projects
/
mesa.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
i965: Validate "General Restrictions on Regioning Parameters"
[mesa.git]
/
src
/
mesa
/
x86-64
/
xform4.S
diff --git
a/src/mesa/x86-64/xform4.S
b/src/mesa/x86-64/xform4.S
index 3f9c9d56ab7dcfdeae89aaa074b15fbd6be3246c..b0aca19c8b0963d00e22c4e42d608e97fc2e4c4e 100644
(file)
--- a/
src/mesa/x86-64/xform4.S
+++ b/
src/mesa/x86-64/xform4.S
@@
-1,6
+1,5
@@
/*
* Mesa 3-D graphics library
/*
* Mesa 3-D graphics library
- * Version: 7.1
*
* Copyright (C) 1999-2007 Brian Paul All Rights Reserved.
*
*
* Copyright (C) 1999-2007 Brian Paul All Rights Reserved.
*
@@
-17,9
+16,10
@@
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
- * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
- * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
- * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
*/
#ifdef USE_X86_64_ASM
*/
#ifdef USE_X86_64_ASM
@@
-29,8
+29,25
@@
.text
.align 16
.text
.align 16
+.globl _mesa_x86_64_cpuid
+.hidden _mesa_x86_64_cpuid
+_mesa_x86_64_cpuid:
+ pushq %rbx
+ movl (%rdi), %eax
+ movl 8(%rdi), %ecx
+
+ cpuid
+
+ movl %ebx, 4(%rdi)
+ movl %eax, (%rdi)
+ movl %ecx, 8(%rdi)
+ movl %edx, 12(%rdi)
+ popq %rbx
+ ret
+.align 16
.globl _mesa_x86_64_transform_points4_general
.globl _mesa_x86_64_transform_points4_general
+.hidden _mesa_x86_64_transform_points4_general
_mesa_x86_64_transform_points4_general:
/*
* rdi = dest
_mesa_x86_64_transform_points4_general:
/*
* rdi = dest
@@
-38,7
+55,7
@@
_mesa_x86_64_transform_points4_general:
* rdx = source
*/
movl V4F_COUNT(%rdx), %ecx /* count */
* rdx = source
*/
movl V4F_COUNT(%rdx), %ecx /* count */
- movz
x V4F_STRIDE(%rdx), %eax
/* stride */
+ movz
bl V4F_STRIDE(%rdx), %eax
/* stride */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
@@
-52,7
+69,7
@@
_mesa_x86_64_transform_points4_general:
movq V4F_START(%rdx), %rdx /* ptr to first src vertex */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
movq V4F_START(%rdx), %rdx /* ptr to first src vertex */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
- prefetch 16(%rdx)
+ prefetch
t1
16(%rdx)
movaps 0(%rsi), %xmm4 /* m3 | m2 | m1 | m0 */
movaps 16(%rsi), %xmm5 /* m7 | m6 | m5 | m4 */
movaps 0(%rsi), %xmm4 /* m3 | m2 | m1 | m0 */
movaps 16(%rsi), %xmm5 /* m7 | m6 | m5 | m4 */
@@
-63,7
+80,7
@@
_mesa_x86_64_transform_points4_general:
p4_general_loop:
movups (%rdx), %xmm8 /* ox | oy | oz | ow */
p4_general_loop:
movups (%rdx), %xmm8 /* ox | oy | oz | ow */
- prefetch
w
16(%rdi)
+ prefetch
t1
16(%rdi)
pshufd $0x00, %xmm8, %xmm0 /* ox | ox | ox | ox */
