aboutsummaryrefslogtreecommitdiffstats
path: root/libavcodec/x86
diff options
context:
space:
mode:
authorLoren Merritt <lorenm@u.washington.edu>2011-05-21 23:36:23 +0200
committerReinhard Tartler <siretart@tauware.de>2011-05-22 19:27:18 +0200
commit422b2362fc83ed3a75532ea68a6d167c52f447ec (patch)
tree6e960264a1be0f40765c1f761c6b574d5e2e7b90 /libavcodec/x86
parent165c7c420d611bfa16d999f2033619c542961926 (diff)
downloadffmpeg-422b2362fc83ed3a75532ea68a6d167c52f447ec.tar.gz
dct32_sse: eliminate some spills
125->104 cycles on penryn (x86_64 only)
Diffstat (limited to 'libavcodec/x86')
-rw-r--r--libavcodec/x86/dct32_sse.asm203
-rw-r--r--libavcodec/x86/fmtconvert.asm13
-rw-r--r--libavcodec/x86/x86util.asm20
3 files changed, 176 insertions, 60 deletions
diff --git a/libavcodec/x86/dct32_sse.asm b/libavcodec/x86/dct32_sse.asm
index 2e1176cd84..bafe00289d 100644
--- a/libavcodec/x86/dct32_sse.asm
+++ b/libavcodec/x86/dct32_sse.asm
@@ -20,7 +20,7 @@
;******************************************************************************
%include "x86inc.asm"
-%include "config.asm"
+%include "x86util.asm"
SECTION_RODATA 32
@@ -37,8 +37,9 @@ ps_cos_vec: dd 0.500603, 0.505471, 0.515447, 0.531043
dd 1.000000, 1.000000, 1.306563, 0.541196
dd 1.000000, 0.707107, 1.000000, -0.707107
dd 1.000000, 0.707107, 1.000000, -0.707107
+ dd 0.707107, 0.707107, 0.707107, 0.707107
-
+align 32
ps_p1p1m1m1: dd 0, 0, 0x80000000, 0x80000000, 0, 0, 0x80000000, 0x80000000
%macro BUTTERFLY_SSE 4
@@ -77,6 +78,18 @@ ps_p1p1m1m1: dd 0, 0, 0x80000000, 0x80000000, 0, 0, 0x80000000, 0x80000000
BUTTERFLY0 %1, %2, %3, %4, 0xb1
%endmacro
+%macro BUTTERFLY3V 5
+ movaps m%5, m%1
+ addps m%1, m%2
+ subps m%5, m%2
+ SWAP %2, %5
+ mulps m%2, [ps_cos_vec+192]
+ movaps m%5, m%3
+ addps m%3, m%4
+ subps m%4, m%5
+ mulps m%4, [ps_cos_vec+192]
+%endmacro
+
%macro PASS6_AND_PERMUTE 0
mov tmpd, [outq+4]
movss m7, [outq+72]
@@ -269,9 +282,131 @@ INIT_XMM
%define BUTTERFLY BUTTERFLY_SSE
%define BUTTERFLY0 BUTTERFLY0_SSE
+%ifdef ARCH_X86_64
+%define SPILL SWAP
+%define UNSPILL SWAP
+
+%macro PASS5 0
+ nop ; FIXME code alignment
+ SWAP 5, 8
+ SWAP 4, 12
+ SWAP 6, 14
+ SWAP 7, 13
+ SWAP 0, 15
+ PERMUTE 9,10, 10,12, 11,14, 12,9, 13,11, 14,13
+ TRANSPOSE4x4PS 8, 9, 10, 11, 0
+ BUTTERFLY3V 8, 9, 10, 11, 0
+ addps m10, m11
+ TRANSPOSE4x4PS 12, 13, 14, 15, 0
+ BUTTERFLY3V 12, 13, 14, 15, 0
+ addps m14, m15
+ addps m12, m14
+ addps m14, m13
+ addps m13, m15
+%endmacro
+
+%macro PASS6 0
+ SWAP 9, 12
+ SWAP 11, 14
+ movss [outq+0x00], m8
+ pshuflw m0, m8, 0xe
+ movss [outq+0x10], m9
+ pshuflw m1, m9, 0xe
+ movss [outq+0x20], m10
+ pshuflw m2, m10, 0xe
+ movss [outq+0x30], m11
+ pshuflw m3, m11, 0xe
