aom/vp8/decoder/onyxd_if.c

547 строки
13 KiB
C
Исходник Обычный вид История

2010-05-18 19:58:33 +04:00
/*
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
2010-05-18 19:58:33 +04:00
*
* 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.
2010-05-18 19:58:33 +04:00
*/
#include "onyxc_int.h"
#if CONFIG_POSTPROC
#include "postproc.h"
#endif
#include "onyxd.h"
#include "onyxd_int.h"
#include "vpx_mem/vpx_mem.h"
#include "alloccommon.h"
#include "vpx_scale/yv12extend.h"
#include "loopfilter.h"
#include "swapyv12buffer.h"
#include "g_common.h"
#include "threading.h"
#include "decoderthreading.h"
#include <stdio.h>
2010-05-18 19:58:33 +04:00
#include "quant_common.h"
#include "vpx_scale/vpxscale.h"
#include "systemdependent.h"
#include "vpx_ports/vpx_timer.h"
#include "detokenize.h"
Add runtime CPU detection support for ARM. The primary goal is to allow a binary to be built which supports NEON, but can fall back to non-NEON routines, since some Android devices do not have NEON, even if they are otherwise ARMv7 (e.g., Tegra). The configure-generated flags HAVE_ARMV7, etc., are used to decide which versions of each function to build, and when CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen at run time. In order for this to work, the CFLAGS must be set to something appropriate (e.g., without -mfpu=neon for ARMv7, and with appropriate -march and -mcpu for even earlier configurations), or the native C code will not be able to run. The ASFLAGS must remain set for the most advanced instruction set required at build time, since the ARM assembler will refuse to emit them otherwise. I have not attempted to make any changes to configure to do this automatically. Doing so will probably require the addition of new configure options. Many of the hooks for RTCD on ARM were already there, but a lot of the code had bit-rotted, and a good deal of the ARM-specific code is not integrated into the RTCD structs at all. I did not try to resolve the latter, merely to add the minimal amount of protection around them to allow RTCD to work. Those functions that were called based on an ifdef at the calling site were expanded to check the RTCD flags at that site, but they should be added to an RTCD struct somewhere in the future. The functions invoked with global function pointers still are, but these should be moved into an RTCD struct for thread safety (I believe every platform currently supported has atomic pointer stores, but this is not guaranteed). The encoder's boolhuff functions did not even have _c and armv7 suffixes, and the correct version was resolved at link time. The token packing functions did have appropriate suffixes, but the version was selected with a define, with no associated RTCD struct. However, for both of these, the only armv7 instruction they actually used was rbit, and this was completely superfluous, so I reworked them to avoid it. The only non-ARMv4 instruction remaining in them is clz, which is ARMv5 (not even ARMv5TE is required). Considering that there are no ARM-specific configs which are not at least ARMv5TE, I did not try to detect these at runtime, and simply enable them for ARMv5 and above. Finally, the NEON register saving code was completely non-reentrant, since it saved the registers to a global, static variable. I moved the storage for this onto the stack. A single binary built with this code was tested on an ARM11 (ARMv6) and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder, and produced identical output, while using the correct accelerated functions on each. I did not test on any earlier processors. Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
2010-10-21 02:39:11 +04:00
#if ARCH_ARM
#include "vpx_ports/arm.h"
#endif
2010-05-18 19:58:33 +04:00
extern void vp8_init_loop_filter(VP8_COMMON *cm);
extern void vp8cx_init_de_quantizer(VP8D_COMP *pbi);
#if CONFIG_DEBUG
void vp8_recon_write_yuv_frame(unsigned char *name, YV12_BUFFER_CONFIG *s)
{
