_kernel_2x2.S
index e8da78d..1d0b23c 100644 (file)
--- a/kernel/x86_64/xgemm3m_kernel_2x2.S
+++ b/kernel/x86_64/xgemm3m_kernel_2x2.S
@@ -78,7 +78,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/xgemm_kernel_1x1.S b/kernel/x86_64/xgemm_kernel_1x1.S
index f04ab07..ee67d8d 100644 (file)
--- a/kernel/x86_64/xgemm_kernel_1x1.S
+++ b/kernel/x86_64/xgemm_kernel_1x1.S
@@ -97,7 +97,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/xgemv_n.S b/kernel/x86_64/xgemv_n.S
index 7d28c11..b66f28d 100644 (file)
--- a/kernel/x86_64/xgemv_n.S
+++ b/kernel/x86_64/xgemv_n.S
@@ -76,7 +76,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/xgemv_t.S b/kernel/x86_64/xgemv_t.S
index e796760..d6d3701 100644 (file)
--- a/kernel/x86_64/xgemv_t.S
+++ b/kernel/x86_64/xgemv_t.S
@@ -75,7 +75,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/xtrsm_kernel_LT_1x1.S b/kernel/x86_64/xtrsm_kernel_LT_1x1.S
index 54d4193..8752063 100644 (file)
--- a/kernel/x86_64/xtrsm_kernel_LT_1x1.S
+++ b/kernel/x86_64/xtrsm_kernel_LT_1x1.S
@@ -90,7 +90,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        subq    $STACKSIZE, %rsp
        movq    %rbx,  0(%rsp)
diff --git a/kernel/x86_64/zamax.S b/kernel/x86_64/zamax.S
index bfd8361..5cb2f60 100644 (file)
--- a/kernel/x86_64/zamax.S
+++ b/kernel/x86_64/zamax.S
@@ -55,7 +55,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        salq    $ZBASE_SHIFT, INCX
 
diff --git a/kernel/x86_64/zasum.S b/kernel/x86_64/zasum.S
index 9ea2aad..3460fce 100644 (file)
--- a/kernel/x86_64/zasum.S
+++ b/kernel/x86_64/zasum.S
@@ -50,7 +50,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        fldz
        testq   M, M
diff --git a/kernel/x86_64/zdot.S b/kernel/x86_64/zdot.S
index f7df919..87c08d7 100644 (file)
--- a/kernel/x86_64/zdot.S
+++ b/kernel/x86_64/zdot.S
@@ -54,9 +54,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
-
 #ifdef WINDOWS_ABI
+       emms
+
        movq    40(%rsp), INCY
 #endif
 
diff --git a/kernel/x86_64/znrm2.S b/kernel/x86_64/znrm2.S
index cb02a5a..0d2aa34 100644 (file)
--- a/kernel/x86_64/znrm2.S
+++ b/kernel/x86_64/znrm2.S
@@ -50,7 +50,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        fldz
        testq   M, M
diff --git a/kernel/x86_64/zscal.S b/kernel/x86_64/zscal.S
index 08c0831..5ed4c45 100644 (file)
--- a/kernel/x86_64/zscal.S
+++ b/kernel/x86_64/zscal.S
@@ -50,7 +50,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        salq    $ZBASE_SHIFT, INCX
 
diff --git a/kernel/x86_64/zsum.S b/kernel/x86_64/zsum.S
index 1c39048..aa02637 100644 (file)
--- a/kernel/x86_64/zsum.S
+++ b/kernel/x86_64/zsum.S
@@ -50,7 +50,9 @@
        PROLOGUE
        PROFCODE
 
-       fninit
+#ifdef WINDOWS_ABI
+       emms
+#endif
 
        fldz
        testq   M, M
Domain: Machine Learning / ML Framework;