framework for assembly version of the detokenizer

adds a compile time option: --enable-arm-asm-detok which pulls in
vp8/decoder/arm/detokenize.asm

currently about break even speed wise, but changes are pending to
the fill code (branch and load 3 bytes versus conditionally always
load one) and the error handling. Currently it doesn't handle zero
runs or overrunning the buffer.

this is really just so i don't have to rebase my changes all the
time to run benchmarks - now just need to replace one file!

Change-Id: I56d0e2354dc0ca3811bffd0e88fe1f952fa6c797
This commit is contained in:
Johann 2010-08-12 09:05:37 -04:00
parent 633646b73b
commit 9602799cd9
7 changed files with 429 additions and 14 deletions

3
configure vendored

@ -38,6 +38,7 @@ Advanced options:
${toggle_realtime_only} enable this option while building for real-time encoding
${toggle_runtime_cpu_detect} runtime cpu detection
${toggle_shared} shared library support
${toggle_arm_asm_detok} assembly version of the detokenizer (ARM platforms only)
Codecs:
Codecs can be selectively enabled or disabled individually, or by family:
@ -242,6 +243,7 @@ CONFIG_LIST="
spatial_resampling
realtime_only
shared
arm_asm_detok
"
CMDLINE_SELECT="
extra_warnings
@ -278,6 +280,7 @@ CMDLINE_SELECT="
spatial_resampling
realtime_only
shared
arm_asm_detok
"
process_cmdline() {