FILE *yuv_file = fopen((char *)name, "ab");
unsigned char *src = s->y_buffer;
int h = s->y_height;
do
{
fwrite(src, s->y_width, 1, yuv_file);
src += s->y_stride;
}
while (--h);
src = s->u_buffer;
h = s->uv_height;
do
{
fwrite(src, s->uv_width, 1, yuv_file);
src += s->uv_stride;
}
while (--h);
src = s->v_buffer;
h = s->uv_height;
do
{
fwrite(src, s->uv_width, 1, yuv_file);
src += s->uv_stride;
}
while (--h);
fclose(yuv_file);
}
#endif
void vp8dx_initialize()
{
static int init_done = 0;
if (!init_done)
{
vp8_initialize_common();
vp8_scale_machine_specific_config();
init_done = 1;
}
}
VP8D_PTR vp8dx_create_decompressor(VP8D_CONFIG *oxcf)
{
VP8D_COMP *pbi = vpx_memalign(32, sizeof(VP8D_COMP));
if (!pbi)
return NULL;
vpx_memset(pbi, 0, sizeof(VP8D_COMP));
if (setjmp(pbi->common.error.jmp))
{
pbi->common.error.setjmp = 0;
vp8dx_remove_decompressor(pbi);
return 0;
}
pbi->common.error.setjmp = 1;
vp8dx_initialize();
vp8_create_common(&pbi->common);
vp8_dmachine_specific_config(pbi);
pbi->common.current_video_frame = 0;
pbi->ready_for_new_data = 1;
pbi->CPUFreq = 0; /*vp8_get_processor_freq();*/
2010-05-18 19:58:33 +04:00
pbi->max_threads = oxcf->max_threads;
vp8_decoder_create_threads(pbi);
/* vp8cx_init_de_quantizer() is first called here. Add check in frame_init_dequantizer() to avoid
* unnecessary calling of vp8cx_init_de_quantizer() for every frame.
*/
2010-05-18 19:58:33 +04:00
vp8cx_init_de_quantizer(pbi);
{
VP8_COMMON *cm = &pbi->common;
vp8_init_loop_filter(cm);
cm->last_frame_type = KEY_FRAME;
cm->last_filter_type = cm->filter_type;
cm->last_sharpness_level = cm->sharpness_level;
}
#if CONFIG_ARM_ASM_DETOK
vp8_init_detokenizer(pbi);
#endif
2010-05-18 19:58:33 +04:00
pbi->common.error.setjmp = 0;
return (VP8D_PTR) pbi;
}
void vp8dx_remove_decompressor(VP8D_PTR ptr)
{
VP8D_COMP *pbi = (VP8D_COMP *) ptr;
if (!pbi)
return;
#if CONFIG_MULTITHREAD
if (pbi->b_multithreaded_rd)
vp8mt_de_alloc_temp_buffers(pbi, pbi->common.mb_rows);
#endif
2010-05-18 19:58:33 +04:00
vp8_decoder_remove_threads(pbi);
vp8_remove_common(&pbi->common);
vpx_free(pbi);
}
void vp8dx_set_setting(VP8D_PTR comp, VP8D_SETTING oxst, int x)
{
VP8D_COMP *pbi = (VP8D_COMP *) comp;
(void) pbi;
(void) x;
switch (oxst)
{
case VP8D_OK:
break;
}
}
int vp8dx_get_setting(VP8D_PTR comp, VP8D_SETTING oxst)
{
VP8D_COMP *pbi = (VP8D_COMP *) comp;
(void) pbi;
switch (oxst)
{
case VP8D_OK:
break;
}
return -1;
}
int vp8dx_get_reference(VP8D_PTR ptr, VP8_REFFRAME ref_frame_flag, YV12_BUFFER_CONFIG *sd)
{
VP8D_COMP *pbi = (VP8D_COMP *) ptr;
VP8_COMMON *cm = &pbi->common;
int ref_fb_idx;
2010-05-18 19:58:33 +04:00
if (ref_frame_flag == VP8_LAST_FLAG)
ref_fb_idx = cm->lst_fb_idx;
2010-05-18 19:58:33 +04:00
else if (ref_frame_flag == VP8_GOLD_FLAG)
ref_fb_idx = cm->gld_fb_idx;
2010-05-18 19:58:33 +04:00
else if (ref_frame_flag == VP8_ALT_FLAG)
ref_fb_idx = cm->alt_fb_idx;
2010-05-18 19:58:33 +04:00
else
return -1;
vp8_yv12_copy_frame_ptr(&cm->yv12_fb[ref_fb_idx], sd);
2010-05-18 19:58:33 +04:00
return 0;
}
int vp8dx_set_reference(VP8D_PTR ptr, VP8_REFFRAME ref_frame_flag, YV12_BUFFER_CONFIG *sd)
{
VP8D_COMP *pbi = (VP8D_COMP *) ptr;
VP8_COMMON *cm = &pbi->common;
int ref_fb_idx;
2010-05-18 19:58:33 +04:00
if (ref_frame_flag == VP8_LAST_FLAG)
ref_fb_idx = cm->lst_fb_idx;
2010-05-18 19:58:33 +04:00
else if (ref_frame_flag == VP8_GOLD_FLAG)
ref_fb_idx = cm->gld_fb_idx;
2010-05-18 19:58:33 +04:00
else if (ref_frame_flag == VP8_ALT_FLAG)
ref_fb_idx = cm->alt_fb_idx;
2010-05-18 19:58:33 +04:00
else
return -1;
vp8_yv12_copy_frame_ptr(sd, &cm->yv12_fb[ref_fb_idx]);
2010-05-18 19:58:33 +04:00
return 0;
}
/*For ARM NEON, d8-d15 are callee-saved registers, and need to be saved by us.*/
2010-05-18 19:58:33 +04:00
#if HAVE_ARMV7
extern void vp8_push_neon(INT64 *store);
extern void vp8_pop_neon(INT64 *store);
#endif
static int get_free_fb (VP8_COMMON *cm)
{
int i;
for (i = 0; i < NUM_YV12_BUFFERS; i++)
if (cm->fb_idx_ref_cnt[i] == 0)
break;
cm->fb_idx_ref_cnt[i] = 1;
