Skip to content

Commit

Permalink
add some ARM optimizations from libav
Browse files Browse the repository at this point in the history
  • Loading branch information
@eighthave
eighthave committed Aug 24, 2012
1 parent 900aec5 commit ba9d15d
Show file tree
Hide file tree
Showing 3 changed files with 282 additions and 0 deletions.
226 changes: 226 additions & 0 deletions ARM_generate_position_independent_code_to_access_data_symbols.patch
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,226 @@
commit 62634158b7cd39ad1e330a87153a97bf3dc6f8de
Author: Mans Rullgard <[email protected]>
Date: Fri Jun 29 13:35:08 2012 +0100

ARM: generate position independent code to access data symbols

This creates proper position independent code when accessing
data symbols if CONFIG_PIC is set.

References to external symbols should now use the movrelx macro.
Some additional code changes are required since this macro may
need a register to hold the GOT pointer.

Signed-off-by: Mans Rullgard <[email protected]>

diff --git a/libavcodec/arm/ac3dsp_armv6.S b/libavcodec/arm/ac3dsp_armv6.S
index 7e2f40e..f6f297a 100644
--- a/libavcodec/arm/ac3dsp_armv6.S
+++ b/libavcodec/arm/ac3dsp_armv6.S
@@ -26,8 +26,8 @@ function ff_ac3_bit_alloc_calc_bap_armv6, export=1
beq 4f
push {r4-r11,lr}
add r5, sp, #40
- movrel r4, X(ff_ac3_bin_to_band_tab)
- movrel lr, X(ff_ac3_band_start_tab)
+ movrelx r4, X(ff_ac3_bin_to_band_tab), r11
+ movrelx lr, X(ff_ac3_band_start_tab)
ldm r5, {r5-r7}
ldrb r4, [r4, r2]
add r1, r1, r2, lsl #1 @ psd + start
diff --git a/libavcodec/arm/fft_fixed_neon.S b/libavcodec/arm/fft_fixed_neon.S
index 4d891ba..faddc00 100644
--- a/libavcodec/arm/fft_fixed_neon.S
+++ b/libavcodec/arm/fft_fixed_neon.S
@@ -214,7 +214,7 @@ function fft\n\()_neon
bl fft\n4\()_neon
mov r0, r4
pop {r4, lr}
- movrel r1, X(ff_cos_\n\()_fixed)
+ movrelx r1, X(ff_cos_\n\()_fixed)
mov r2, #\n4/2
b fft_pass_neon
endfunc
diff --git a/libavcodec/arm/fft_neon.S b/libavcodec/arm/fft_neon.S
index aa06e6d..c4d8918 100644
--- a/libavcodec/arm/fft_neon.S
+++ b/libavcodec/arm/fft_neon.S
@@ -143,7 +143,7 @@ function fft16_neon
vswp d29, d30 @ q14{r12,i12,i14,r15} q15{r13,i13,i15,r14}
vadd.f32 q0, q12, q13 @ {t1,t2,t5,t6}
vadd.f32 q1, q14, q15 @ {t1a,t2a,t5a,t6a}
- movrel r2, X(ff_cos_16)
+ movrelx r2, X(ff_cos_16)
vsub.f32 q13, q12, q13 @ {t3,t4,t7,t8}
vrev64.32 d1, d1
vsub.f32 q15, q14, q15 @ {t3a,t4a,t7a,t8a}
@@ -290,7 +290,7 @@ function fft\n\()_neon
bl fft\n4\()_neon
mov r0, r4
pop {r4, lr}
- movrel r1, X(ff_cos_\n)
+ movrelx r1, X(ff_cos_\n)
mov r2, #\n4/2
b fft_pass_neon
endfunc
diff --git a/libavcodec/arm/sbrdsp_neon.S b/libavcodec/arm/sbrdsp_neon.S
index 4b681bf..610397f 100644
--- a/libavcodec/arm/sbrdsp_neon.S
+++ b/libavcodec/arm/sbrdsp_neon.S
@@ -307,8 +307,8 @@ function ff_sbr_hf_apply_noise_0_neon, export=1
vmov.i32 d3, #0
.Lhf_apply_noise_0:
push {r4,lr}
+ movrelx r4, X(ff_sbr_noise_table)
ldr r12, [sp, #12]
- movrel r4, X(ff_sbr_noise_table)
add r3, r3, #1
bfc r3, #9, #23
sub r12, r12, #1
@@ -355,8 +355,8 @@ function ff_sbr_hf_apply_noise_1_neon, export=1
eor lr, r12, #1<<31
vmov d3, r12, lr
.Lhf_apply_noise_1:
+ movrelx r4, X(ff_sbr_noise_table)
ldr r12, [sp, #12]
- movrel r4, X(ff_sbr_noise_table)
add r3, r3, #1
bfc r3, #9, #23
sub r12, r12, #1
diff --git a/libavcodec/arm/vp3dsp_neon.S b/libavcodec/arm/vp3dsp_neon.S
index 8d22d00..2a9b25f 100644
--- a/libavcodec/arm/vp3dsp_neon.S
+++ b/libavcodec/arm/vp3dsp_neon.S
@@ -116,9 +116,8 @@ function vp3_idct_start_neon
vadd.s16 q1, q8, q12
vsub.s16 q8, q8, q12
vld1.64 {d28-d31}, [r2,:128]!
-endfunc