@ -0,0 +1,333 @@
;
; Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
;
; Use of this source code is governed by a BSD-style license
; that can be found in the LICENSE file in the root of the source
; tree. An additional intellectual property rights grant can be found
; in the file PATENTS. All contributing project authors may
; be found in the AUTHORS file in the root of the source tree.
;
EXPORT |vp8_decode_mb_tokens_v6|
AREA |.text|, CODE, READONLY ; name this block of code
INCLUDE vpx_asm_offsets.asm
l_qcoeff EQU 0
l_i EQU 4
l_type EQU 8
l_stop EQU 12
l_c EQU 16
l_l_ptr EQU 20
l_a_ptr EQU 24
l_bc EQU 28
l_coef_ptr EQU 32
l_stacksize EQU 64
;; constant offsets -- these should be created at build time
c_onyxblock2left_offset EQU 25
c_onyxblock2above_offset EQU 50
c_entropy_nodes EQU 11
c_dct_eob_token EQU 11
|vp8_decode_mb_tokens_v6| PROC
stmdb sp!, {r4 - r11, lr}
sub sp, sp, #l_stacksize
mov r7, r1 ; type
mov r9, r0 ; detoken
ldr r1, [r9, #detok_current_bc]
ldr r0, [r9, #detok_qcoeff_start_ptr]
mov r11, #0 ; i
mov r3, #0x10 ; stop
cmp r7, #1 ; type ?= 1
addeq r11, r11, #24 ; i = 24
addeq r3, r3, #8 ; stop = 24
addeq r0, r0, #3, 24 ; qcoefptr += 24*16 ?CHECKME
str r0, [sp, #l_qcoeff]
str r11, [sp, #l_i]
str r7, [sp, #l_type]
str r3, [sp, #l_stop]
str r1, [sp, #l_bc]
add lr, r9, r7, lsl #2 ; detoken + type*4
ldr r8, [r1, #bool_decoder_user_buffer]
ldr r10, [lr, #detok_coef_probs] ; coef_probs[type]
ldr r5, [r1, #bool_decoder_count]
ldr r6, [r1, #bool_decoder_range]
ldr r4, [r1, #bool_decoder_value]
str r10, [sp, #l_coef_ptr]
;align 4
BLOCK_LOOP
ldr r3, [r9, #detok_ptr_onyxblock2context_leftabove]
ldr r2, [r9, #detok_A]
ldr r1, [r9, #detok_L]
ldrb r12, [r3, r11]! ; onyxblock2context[i]
cmp r7, #0 ; c = !type
moveq r7, #1
movne r7, #0
ldr r0, [r2, r12, lsl #2] ; A[onyxblock2context[i]]
add r1, r1, r12, lsl #4 ; L + onyxblock2context[i] << 4
; A is ptr to ptr (**)
; L is ptr to data (*[4])
ldrb r2, [r3, #c_onyxblock2above_offset] ; + above offset
ldrb r3, [r3, #c_onyxblock2left_offset] ; + left offset
mov lr, #c_entropy_nodes ; ENTROPY_NODES = 11
;; ;++
ldr r2, [r0, r2, lsl #2]! ; A + above offset
ldr r3, [r1, r3, lsl #2]! ; L + left offset
; VP8_COMBINEENTROPYCONTETEXTS(t, *a, *l) => t = ((*a) != 0) + ((*l) !=0)
cmp r2, #0 ; *a ?= 0
movne r2, #1 ; haha if a == 0 no need to set up another var to state that pretty sweet :)
cmp r3, #0 ; *l ?= 0
addne r2, r2, #1 ; t
str r1, [sp, #l_l_ptr] ; save &l
str r0, [sp, #l_a_ptr] ; save &a
smlabb r0, r2, lr, r10 ; Prob = coef_probs + (t * ENTROPY_NODES)
mov r1, #0 ; t = 0
str r7, [sp, #l_c]
;align 4
COEFF_LOOP
ldr r3, [r9, #detok_ptr_onyx_coef_bands_x]
ldr lr, [r9, #detok_onyx_coef_tree_ptr]
; onyx_coef_bands_x is UINT16
add r3, r3, r7, lsl #1 ; coef_bands_x[c]
ldrh r3, [r3] ; UINT16
;++
add r0, r0, r3 ; Prob += coef_bands_x[c]
;align 4
get_token_loop
ldrb r2, [r0, +r1, asr #1] ; Prob[t >> 1]
mov r3, r6, lsl #8 ; range << 8
sub r3, r3, #256 ; (range << 8) - (1 << 8)
mov r10, #1 ; 1
smlawb r2, r3, r2, r10 ; split = 1 + (((range-1) * probability) >> 8)
ldrb r12, [r8] ; load cx data byte in stall slot : r8 = bufptr
;++
subs r3, r4, r2, lsl #24 ; value-(split<<24): used later to calculate shift for NORMALIZE
addhs r1, r1, #1 ; t += 1
movhs r4, r3 ; value -= bigsplit (split << 24)
subhs r2, r6, r2 ; range -= split
; movlo r6, r2 ; range = split
ldrsb r1, [lr, r1] ; t = onyx_coef_tree_ptr[t]
; NORMALIZE
clz r3, r2 ; vp8dx_bitreader_norm[range] + 24
sub r3, r3, #24 ; vp8dx_bitreader_norm[range]
subs r5, r5, r3 ; count -= shift
mov r6, r2, lsl r3 ; range <<= shift
mov r4, r4, lsl r3 ; value <<= shift
; if count <= 0, += BR_COUNT; value |= *bufptr++ << (BR_COUNT-count); BR_COUNT = 8, but need to upshift values by +16
addle r5, r5, #8 ; count += 8
rsble r3, r5, #24 ; 24 - count
addle r8, r8, #1 ; bufptr++
orrle r4, r4, r12, lsl r3 ; value |= *bufptr << shift + 16
cmp r1, #0 ; t ?= 0
bgt get_token_loop ; while (t > 0)
cmn r1, #c_dct_eob_token ; if(t == -DCT_EOB_TOKEN)
beq END_OF_BLOCK ; break
rsb lr, r1, #0 ; v = -t;
cmp lr, #4 ; if(v > FOUR_TOKEN)
ble SKIP_EXTRABITS
ldr r3, [r9, #detok_teb_base_ptr]
mov r11, #1 ; 1 in split = 1 + ... nope, v+= 1 << bits_count