return i;
}
static void ref_cnt_fb (int *buf, int *idx, int new_idx)
{
if (buf[*idx] > 0)
buf[*idx]--;
*idx = new_idx;
buf[new_idx]++;
}
/* If any buffer copy / swapping is signalled it should be done here. */
static int swap_frame_buffers (VP8_COMMON *cm)
{
int fb_to_update_with, err = 0;
if (cm->refresh_last_frame)
fb_to_update_with = cm->lst_fb_idx;
else
fb_to_update_with = cm->new_fb_idx;
/* The alternate reference frame or golden frame can be updated
* using the new, last, or golden/alt ref frame. If it
* is updated using the newly decoded frame it is a refresh.
* An update using the last or golden/alt ref frame is a copy.
*/
if (cm->copy_buffer_to_arf)
{
int new_fb = 0;
if (cm->copy_buffer_to_arf == 1)
new_fb = fb_to_update_with;
else if (cm->copy_buffer_to_arf == 2)
new_fb = cm->gld_fb_idx;
else
err = -1;
ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->alt_fb_idx, new_fb);
}
if (cm->copy_buffer_to_gf)
{
int new_fb = 0;
if (cm->copy_buffer_to_gf == 1)
new_fb = fb_to_update_with;
else if (cm->copy_buffer_to_gf == 2)
new_fb = cm->alt_fb_idx;
else
err = -1;
ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->gld_fb_idx, new_fb);
}
if (cm->refresh_golden_frame)
ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->gld_fb_idx, cm->new_fb_idx);
if (cm->refresh_alt_ref_frame)
ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->alt_fb_idx, cm->new_fb_idx);
if (cm->refresh_last_frame)
{
ref_cnt_fb (cm->fb_idx_ref_cnt, &cm->lst_fb_idx, cm->new_fb_idx);
cm->frame_to_show = &cm->yv12_fb[cm->lst_fb_idx];
}
else
cm->frame_to_show = &cm->yv12_fb[cm->new_fb_idx];
cm->fb_idx_ref_cnt[cm->new_fb_idx]--;
return err;
}
2010-05-18 19:58:33 +04:00
int vp8dx_receive_compressed_data(VP8D_PTR ptr, unsigned long size, const unsigned char *source, INT64 time_stamp)
{
Add runtime CPU detection support for ARM. The primary goal is to allow a binary to be built which supports NEON, but can fall back to non-NEON routines, since some Android devices do not have NEON, even if they are otherwise ARMv7 (e.g., Tegra). The configure-generated flags HAVE_ARMV7, etc., are used to decide which versions of each function to build, and when CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen at run time. In order for this to work, the CFLAGS must be set to something appropriate (e.g., without -mfpu=neon for ARMv7, and with appropriate -march and -mcpu for even earlier configurations), or the native C code will not be able to run. The ASFLAGS must remain set for the most advanced instruction set required at build time, since the ARM assembler will refuse to emit them otherwise. I have not attempted to make any changes to configure to do this automatically. Doing so will probably require the addition of new configure options. Many of the hooks for RTCD on ARM were already there, but a lot of the code had bit-rotted, and a good deal of the ARM-specific code is not integrated into the RTCD structs at all. I did not try to resolve the latter, merely to add the minimal amount of protection around them to allow RTCD to work. Those functions that were called based on an ifdef at the calling site were expanded to check the RTCD flags at that site, but they should be added to an RTCD struct somewhere in the future. The functions invoked with global function pointers still are, but these should be moved into an RTCD struct for thread safety (I believe every platform currently supported has atomic pointer stores, but this is not guaranteed). The encoder's boolhuff functions did not even have _c and armv7 suffixes, and the correct version was resolved at link time. The token packing functions did have appropriate suffixes, but the version was selected with a define, with no associated RTCD struct. However, for both of these, the only armv7 instruction they actually used was rbit, and this was completely superfluous, so I reworked them to avoid it. The only non-ARMv4 instruction remaining in them is clz, which is