addq %rax, %rdx
pshufd $0x00, %xmm8, %xmm0 /* ox | ox | ox | ox */
addq %rax, %rdx
@@
-76,7
+93,7
@@
p4_general_loop:
addps %xmm1, %xmm0 /* ox*m3+oy*m7 | ... */
mulps %xmm7, %xmm3 /* ow*m15 | ow*m14 | ow*m13 | ow*m12 */
addps %xmm2, %xmm0 /* ox*m3+oy*m7+oz*m11 | ... */
addps %xmm1, %xmm0 /* ox*m3+oy*m7 | ... */
mulps %xmm7, %xmm3 /* ow*m15 | ow*m14 | ow*m13 | ow*m12 */
addps %xmm2, %xmm0 /* ox*m3+oy*m7+oz*m11 | ... */
- prefetch 16(%rdx)
+ prefetch
t1
16(%rdx)
addps %xmm3, %xmm0 /* ox*m3+oy*m7+oz*m11+ow*m15 | ... */
movaps %xmm0, (%rdi) /* ->D(3) | ->D(2) | ->D(1) | ->D(0) */
addps %xmm3, %xmm0 /* ox*m3+oy*m7+oz*m11+ow*m15 | ... */
movaps %xmm0, (%rdi) /* ->D(3) | ->D(2) | ->D(1) | ->D(0) */
@@
-101,11
+118,12
@@
p4_constants:
.byte 0x00, 0x00, 0x00, 0x00
.byte 0x00, 0x00, 0x00, 0x00
.byte 0x00, 0x00, 0x00, 0x00
.byte 0x00, 0x00, 0x00, 0x00
.byte 0x00, 0x00, 0x00, 0x00
.byte 0x00, 0x00, 0x00, 0x00
-.float
0f+
1.0
+.float 1.0
.text
.align 16
.globl _mesa_x86_64_transform_points4_3d
.text
.align 16
.globl _mesa_x86_64_transform_points4_3d
+.hidden _mesa_x86_64_transform_points4_3d
/*
* this is slower than _mesa_x86_64_transform_points4_general
* because it ensures that the last matrix row (or is it column?) is 0,0,0,1
/*
* this is slower than _mesa_x86_64_transform_points4_general
* because it ensures that the last matrix row (or is it column?) is 0,0,0,1
@@
-120,7
+138,7
@@
_mesa_x86_64_transform_points4_3d:
movaps 16(%rax), %xmm10
movl V4F_COUNT(%rdx), %ecx /* count */
movaps 16(%rax), %xmm10
movl V4F_COUNT(%rdx), %ecx /* count */
- movz
x V4F_STRIDE(%rdx), %eax
/* stride */
+ movz
bl V4F_STRIDE(%rdx), %eax
/* stride */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
@@
-132,7
+150,7
@@
_mesa_x86_64_transform_points4_3d:
movq V4F_START(%rdx), %rdx /* ptr to first src vertex */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
movq V4F_START(%rdx), %rdx /* ptr to first src vertex */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
- prefetch 16(%rdx)
+ prefetch
t1
16(%rdx)
movaps 0(%rsi), %xmm4 /* m3 | m2 | m1 | m0 */
movaps 16(%rsi), %xmm5 /* m7 | m6 | m5 | m4 */
movaps 0(%rsi), %xmm4 /* m3 | m2 | m1 | m0 */
movaps 16(%rsi), %xmm5 /* m7 | m6 | m5 | m4 */
@@
-148,7
+166,7
@@
_mesa_x86_64_transform_points4_3d:
p4_3d_loop:
movups (%rdx), %xmm8 /* ox | oy | oz | ow */
p4_3d_loop:
movups (%rdx), %xmm8 /* ox | oy | oz | ow */
- prefetch
w
16(%rdi)
+ prefetch
t1
16(%rdi)
pshufd $0x00, %xmm8, %xmm0 /* ox | ox | ox | ox */
addq %rax, %rdx
pshufd $0x00, %xmm8, %xmm0 /* ox | ox | ox | ox */
addq %rax, %rdx
@@
-161,7
+179,7
@@
p4_3d_loop:
addps %xmm1, %xmm0 /* ox*m3+oy*m7 | ... */
mulps %xmm7, %xmm3 /* ow*m15 | ow*m14 | ow*m13 | ow*m12 */
addps %xmm2, %xmm0 /* ox*m3+oy*m7+oz*m11 | ... */
addps %xmm1, %xmm0 /* ox*m3+oy*m7 | ... */
mulps %xmm7, %xmm3 /* ow*m15 | ow*m14 | ow*m13 | ow*m12 */
addps %xmm2, %xmm0 /* ox*m3+oy*m7+oz*m11 | ... */
- prefetch 16(%rdx)
+ prefetch
t1
16(%rdx)
addps %xmm3, %xmm0 /* ox*m3+oy*m7+oz*m11+ow*m15 | ... */
movaps %xmm0, (%rdi) /* ->D(3) | ->D(2) | ->D(1) | ->D(0) */
addps %xmm3, %xmm0 /* ox*m3+oy*m7+oz*m11+ow*m15 | ... */
movaps %xmm0, (%rdi) /* ->D(3) | ->D(2) | ->D(1) | ->D(0) */
@@
-177,10
+195,11
@@
p4_3d_done:
.align 16
.globl _mesa_x86_64_transform_points4_identity
.align 16
.globl _mesa_x86_64_transform_points4_identity
+.hidden _mesa_x86_64_transform_points4_identity
_mesa_x86_64_transform_points4_identity:
movl V4F_COUNT(%rdx), %ecx /* count */
_mesa_x86_64_transform_points4_identity:
movl V4F_COUNT(%rdx), %ecx /* count */
- movz
x V4F_STRIDE(%rdx), %eax
/* stride */
+ movz
bl V4F_STRIDE(%rdx), %eax
/* stride */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
@@
-191,8
+210,8
@@
_mesa_x86_64_transform_points4_identity:
movq V4F_START(%rdx), %rsi /* ptr to first src vertex */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
movq V4F_START(%rdx), %rsi /* ptr to first src vertex */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
- prefetch 64(%rsi)
- prefetch
w
64(%rdi)
+ prefetch
t1
64(%rsi)
+ prefetch
t1
64(%rdi)
add %ecx, %ecx
add %ecx, %ecx
@@
-204,11
+223,12
@@
p4_identity_done:
.align 16
.align 16
-.globl _mesa_x86_64_transform_points4_3d_no_rot
-_mesa_x86_64_transform_points4_3d_no_rot:
+.globl _mesa_3dnow_transform_points4_3d_no_rot
+.hidden _mesa_3dnow_transform_points4_3d_no_rot
+_mesa_3dnow_transform_points4_3d_no_rot:
movl V4F_COUNT(%rdx), %ecx /* count */
movl V4F_COUNT(%rdx), %ecx /* count */
- movz
x V4F_STRIDE(%rdx), %eax
/* stride */
+ movz
bl V4F_STRIDE(%rdx), %eax
/* stride */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
@@
-222,7
+242,7
@@
_mesa_x86_64_transform_points4_3d_no_rot:
movq V4F_START(%rdx), %rdx /* ptr to first src vertex */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
movq V4F_START(%rdx), %rdx /* ptr to first src vertex */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
- prefetch (%rdx)
+ prefetch
t1
(%rdx)
movd (%rsi), %mm0 /* | m00 */
.byte 0x66, 0x66, 0x90 /* manual align += 3 */
movd (%rsi), %mm0 /* | m00 */
.byte 0x66, 0x66, 0x90 /* manual align += 3 */
@@
-235,7
+255,7
@@
_mesa_x86_64_transform_points4_3d_no_rot:
p4_3d_no_rot_loop:
p4_3d_no_rot_loop:
- prefetch
w
32(%rdi)