add r7, r3, lr, lsl #4 ; detok_teb_base_ptr + (v << 4)
ldrsh lr, [r7, #tokenextrabits_min_val] ; v = teb_ptr->min_val
ldrsh r0, [r7, #tokenextrabits_length] ; bits_count = teb_ptr->Length
extrabits_loop
add r3, r0, r7 ; &teb_ptr->Probs[bits_count]
ldrb r2, [r3, #4] ; probability. why +4?
mov r3, r6, lsl #8 ; range << 8
sub r3, r3, #256 ; range << 8 + 1 << 8
smlawb r2, r3, r2, r11 ; split = 1 + (((range-1) * probability) >> 8)
ldrb r12, [r8] ; *bufptr
;++
subs r10, r4, r2, lsl #24 ; value - (split<<24)
movhs r4, r10 ; value = value - (split << 24)
subhs r2, r6, r2 ; range = range - split
addhs lr, lr, r11, lsl r0 ; v += ((UINT16)1<<bits_count)
; NORMALIZE
clz r3, r2 ; shift - leading zeros in split
sub r3, r3, #24 ; don't count first 3 bytes
subs r5, r5, r3 ; count -= shift
mov r6, r2, lsl r3 ; range = range << shift
mov r4, r4, lsl r3 ; value <<= shift
addle r5, r5, #8 ; count += BR_COUNT
addle r8, r8, #1 ; bufptr++
rsble r3, r5, #24 ; BR_COUNT - count
orrle r4, r4, r12, lsl r3 ; value |= *bufptr << (BR_COUNT - count)
subs r0, r0, #1 ; bits_count --
bpl extrabits_loop
SKIP_EXTRABITS
ldr r11, [sp, #l_qcoeff]
ldr r0, [sp, #l_coef_ptr] ; Prob = coef_probs
cmp r1, #0 ; check for nonzero token - if (t)
beq SKIP_EOB_CHECK ; if t is zero, we will skip the eob table chec
add r3, r6, #1 ; range + 1
mov r2, r3, lsr #1 ; split = (range + 1) >> 1
subs r3, r4, r2, lsl #24 ; value - (split<<24)
movhs r4, r3 ; value -= (split << 24)
subhs r2, r6, r2 ; range -= split
mvnhs r3, lr ; -v
addhs lr, r3, #1 ; v = (v ^ -1) + 1
; NORMALIZE
clz r3, r2 ; leading 0s in split
sub r3, r3, #24 ; shift
subs r5, r5, r3 ; count -= shift
mov r6, r2, lsl r3 ; range <<= shift
mov r4, r4, lsl r3 ; value <<= shift
ldrleb r2, [r8], #1 ; *(bufptr++)
addle r5, r5, #8 ; count += 8
rsble r3, r5, #24 ; BR_COUNT - count
orrle r4, r4, r2, lsl r3 ; value |= *bufptr << (BR_COUNT - count)
add r0, r0, #0xB ; Prob += ENTROPY_NODES (11)
cmn r1, #1 ; t < -ONE_TOKEN
addlt r0, r0, #0xB ; Prob += ENTROPY_NODES (11)
mvn r1, #1 ; t = -1 ???? C is -2
SKIP_EOB_CHECK
ldr r7, [sp, #l_c] ; c
ldr r3, [r9, #detok_scan]
add r1, r1, #2 ; t+= 2
cmp r7, #(0x10 - 1) ; c should will be one higher
ldr r3, [r3, +r7, lsl #2] ; scan[c] this needs pre-inc c value
add r7, r7, #1 ; c++
add r3, r11, r3, lsl #1 ; qcoeff + scan[c]
str r7, [sp, #l_c] ; store c
strh lr, [r3] ; qcoef_ptr[scan[c]] = v
blt COEFF_LOOP
sub r7, r7, #1 ; if(t != -DCT_EOB_TOKEN) --c ; never stored! no condition!
END_OF_BLOCK
ldr r3, [sp, #l_type] ; type
ldr r10, [sp, #l_coef_ptr] ; coef_ptr
ldr r0, [sp, #l_qcoeff] ; qcoeff
ldr r11, [sp, #l_i] ; i
ldr r12, [sp, #l_stop] ; stop
cmp r3, #0 ; type ?= 0
moveq r1, #1
movne r1, #0
add r3, r11, r9 ; detok + i
cmp r7, r1 ; c ?= !type
strb r7, [r3, #detok_eob] ; eob[i] = c
ldr r7, [sp, #l_l_ptr] ; l
ldr r2, [sp, #l_a_ptr] ; a
movne r3, #1 ; t
moveq r3, #0
add r0, r0, #0x20 ; qcoeff += 32 (16 * 2?)
add r11, r11, #1 ; i++
str r3, [r7] ; *l = t
str r3, [r2] ; *a = t
str r0, [sp, #l_qcoeff] ; qcoeff
str r11, [sp, #l_i] ; i
cmp r11, r12 ; i >= stop ? VERIFY should be strictly LT(<)?
ldr r7, [sp, #l_type] ; type
mov lr, #0xB ; 11 (ENTORPY_NODES?)
blt BLOCK_LOOP
cmp r11, #0x19 ; i ?= 25
bne ln2_decode_mb_to
ldr r12, [r9, #detok_qcoeff_start_ptr]
ldr r10, [r9, #detok_coef_probs]
mov r7, #0 ; type/i = 0
mov r3, #0x10 ; stop = 0
str r12, [sp, #l_qcoeff] ; qcoeff_ptr = qcoeff_start_ptr
str r7, [sp, #l_i]
str r7, [sp, #l_type]
str r3, [sp, #l_stop]
str r10, [sp, #l_coef_ptr] ; coef_probs = coef_probs[type] (0)
b BLOCK_LOOP
ln2_decode_mb_to
cmp r11, #0x10 ; i ?= 16
bne ln1_decode_mb_to
mov r10, #detok_coef_probs
add r10, r10, #2*4 ; coef_probs[type]
ldr r10, [r9, r10] ; detok + 48 - THIS IS PROBABLY THE ISSUE: NEW STRUCTURE
mov r7, #2 ; type = 2
mov r3, #0x18 ; stop = 24
str r7, [sp, #l_type]
str r3, [sp, #l_stop]
str r10, [sp, #l_coef_ptr] ; coef_probs = coef_probs[type] - didn't want to add 2 to coef_probs
b BLOCK_LOOP
ln1_decode_mb_to
ldr r2, [sp, #l_bc]
mov r0, #0
nop
str r8, [r2, #bool_decoder_user_buffer]
str r5, [r2, #bool_decoder_count]
str r4, [r2, #bool_decoder_value]
str r6, [r2, #bool_decoder_range]
add sp, sp, #l_stacksize
ldmia sp!, {r4 - r11, pc}
ENDP ; |vp8_decode_mb_tokens_v6|
END