+ movss [outq+0x40], m12
+ pshuflw m4, m12, 0xe
+ movss [outq+0x50], m13
+ pshuflw m5, m13, 0xe
+ movss [outq+0x60], m14
+ pshuflw m6, m14, 0xe
+ movaps [outq+0x70], m15
+ pshuflw m7, m15, 0xe
+ addss m0, m1
+ addss m1, m2
+ movss [outq+0x08], m0
+ addss m2, m3
+ movss [outq+0x18], m1
+ addss m3, m4
+ movss [outq+0x28], m2
+ addss m4, m5
+ movss [outq+0x38], m3
+ addss m5, m6
+ movss [outq+0x48], m4
+ addss m6, m7
+ movss [outq+0x58], m5
+ movss [outq+0x68], m6
+ movss [outq+0x78], m7
+
+ PERMUTE 1,8, 3,9, 5,10, 7,11, 9,12, 11,13, 13,14, 8,1, 10,3, 12,5, 14,7
+ movhlps m0, m1
+ pshufd m1, m1, 3
+ SWAP 0, 2, 4, 6, 8, 10, 12, 14
+ SWAP 1, 3, 5, 7, 9, 11, 13, 15
+%rep 7
+ movhlps m0, m1
+ pshufd m1, m1, 3
+ addss m15, m1
+ SWAP 0, 2, 4, 6, 8, 10, 12, 14
+ SWAP 1, 3, 5, 7, 9, 11, 13, 15
+%endrep
+%assign i 4
+%rep 15
+ addss m0, m1
+ movss [outq+i], m0
+ SWAP 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
+ %assign i i+8
+%endrep
+%endmacro
+
+%else ; ARCH_X86_32
+%macro SPILL 2 ; xmm#, mempos
+ movaps [outq+(%2-8)*16], m%1
+%endmacro
+%macro UNSPILL 2
+ movaps m%1, [outq+(%2-8)*16]
+%endmacro
+
+%define PASS6 PASS6_AND_PERMUTE
+%macro PASS5 0
+ movaps m2, [ps_cos_vec+160]
+ shufps m3, m3, 0xcc
+
+ BUTTERFLY3 m5, m3, m2, m1
+ SPILL 5, 8
+
+ UNSPILL 1, 9
+ BUTTERFLY3 m1, m3, m2, m5
+ SPILL 1, 14
+
+ BUTTERFLY3 m4, m3, m2, m5
+ SPILL 4, 12
+
+ BUTTERFLY3 m7, m3, m2, m5
+ SPILL 7, 13
+
+ UNSPILL 5, 10
+ BUTTERFLY3 m5, m3, m2, m7
+ SPILL 5, 10
+
+ UNSPILL 4, 11
+ BUTTERFLY3 m4, m3, m2, m7
+ SPILL 4, 11
+
+ BUTTERFLY3 m6, m3, m2, m7
+ SPILL 6, 9
+
+ BUTTERFLY3 m0, m3, m2, m7
+ SPILL 0, 15
+%endmacro
+%endif
+
+
INIT_XMM
; void ff_dct32_float_sse(FFTSample *out, const FFTSample *in)
-cglobal dct32_float_sse, 2,3,8, out, in, tmp
+cglobal dct32_float_sse, 2,3,16, out, in, tmp
; pass 1
movaps m0, [inq+0]
@@ -287,8 +422,8 @@ cglobal dct32_float_sse, 2,3,8, out, in, tmp
; pass 2
movaps m2, [ps_cos_vec+64]
BUTTERFLY m1, m4, m2, m3
- movaps [outq+48], m1
- movaps [outq+ 0], m4
+ SPILL 1, 11
+ SPILL 4, 8
; pass 1
movaps m1, [inq+16]
@@ -313,17 +448,17 @@ cglobal dct32_float_sse, 2,3,8, out, in, tmp
movaps m2, [ps_cos_vec+96]
shufps m1, m1, 0x1b
BUTTERFLY m0, m1, m2, m3
- movaps [outq+112], m0
- movaps [outq+ 96], m1
+ SPILL 0, 15
+ SPILL 1, 14
- movaps m0, [outq+0]
+ UNSPILL 0, 8
shufps m5, m5, 0x1b
BUTTERFLY m0, m5, m2, m3
- movaps m1, [outq+48]
+ UNSPILL 1, 11
shufps m6, m6, 0x1b
BUTTERFLY m1, m6, m2, m3
- movaps [outq+48], m1
+ SPILL 1, 11
shufps m4, m4, 0x1b
BUTTERFLY m7, m4, m2, m3
@@ -335,57 +470,25 @@ cglobal dct32_float_sse, 2,3,8, out, in, tmp
BUTTERFLY2 m5, m3, m2, m1
BUTTERFLY2 m0, m3, m2, m1
- movaps [outq+16], m0
+ SPILL 0, 9
BUTTERFLY2 m6, m3, m2, m1