-function vp3_idct_core_neon
+vp3_idct_core_neon:
vmull.s16 q2, d18, xC1S7 // (ip[1] * C1) << 16
vmull.s16 q3, d19, xC1S7
vmull.s16 q4, d2, xC4S4 // ((ip[0] + ip[4]) * C4) << 16
diff --git a/libavcodec/arm/vp8_armv6.S b/libavcodec/arm/vp8_armv6.S
index 1fa6d15..1b668bc 100644
--- a/libavcodec/arm/vp8_armv6.S
+++ b/libavcodec/arm/vp8_armv6.S
@@ -65,7 +65,7 @@ T orrcs \cw, \cw, \t1

function ff_decode_block_coeffs_armv6, export=1
push {r0,r1,r4-r11,lr}
- movrel lr, X(ff_vp56_norm_shift)
+ movrelx lr, X(ff_vp56_norm_shift)
ldrd r4, r5, [sp, #44] @ token_prob, qmul
cmp r3, #0
ldr r11, [r5]
@@ -206,7 +206,7 @@ A orrcs r8, r8, r10, lsl r6
mov r9, #8
it ge
addge r12, r12, #1
- movrel r4, X(ff_vp8_dct_cat_prob)
+ movrelx r4, X(ff_vp8_dct_cat_prob), r1
lsl r9, r9, r12
ldr r4, [r4, r12, lsl #2]
add r12, r9, #3
diff --git a/libavcodec/arm/asm.S b/libavcodec/arm/asm.S
index 6038a63..1508180 100644
--- a/libavcodec/arm/asm.S
+++ b/libavcodec/arm/asm.S
@@ -62,7 +62,14 @@ ELF .eabi_attribute 25, \val
.endm

.macro function name, export=0
+ .set .Lpic_idx, 0
+ .set .Lpic_gp, 0
.macro endfunc
+ .if .Lpic_idx
+ .altmacro
+ put_pic %(.Lpic_idx - 1)
+ .noaltmacro
+ .endif
ELF .size \name, . - \name
.endfunc
.purgem endfunc
@@ -106,8 +113,44 @@ ELF .size \name, . - \name
#endif
.endm

+.macro put_pic num
+ put_pic_\num
+.endm
+
+.macro do_def_pic num, val, label
+ .macro put_pic_\num
+ .if \num
+ .altmacro
+ put_pic %(\num - 1)
+ .noaltmacro
+ .endif
+\label: .word \val
+ .purgem put_pic_\num
+ .endm
+.endm
+
+.macro def_pic val, label
+ .altmacro
+ do_def_pic %.Lpic_idx, \val, \label
+ .noaltmacro
+ .set .Lpic_idx, .Lpic_idx + 1
+.endm
+
+.macro ldpic rd, val, indir=0
+ ldr \rd, .Lpicoff\@
+.Lpic\@:
+ .if \indir
+ ldr \rd, [pc, \rd]
+ .else
+ add \rd, pc, \rd
+ .endif
+ def_pic \val - (.Lpic\@ + (8 >> CONFIG_THUMB)), .Lpicoff\@
+.endm
+
.macro movrel rd, val
-#if HAVE_ARMV6T2 && !CONFIG_PIC && !defined(__APPLE__)
+#if CONFIG_PIC
+ ldpic \rd, \val
+#elif HAVE_ARMV6T2 && !defined(__APPLE__)
movw \rd, #:lower16:\val
movt \rd, #:upper16:\val
#else
@@ -115,6 +158,34 @@ ELF .size \name, . - \name
#endif
.endm