ARMv5 (not even ARMv5TE is required). Considering that there are no ARM-specific configs which are not at least ARMv5TE, I did not try to detect these at runtime, and simply enable them for ARMv5 and above. Finally, the NEON register saving code was completely non-reentrant, since it saved the registers to a global, static variable. I moved the storage for this onto the stack. A single binary built with this code was tested on an ARM11 (ARMv6) and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder, and produced identical output, while using the correct accelerated functions on each. I did not test on any earlier processors. Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
2010-10-21 02:39:11 +04:00
#if HAVE_ARMV7
INT64 dx_store_reg[8];
#endif
2010-05-18 19:58:33 +04:00
VP8D_COMP *pbi = (VP8D_COMP *) ptr;
VP8_COMMON *cm = &pbi->common;
int retcode = 0;
struct vpx_usec_timer timer;
/*if(pbi->ready_for_new_data == 0)
return -1;*/
2010-05-18 19:58:33 +04:00
if (ptr == 0)
{
return -1;
}
pbi->common.error.error_code = VPX_CODEC_OK;
Add runtime CPU detection support for ARM. The primary goal is to allow a binary to be built which supports NEON, but can fall back to non-NEON routines, since some Android devices do not have NEON, even if they are otherwise ARMv7 (e.g., Tegra). The configure-generated flags HAVE_ARMV7, etc., are used to decide which versions of each function to build, and when CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen at run time. In order for this to work, the CFLAGS must be set to something appropriate (e.g., without -mfpu=neon for ARMv7, and with appropriate -march and -mcpu for even earlier configurations), or the native C code will not be able to run. The ASFLAGS must remain set for the most advanced instruction set required at build time, since the ARM assembler will refuse to emit them otherwise. I have not attempted to make any changes to configure to do this automatically. Doing so will probably require the addition of new configure options. Many of the hooks for RTCD on ARM were already there, but a lot of the code had bit-rotted, and a good deal of the ARM-specific code is not integrated into the RTCD structs at all. I did not try to resolve the latter, merely to add the minimal amount of protection around them to allow RTCD to work. Those functions that were called based on an ifdef at the calling site were expanded to check the RTCD flags at that site, but they should be added to an RTCD struct somewhere in the future. The functions invoked with global function pointers still are, but these should be moved into an RTCD struct for thread safety (I believe every platform currently supported has atomic pointer stores, but this is not guaranteed). The encoder's boolhuff functions did not even have _c and armv7 suffixes, and the correct version was resolved at link time. The token packing functions did have appropriate suffixes, but the version was selected with a define, with no associated RTCD struct. However, for both of these, the only armv7 instruction they actually used was rbit, and this was completely superfluous, so I reworked them to avoid it. The only non-ARMv4 instruction remaining in them is clz, which is ARMv5 (not even ARMv5TE is required). Considering that there are no ARM-specific configs which are not at least ARMv5TE, I did not try to detect these at runtime, and simply enable them for ARMv5 and above. Finally, the NEON register saving code was completely non-reentrant, since it saved the registers to a global, static variable. I moved the storage for this onto the stack. A single binary built with this code was tested on an ARM11 (ARMv6) and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder, and produced identical output, while using the correct accelerated functions on each. I did not test on any earlier processors. Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
2010-10-21 02:39:11 +04:00
#if HAVE_ARMV7