- movaps [outq+32], m6
+ SPILL 6, 10
- movaps m0, [outq+48]
+ UNSPILL 0, 11
BUTTERFLY2 m0, m3, m2, m1
- movaps [outq+48], m0
+ SPILL 0, 11
BUTTERFLY2 m4, m3, m2, m1
BUTTERFLY2 m7, m3, m2, m1
- movaps m6, [outq+96]
+ UNSPILL 6, 14
BUTTERFLY2 m6, m3, m2, m1
- movaps m0, [outq+112]
+ UNSPILL 0, 15
BUTTERFLY2 m0, m3, m2, m1
- ; pass 5
- movaps m2, [ps_cos_vec+160]
- shufps m3, m3, 0xcc
-
- BUTTERFLY3 m5, m3, m2, m1
- movaps [outq+0], m5
-
- movaps m1, [outq+16]
- BUTTERFLY3 m1, m3, m2, m5
- movaps [outq+96], m1
-
- BUTTERFLY3 m4, m3, m2, m5
- movaps [outq+64], m4
-
- BUTTERFLY3 m7, m3, m2, m5
- movaps [outq+80], m7
-
- movaps m5, [outq+32]
- BUTTERFLY3 m5, m3, m2, m7
- movaps [outq+32], m5
-
- movaps m4, [outq+48]
- BUTTERFLY3 m4, m3, m2, m7
- movaps [outq+48], m4
-
- BUTTERFLY3 m6, m3, m2, m7
- movaps [outq+16], m6
-
- BUTTERFLY3 m0, m3, m2, m7
- movaps [outq+112], m0
-
-
- ; pass 6, no SIMD...
- PASS6_AND_PERMUTE
+ PASS5
+ PASS6
RET
diff --git a/libavcodec/x86/fmtconvert.asm b/libavcodec/x86/fmtconvert.asm
index 13d6cc0130..efab87d570 100644
--- a/libavcodec/x86/fmtconvert.asm
+++ b/libavcodec/x86/fmtconvert.asm
@@ -95,13 +95,6 @@ FLOAT_TO_INT16_INTERLEAVE6 3dn2
; void ff_float_interleave6(float *dst, const float **src, unsigned int len);
;-----------------------------------------------------------------------------
-%macro BUTTERFLYPS 3
- movaps m%3, m%1
- unpcklps m%1, m%2
- unpckhps m%3, m%2
- SWAP %2, %3
-%endmacro
-
%macro FLOAT_INTERLEAVE6 2
cglobal float_interleave6_%1, 2,7,%2, dst, src, src1, src2, src3, src4, src5
%ifdef ARCH_X86_64
@@ -130,9 +123,9 @@ cglobal float_interleave6_%1, 2,7,%2, dst, src, src1, src2, src3, src4, src5
movaps m4, [srcq+src4q]
movaps m5, [srcq+src5q]
- BUTTERFLYPS 0, 1, 6
- BUTTERFLYPS 2, 3, 6
- BUTTERFLYPS 4, 5, 6
+ SBUTTERFLYPS 0, 1, 6
+ SBUTTERFLYPS 2, 3, 6
+ SBUTTERFLYPS 4, 5, 6
movaps m6, m4
shufps m4, m0, 0xe4
diff --git a/libavcodec/x86/x86util.asm b/libavcodec/x86/x86util.asm
index 7bd985a33b..141e96000c 100644
--- a/libavcodec/x86/x86util.asm
+++ b/libavcodec/x86/x86util.asm
@@ -41,6 +41,13 @@
SWAP %2, %4, %3
%endmacro
+%macro SBUTTERFLYPS 3
+ movaps m%3, m%1
+ unpcklps m%1, m%2
+ unpckhps m%3, m%2
+ SWAP %2, %3
+%endmacro
+
%macro TRANSPOSE4x4B 5
SBUTTERFLY bw, %1, %2, %5
SBUTTERFLY bw, %3, %4, %5
@@ -74,6 +81,19 @@
SWAP %2, %3
%endmacro
+; identical behavior to TRANSPOSE4x4D, but using SSE1 float ops
+%macro TRANSPOSE4x4PS 5
+ SBUTTERFLYPS %1, %2, %5
+ SBUTTERFLYPS %3, %4, %5
+ movaps m%5, m%1
+ movlhps m%1, m%3
+ movhlps m%3, m%5
+ movaps m%5, m%2
+ movlhps m%2, m%4
+ movhlps m%4, m%5
+ SWAP %2, %3
+%endmacro
+
%macro TRANSPOSE8x8W 9-11
%ifdef ARCH_X86_64
SBUTTERFLY wd, %1, %2, %9