+ prefetch
t1
32(%rdi)
movq (%rdx), %mm4 /* x1 | x0 */
movq 8(%rdx), %mm5 /* x3 | x2 */
movq (%rdx), %mm4 /* x1 | x0 */
movq 8(%rdx), %mm5 /* x3 | x2 */
@@
-259,7
+279,7
@@
p4_3d_no_rot_loop:
addq $16, %rdi
decl %ecx
addq $16, %rdi
decl %ecx
- prefetch 32(%rdx)
+ prefetch
t1
32(%rdx)
jnz p4_3d_no_rot_loop
p4_3d_no_rot_done:
jnz p4_3d_no_rot_loop
p4_3d_no_rot_done:
@@
-268,11
+288,12
@@
p4_3d_no_rot_done:
.align 16
.align 16
-.globl _mesa_x86_64_transform_points4_perspective
-_mesa_x86_64_transform_points4_perspective:
+.globl _mesa_3dnow_transform_points4_perspective
+.hidden _mesa_3dnow_transform_points4_perspective
+_mesa_3dnow_transform_points4_perspective:
movl V4F_COUNT(%rdx), %ecx /* count */
movl V4F_COUNT(%rdx), %ecx /* count */
- movz
x V4F_STRIDE(%rdx), %eax
/* stride */
+ movz
bl V4F_STRIDE(%rdx), %eax
/* stride */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
@@
-290,7
+311,7
@@
_mesa_x86_64_transform_points4_perspective:
punpckldq 20(%rsi), %mm0 /* m11 | m00 */
movq 32(%rsi), %mm2 /* m21 | m20 */
punpckldq 20(%rsi), %mm0 /* m11 | m00 */
movq 32(%rsi), %mm2 /* m21 | m20 */
- prefetch (%rdx)
+ prefetch
t1
(%rdx)
movd 40(%rsi), %mm1 /* | m22 */
movd 40(%rsi), %mm1 /* | m22 */
@@
-300,7
+321,7
@@
_mesa_x86_64_transform_points4_perspective:
p4_perspective_loop:
p4_perspective_loop:
- prefetch
w 32(%rdi)
/* prefetch 2 vertices ahead */
+ prefetch
t1 32(%rdi)
/* prefetch 2 vertices ahead */
movq (%rdx), %mm4 /* x1 | x0 */
movq 8(%rdx), %mm5 /* x3 | x2 */
movq (%rdx), %mm4 /* x1 | x0 */
movq 8(%rdx), %mm5 /* x3 | x2 */
@@
-326,7
+347,7
@@
p4_perspective_loop:
addq $16, %rdi
decl %ecx
addq $16, %rdi
decl %ecx
- prefetch
32(%rdx)
/* hopefully stride is zero */
+ prefetch
t1 32(%rdx)
/* hopefully stride is zero */
jnz p4_perspective_loop
p4_perspective_done:
jnz p4_perspective_loop
p4_perspective_done:
@@
-334,11
+355,12
@@
p4_perspective_done:
ret
.align 16
ret
.align 16
-.globl _mesa_x86_64_transform_points4_2d_no_rot
-_mesa_x86_64_transform_points4_2d_no_rot:
+.globl _mesa_3dnow_transform_points4_2d_no_rot
+.hidden _mesa_3dnow_transform_points4_2d_no_rot
+_mesa_3dnow_transform_points4_2d_no_rot:
movl V4F_COUNT(%rdx), %ecx /* count */
movl V4F_COUNT(%rdx), %ecx /* count */
- movz
x V4F_STRIDE(%rdx), %eax
/* stride */
+ movz
bl V4F_STRIDE(%rdx), %eax
/* stride */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
@@
-352,14
+374,14
@@
_mesa_x86_64_transform_points4_2d_no_rot:
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
movd (%rsi), %mm0 /* | m00 */
movq V4F_START(%rdi), %rdi /* ptr to first dest vertex */