#if CONFIG_RUNTIME_CPU_DETECT
if (cm->rtcd.flags & HAS_NEON)
#endif
{
vp8_push_neon(dx_store_reg);
}
#endif
cm->new_fb_idx = get_free_fb (cm);
2010-05-18 19:58:33 +04:00
if (setjmp(pbi->common.error.jmp))
{
Add runtime CPU detection support for ARM. The primary goal is to allow a binary to be built which supports NEON, but can fall back to non-NEON routines, since some Android devices do not have NEON, even if they are otherwise ARMv7 (e.g., Tegra). The configure-generated flags HAVE_ARMV7, etc., are used to decide which versions of each function to build, and when CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen at run time. In order for this to work, the CFLAGS must be set to something appropriate (e.g., without -mfpu=neon for ARMv7, and with appropriate -march and -mcpu for even earlier configurations), or the native C code will not be able to run. The ASFLAGS must remain set for the most advanced instruction set required at build time, since the ARM assembler will refuse to emit them otherwise. I have not attempted to make any changes to configure to do this automatically. Doing so will probably require the addition of new configure options. Many of the hooks for RTCD on ARM were already there, but a lot of the code had bit-rotted, and a good deal of the ARM-specific code is not integrated into the RTCD structs at all. I did not try to resolve the latter, merely to add the minimal amount of protection around them to allow RTCD to work. Those functions that were called based on an ifdef at the calling site were expanded to check the RTCD flags at that site, but they should be added to an RTCD struct somewhere in the future. The functions invoked with global function pointers still are, but these should be moved into an RTCD struct for thread safety (I believe every platform currently supported has atomic pointer stores, but this is not guaranteed). The encoder's boolhuff functions did not even have _c and armv7 suffixes, and the correct version was resolved at link time. The token packing functions did have appropriate suffixes, but the version was selected with a define, with no associated RTCD struct. However, for both of these, the only armv7 instruction they actually used was rbit, and this was completely superfluous, so I reworked them to avoid it. The only non-ARMv4 instruction remaining in them is clz, which is ARMv5 (not even ARMv5TE is required). Considering that there are no ARM-specific configs which are not at least ARMv5TE, I did not try to detect these at runtime, and simply enable them for ARMv5 and above. Finally, the NEON register saving code was completely non-reentrant, since it saved the registers to a global, static variable. I moved the storage for this onto the stack. A single binary built with this code was tested on an ARM11 (ARMv6) and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder, and produced identical output, while using the correct accelerated functions on each. I did not test on any earlier processors. Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
2010-10-21 02:39:11 +04:00
#if HAVE_ARMV7
#if CONFIG_RUNTIME_CPU_DETECT
if (cm->rtcd.flags & HAS_NEON)
#endif
{
vp8_pop_neon(dx_store_reg);
}
#endif
2010-05-18 19:58:33 +04:00
pbi->common.error.setjmp = 0;
if (cm->fb_idx_ref_cnt[cm->new_fb_idx] > 0)
cm->fb_idx_ref_cnt[cm->new_fb_idx]--;
2010-05-18 19:58:33 +04:00
return -1;
}
pbi->common.error.setjmp = 1;
vpx_usec_timer_start(&timer);
/*cm->current_video_frame++;*/
2010-05-18 19:58:33 +04:00
pbi->Source = source;
pbi->source_sz = size;
retcode = vp8_decode_frame(pbi);
if (retcode < 0)
{
#if HAVE_ARMV7