+.macro movrelx rd, val, gp
+#if CONFIG_PIC && defined(__ELF__)
+ .ifnb \gp
+ .if .Lpic_gp
+ .unreq gp
+ .endif
+ gp .req \gp
+ ldpic gp, _GLOBAL_OFFSET_TABLE_
+ .elseif !.Lpic_gp
+ gp .req r12
+ ldpic gp, _GLOBAL_OFFSET_TABLE_
+ .endif
+ .set .Lpic_gp, 1
+ ldr \rd, .Lpicoff\@
+ ldr \rd, [gp, \rd]
+ def_pic \val(GOT), .Lpicoff\@
+#elif CONFIG_PIC && defined(__APPLE__)
+ ldpic \rd, .Lpic\@, indir=1
+ .non_lazy_symbol_pointer
+.Lpic\@:
+ .indirect_symbol \val
+ .word 0
+ .text
+#else
+ movrel \rd, \val
+#endif
+.endm
+
.macro ldr_pre rt, rn, rm:vararg
A ldr \rt, [\rn, \rm]!
T add \rn, \rn, \rm
52 changes: 52 additions & 0 deletions ARM_intmath_use_native-size_return_types_for_clipping_functions.patch
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
commit 87fa05a0dae629cdad390adaa1054db3f4ecc3c7
Author: Mans Rullgard <[email protected]>
Date: Sat Aug 11 04:08:15 2012 +0100

ARM: intmath: use native-size return types for clipping functions

This avoids having the compiler redundantly mask the values to
the smaller size.

Signed-off-by: Mans Rullgard <[email protected]>

diff --git a/libavutil/arm/intmath.h b/libavutil/arm/intmath.h
index d5a343c..88e9c26 100644
--- a/libavutil/arm/intmath.h
+++ b/libavutil/arm/intmath.h
@@ -44,7 +44,7 @@ static av_always_inline av_const int FASTDIV(int a, int b)
}

#define av_clip_uint8 av_clip_uint8_arm
-static av_always_inline av_const uint8_t av_clip_uint8_arm(int a)
+static av_always_inline av_const unsigned av_clip_uint8_arm(int a)
{
unsigned x;
__asm__ ("usat %0, #8, %1" : "=r"(x) : "r"(a));
@@ -52,15 +52,15 @@ static av_always_inline av_const uint8_t av_clip_uint8_arm(int a)
}

#define av_clip_int8 av_clip_int8_arm
-static av_always_inline av_const uint8_t av_clip_int8_arm(int a)
+static av_always_inline av_const int av_clip_int8_arm(int a)
{
- unsigned x;
+ int x;
__asm__ ("ssat %0, #8, %1" : "=r"(x) : "r"(a));
return x;
}

#define av_clip_uint16 av_clip_uint16_arm
-static av_always_inline av_const uint16_t av_clip_uint16_arm(int a)
+static av_always_inline av_const unsigned av_clip_uint16_arm(int a)
{
unsigned x;
__asm__ ("usat %0, #16, %1" : "=r"(x) : "r"(a));
@@ -68,7 +68,7 @@ static av_always_inline av_const uint16_t av_clip_uint16_arm(int a)
}

#define av_clip_int16 av_clip_int16_arm
-static av_always_inline av_const int16_t av_clip_int16_arm(int a)
+static av_always_inline av_const int av_clip_int16_arm(int a)
{
int x;
__asm__ ("ssat %0, #16, %1" : "=r"(x) : "r"(a));
4 changes: 4 additions & 0 deletions configure_ffmpeg.sh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,10 @@ fi
# could happen unnoticed.
patch -N -p1 --reject-file=- < redact-plugins.patch
patch -N -p1 --reject-file=- < arm-asm-fix.patch
patch -d ffmpeg -N -p1 --reject-file=- < \
ARM_generate_position_independent_code_to_access_data_symbols.patch
patch -d ffmpeg -N -p1 --reject-file=- < \
ARM_intmath_use_native-size_return_types_for_clipping_functions.patch

pushd ffmpeg

Expand Down

0 comments on commit ba9d15d

Please sign in to comment.