@ -0,0 +1,22 @@
/*
* Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef DETOKENIZE_ARM_H
#define DETOKENIZE_ARM_H
#if HAVE_ARMV6
#if CONFIG_ARM_ASM_DETOK
void vp8_init_detokenizer(VP8D_COMP *dx);
void vp8_decode_mb_tokens_v6(DETOK *detoken, int type);
#endif
#endif
#endif

@ -14,6 +14,7 @@
#include "onyxd_int.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/mem.h"
#include "detokenize.h"
#define BOOL_DATA UINT8
@ -103,6 +104,34 @@ void vp8_reset_mb_tokens_context(MACROBLOCKD *x)
*l = 0;
}
}
#if CONFIG_ARM_ASM_DETOK
DECLARE_ALIGNED(16, const UINT8, vp8_block2context_leftabove[25*3]) =
{
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, //end of vp8_block2context
0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 0, 0, 1, 1, 0, 0, 1, 1, 0, //end of vp8_block2left
0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 0, 1, 0, 1, 0, 1, 0 //end of vp8_block2above
};
void vp8_init_detokenizer(VP8D_COMP *dx)
{
const VP8_COMMON *const oc = & dx->common;
MACROBLOCKD *x = & dx->mb;
dx->detoken.vp8_coef_tree_ptr = vp8_coef_tree;
dx->detoken.ptr_onyxblock2context_leftabove = vp8_block2context_leftabove;
dx->detoken.ptr_onyx_coef_bands_x = vp8_coef_bands_x;
dx->detoken.scan = vp8_default_zig_zag1d;
dx->detoken.teb_base_ptr = vp8d_token_extra_bits2;
dx->detoken.qcoeff_start_ptr = &x->qcoeff[0];
dx->detoken.coef_probs[0] = (oc->fc.coef_probs [0] [ 0 ] [0]);
dx->detoken.coef_probs[1] = (oc->fc.coef_probs [1] [ 0 ] [0]);
dx->detoken.coef_probs[2] = (oc->fc.coef_probs [2] [ 0 ] [0]);
dx->detoken.coef_probs[3] = (oc->fc.coef_probs [3] [ 0 ] [0]);
}
#endif
DECLARE_ALIGNED(16, extern const unsigned int, vp8dx_bitreader_norm[256]);
#define FILL \
if(count < 0) \
@ -200,6 +229,35 @@ DECLARE_ALIGNED(16, extern const unsigned int, vp8dx_bitreader_norm[256]);
}\
NORMALIZE
#if CONFIG_ARM_ASM_DETOK
int vp8_decode_mb_tokens(VP8D_COMP *dx, MACROBLOCKD *x)
{
int eobtotal = 0;
int i, type;
dx->detoken.current_bc = x->current_bc;
dx->detoken.A = x->above_context;
dx->detoken.L = x->left_context;
type = 3;
if (x->mbmi.mode != B_PRED && x->mbmi.mode != SPLITMV)
{
type = 1;
eobtotal -= 16;
}
vp8_decode_mb_tokens_v6(&dx->detoken, type);
for (i = 0; i < 25; i++)
{
x->block[i].eob = dx->detoken.eob[i];
eobtotal += dx->detoken.eob[i];
}
return eobtotal;
}
#else
int vp8_decode_mb_tokens(VP8D_COMP *dx, MACROBLOCKD *x)
{
ENTROPY_CONTEXT **const A = x->above_context;
@ -395,3 +453,4 @@ BLOCK_FINISHED:
return eobtotal;
}
#endif //!CONFIG_ASM_DETOK