movd (%rsi), %mm0 /* | m00 */
- prefetch (%rdx)
+ prefetch
t1
(%rdx)
punpckldq 20(%rsi), %mm0 /* m11 | m00 */
movq 48(%rsi), %mm1 /* m31 | m30 */
p4_2d_no_rot_loop:
punpckldq 20(%rsi), %mm0 /* m11 | m00 */
movq 48(%rsi), %mm1 /* m31 | m30 */
p4_2d_no_rot_loop:
- prefetch
w 32(%rdi)
/* prefetch 2 vertices ahead */
+ prefetch
t1 32(%rdi)
/* prefetch 2 vertices ahead */
movq (%rdx), %mm4 /* x1 | x0 */
movq 8(%rdx), %mm5 /* x3 | x2 */
movq (%rdx), %mm4 /* x1 | x0 */
movq 8(%rdx), %mm5 /* x3 | x2 */
@@
-372,7
+394,7
@@
p4_2d_no_rot_loop:
addq %rax, %rdx
pfmul %mm1, %mm6 /* x3*m31 | x3*m30 */
addq %rax, %rdx
pfmul %mm1, %mm6 /* x3*m31 | x3*m30 */
- prefetch
32(%rdx)
/* hopefully stride is zero */
+ prefetch
t1 32(%rdx)
/* hopefully stride is zero */
pfadd %mm4, %mm6 /* x1*m11+x3*m31 | x0*m00+x3*m30 */
movq %mm6, (%rdi) /* write r0, r1 */
pfadd %mm4, %mm6 /* x1*m11+x3*m31 | x0*m00+x3*m30 */
movq %mm6, (%rdi) /* write r0, r1 */
@@
-389,11
+411,12
@@
p4_2d_no_rot_done:
.align 16
.align 16
-.globl _mesa_x86_64_transform_points4_2d
-_mesa_x86_64_transform_points4_2d:
+.globl _mesa_3dnow_transform_points4_2d
+.hidden _mesa_3dnow_transform_points4_2d
+_mesa_3dnow_transform_points4_2d:
movl V4F_COUNT(%rdx), %ecx /* count */
movl V4F_COUNT(%rdx), %ecx /* count */
- movz
x V4F_STRIDE(%rdx), %eax
/* stride */
+ movz
bl V4F_STRIDE(%rdx), %eax
/* stride */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
movl %ecx, V4F_COUNT(%rdi) /* set dest count */
movl $4, V4F_SIZE(%rdi) /* set dest size */
@@
-410,7
+433,7
@@
_mesa_x86_64_transform_points4_2d:
movd (%rsi), %mm0 /* | m00 */
movd 4(%rsi), %mm1 /* | m01 */
movd (%rsi), %mm0 /* | m00 */
movd 4(%rsi), %mm1 /* | m01 */
- prefetch (%rdx)
+ prefetch
t1
(%rdx)
punpckldq 16(%rsi), %mm0 /* m10 | m00 */
.byte 0x66, 0x66, 0x90 /* manual align += 4 */
punpckldq 16(%rsi), %mm0 /* m10 | m00 */
.byte 0x66, 0x66, 0x90 /* manual align += 4 */
@@
-420,7
+443,7
@@
_mesa_x86_64_transform_points4_2d:
p4_2d_loop:
p4_2d_loop:
- prefetch
w 32(%rdi)
/* prefetch 2 vertices ahead */
+ prefetch
t1 32(%rdi)
/* prefetch 2 vertices ahead */
movq (%rdx), %mm3 /* x1 | x0 */
movq 8(%rdx), %mm5 /* x3 | x2 */
movq (%rdx), %mm3 /* x1 | x0 */
movq 8(%rdx), %mm5 /* x3 | x2 */
@@
-437,7
+460,7
@@
p4_2d_loop:
pfacc %mm4, %mm3 /* x0*m01+x1*m11 | x0*m00+x1*m10 */
pfmul %mm2, %mm6 /* x3*m31 | x3*m30 */
pfacc %mm4, %mm3 /* x0*m01+x1*m11 | x0*m00+x1*m10 */
pfmul %mm2, %mm6 /* x3*m31 | x3*m30 */
- prefetch
32(%rdx)
/* hopefully stride is zero */
+ prefetch
t1 32(%rdx)
/* hopefully stride is zero */
pfadd %mm6, %mm3 /* r1 | r0 */
pfadd %mm6, %mm3 /* r1 | r0 */