floraachy/OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix declaration of input arguments in the x86_64 microkernels for DOT and AXPY (#1965)

Browse Source 
* Tag operands 0 and 1 as both input and output

For #1964 (basically a continuation of coding problems first seen in #1292)
This commit is contained in:
Martin Kroeker
2019-01-17 23:20:32 +01:00
committed by GitHub
parent 24e697eadb
commit d5e6940253
37 changed files with 202 additions and 202 deletions
Download Patch File Download Diff File Expand all files Collapse all files
kernel/x86_64/caxpy_microk_bulldozer-2.c
+7 -7
View File
@@ -114,9 +114,9 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
@@ -180,10 +180,10 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
kernel/x86_64/caxpy_microk_haswell-2.c
+3 -3
View File
@@ -112,9 +112,9 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
kernel/x86_64/caxpy_microk_sandy-2.c
+4 -4
View File
@@ -95,10 +95,10 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
kernel/x86_64/caxpy_microk_steamroller-2.c
+7 -7
View File
@@ -113,10 +113,10 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
@@ -181,9 +181,9 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
kernel/x86_64/cdot_microk_bulldozer-2.c
+7 -7
View File
@@ -97,9 +97,9 @@ static void cdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
@@ -175,10 +175,10 @@ static void cdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/cdot_microk_haswell-2.c
+3 -3
View File
@@ -98,9 +98,9 @@ static void cdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/cdot_microk_sandy-2.c
+4 -4
View File
@@ -105,10 +105,10 @@ static void cdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/cdot_microk_steamroller-2.c
+7 -7
View File
@@ -97,9 +97,9 @@ static void cdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
@@ -175,10 +175,10 @@ static void cdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/daxpy_microk_bulldozer-2.c
+3 -3
View File
@@ -64,9 +64,9 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/daxpy_microk_haswell-2.c
+4 -4
View File
@@ -59,10 +59,10 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/daxpy_microk_nehalem-2.c
+3 -3
View File
@@ -73,9 +73,9 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/daxpy_microk_piledriver-2.c
+8 -8
View File
@@ -78,10 +78,10 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"subq $16, %1 \n\t" "subq $16, %1 \n\t"
"jnz 1b \n\t" "jnz 1b \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
@@ -140,10 +140,10 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"subq $16, %1 \n\t" "subq $16, %1 \n\t"
"jnz 1b \n\t" "jnz 1b \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/daxpy_microk_sandy-2.c
+4 -4
View File
@@ -99,10 +99,10 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/daxpy_microk_steamroller-2.c
+8 -8
View File
@@ -78,10 +78,10 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"subq $16, %1 \n\t" "subq $16, %1 \n\t"
"jnz 1b \n\t" "jnz 1b \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
@@ -140,10 +140,10 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"subq $16, %1 \n\t" "subq $16, %1 \n\t"
"jnz 1b \n\t" "jnz 1b \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/ddot_microk_bulldozer-2.c
+4 -4
View File
@@ -65,10 +65,10 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovsd %%xmm4, (%4) \n\t" "vmovsd %%xmm4, (%4) \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/ddot_microk_haswell-2.c
+3 -3
View File
@@ -77,9 +77,9 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/ddot_microk_nehalem-2.c
+4 -4
View File
@@ -75,10 +75,10 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"movsd %%xmm4, (%4) \n\t" "movsd %%xmm4, (%4) \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/ddot_microk_piledriver-2.c
+8 -8
View File
@@ -81,10 +81,10 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovsd %%xmm4, (%4) \n\t" "vmovsd %%xmm4, (%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
@@ -145,10 +145,10 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovsd %%xmm4, (%4) \n\t" "vmovsd %%xmm4, (%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/ddot_microk_sandy-2.c
+4 -4
View File
@@ -81,10 +81,10 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovsd %%xmm4, (%4) \n\t" "vmovsd %%xmm4, (%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/ddot_microk_steamroller-2.c
+4 -4
View File
@@ -78,10 +78,10 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovsd %%xmm4, (%4) \n\t" "vmovsd %%xmm4, (%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/saxpy_microk_haswell-2.c
+4 -4
View File
@@ -59,10 +59,10 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/saxpy_microk_nehalem-2.c
+3 -3
View File
@@ -73,9 +73,9 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/saxpy_microk_piledriver-2.c
+8 -8
View File
@@ -78,10 +78,10 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
@@ -139,10 +139,10 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/saxpy_microk_sandy-2.c
+4 -4
View File
@@ -99,10 +99,10 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha) // 4 "r" (alpha) // 4
kernel/x86_64/sdot_microk_bulldozer-2.c
+4 -4
View File
@@ -66,10 +66,10 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovss %%xmm4, (%4) \n\t" "vmovss %%xmm4, (%4) \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/sdot_microk_haswell-2.c
+4 -4
View File
@@ -79,10 +79,10 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovss %%xmm4, (%4) \n\t" "vmovss %%xmm4, (%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/sdot_microk_nehalem-2.c
+4 -4
View File
@@ -75,10 +75,10 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"movss %%xmm4, (%4) \n\t" "movss %%xmm4, (%4) \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/sdot_microk_sandy-2.c
+4 -4
View File
@@ -82,10 +82,10 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovss %%xmm4, (%4) \n\t" "vmovss %%xmm4, (%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/sdot_microk_steamroller-2.c
+8 -8
View File
@@ -80,10 +80,10 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovss %%xmm4, (%4) \n\t" "vmovss %%xmm4, (%4) \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
@@ -143,10 +143,10 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovss %%xmm4, (%4) \n\t" "vmovss %%xmm4, (%4) \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/zaxpy_microk_bulldozer-2.c
+8 -8
View File
@@ -113,10 +113,10 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
@@ -180,10 +180,10 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
kernel/x86_64/zaxpy_microk_haswell-2.c
+4 -4
View File
@@ -111,10 +111,10 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
kernel/x86_64/zaxpy_microk_sandy-2.c
+8 -8
View File
@@ -99,10 +99,10 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
@@ -176,10 +176,10 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
kernel/x86_64/zaxpy_microk_steamroller-2.c
+8 -8
View File
@@ -113,10 +113,10 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
@@ -180,10 +180,10 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
"jnz 1b \n\t" "jnz 1b \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (alpha), // 4 "r" (alpha), // 4
kernel/x86_64/zdot_microk_bulldozer-2.c
+8 -8
View File
@@ -96,10 +96,10 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
@@ -175,10 +175,10 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/zdot_microk_haswell-2.c
+8 -8
View File
@@ -101,10 +101,10 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
@@ -186,10 +186,10 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/zdot_microk_sandy-2.c
+8 -8
View File
@@ -107,10 +107,10 @@ if ( n < 1280 )
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
@@ -199,10 +199,10 @@ if ( n < 1280 )
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
kernel/x86_64/zdot_microk_steamroller-2.c
+8 -8
View File
@@ -95,10 +95,10 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4
@@ -172,10 +172,10 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
"vmovups %%xmm4, 16(%4) \n\t" "vmovups %%xmm4, 16(%4) \n\t"
"vzeroupper \n\t" "vzeroupper \n\t"
: :
: "+r" (i), // 0
"r" (i), // 0 "+r" (n) // 1
"r" (n), // 1 :
"r" (x), // 2 "r" (x), // 2
"r" (y), // 3 "r" (y), // 3
"r" (dot) // 4 "r" (dot) // 4