@ -9,12 +9,16 @@
*/
#ifndef detokenize_h
#define detokenize_h 1
#ifndef DETOKENIZE_H
#define DETOKENIZE_H
#include "onyxd_int.h"
#if ARCH_ARM
#include "arm/detokenize_arm.h"
#endif
void vp8_reset_mb_tokens_context(MACROBLOCKD *x);
int vp8_decode_mb_tokens(VP8D_COMP *, MACROBLOCKD *);
#endif /* detokenize_h */
#endif /* DETOKENIZE_H */

@ -29,13 +29,11 @@
#include "vpx_scale/vpxscale.h"
#include "systemdependent.h"
#include "vpx_ports/vpx_timer.h"
#include "detokenize.h"
extern void vp8_init_loop_filter(VP8_COMMON *cm);
extern void vp8cx_init_de_quantizer(VP8D_COMP *pbi);
// DEBUG code
#if CONFIG_DEBUG
void vp8_recon_write_yuv_frame(unsigned char *name, YV12_BUFFER_CONFIG *s)
{
@ -129,6 +127,9 @@ VP8D_PTR vp8dx_create_decompressor(VP8D_CONFIG *oxcf)
cm->last_sharpness_level = cm->sharpness_level;
}
#if CONFIG_ARM_ASM_DETOK
vp8_init_detokenizer(pbi);
#endif
pbi->common.error.setjmp = 0;
return (VP8D_PTR) pbi;
}

@ -11,24 +11,17 @@
#VP8_DX_SRCS list is modified according to different platforms.
#File list for arm
# decoder
#VP8_DX_SRCS-$(HAVE_ARMV6) += decoder/arm/decodframe_arm.c
VP8_DX_SRCS-$(HAVE_ARMV6) += decoder/arm/dequantize_arm.c
VP8_DX_SRCS-$(HAVE_ARMV6) += decoder/arm/dsystemdependent.c
#VP8_DX_SRCS_REMOVE-$(HAVE_ARMV6) += decoder/decodframe.c
VP8_DX_SRCS_REMOVE-$(HAVE_ARMV6) += decoder/dequantize.c
VP8_DX_SRCS_REMOVE-$(HAVE_ARMV6) += decoder/generic/dsystemdependent.c
VP8_DX_SRCS-$(CONFIG_ARM_ASM_DETOK) += decoder/arm/detokenize$(ASM)
#File list for armv6
# decoder
VP8_DX_SRCS-$(HAVE_ARMV6) += decoder/arm/armv6/dequant_dc_idct_v6$(ASM)
VP8_DX_SRCS-$(HAVE_ARMV6) += decoder/arm/armv6/dequant_idct_v6$(ASM)
VP8_DX_SRCS-$(HAVE_ARMV6) += decoder/arm/armv6/dequantize_v6$(ASM)
#File list for neon
# decoder
VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/dequant_dc_idct_neon$(ASM)
VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/dequant_idct_neon$(ASM)
VP8_DX_SRCS-$(HAVE_ARMV7) += decoder/arm/neon/dequantizeb_neon$(ASM)