Add runtime CPU detection support for ARM. The primary goal is to allow a binary to be built which supports NEON, but can fall back to non-NEON routines, since some Android devices do not have NEON, even if they are otherwise ARMv7 (e.g., Tegra). The configure-generated flags HAVE_ARMV7, etc., are used to decide which versions of each function to build, and when CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen at run time. In order for this to work, the CFLAGS must be set to something appropriate (e.g., without -mfpu=neon for ARMv7, and with appropriate -march and -mcpu for even earlier configurations), or the native C code will not be able to run. The ASFLAGS must remain set for the most advanced instruction set required at build time, since the ARM assembler will refuse to emit them otherwise. I have not attempted to make any changes to configure to do this automatically. Doing so will probably require the addition of new configure options. Many of the hooks for RTCD on ARM were already there, but a lot of the code had bit-rotted, and a good deal of the ARM-specific code is not integrated into the RTCD structs at all. I did not try to resolve the latter, merely to add the minimal amount of protection around them to allow RTCD to work. Those functions that were called based on an ifdef at the calling site were expanded to check the RTCD flags at that site, but they should be added to an RTCD struct somewhere in the future. The functions invoked with global function pointers still are, but these should be moved into an RTCD struct for thread safety (I believe every platform currently supported has atomic pointer stores, but this is not guaranteed). The encoder's boolhuff functions did not even have _c and armv7 suffixes, and the correct version was resolved at link time. The token packing functions did have appropriate suffixes, but the version was selected with a define, with no associated RTCD struct. However, for both of these, the only armv7 instruction they actually used was rbit, and this was completely superfluous, so I reworked them to avoid it. The only non-ARMv4 instruction remaining in them is clz, which is ARMv5 (not even ARMv5TE is required). Considering that there are no ARM-specific configs which are not at least ARMv5TE, I did not try to detect these at runtime, and simply enable them for ARMv5 and above. Finally, the NEON register saving code was completely non-reentrant, since it saved the registers to a global, static variable. I moved the storage for this onto the stack. A single binary built with this code was tested on an ARM11 (ARMv6) and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder, and produced identical output, while using the correct accelerated functions on each. I did not test on any earlier processors. Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
2010-10-21 02:39:11 +04:00
#if CONFIG_RUNTIME_CPU_DETECT
if (cm->rtcd.flags & HAS_NEON)
#endif
{
vp8_pop_neon(dx_store_reg);
}
2010-05-18 19:58:33 +04:00
#endif
pbi->common.error.error_code = VPX_CODEC_ERROR;
pbi->common.error.setjmp = 0;
if (cm->fb_idx_ref_cnt[cm->new_fb_idx] > 0)
cm->fb_idx_ref_cnt[cm->new_fb_idx]--;
2010-05-18 19:58:33 +04:00
return retcode;
}
if (pbi->b_multithreaded_rd && cm->multi_token_partition != ONE_PARTITION)
2010-05-18 19:58:33 +04:00
{
if (swap_frame_buffers (cm))
{
Add runtime CPU detection support for ARM. The primary goal is to allow a binary to be built which supports NEON, but can fall back to non-NEON routines, since some Android devices do not have NEON, even if they are otherwise ARMv7 (e.g., Tegra). The configure-generated flags HAVE_ARMV7, etc., are used to decide which versions of each function to build, and when CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen at run time. In order for this to work, the CFLAGS must be set to something appropriate (e.g., without -mfpu=neon for ARMv7, and with appropriate -march and -mcpu for even earlier configurations), or the native C code will not be able to run. The ASFLAGS must remain set for the most advanced instruction set required at build time, since the ARM assembler will refuse to emit them otherwise. I have not attempted to make any changes to configure to do this automatically. Doing so will probably require the addition of new configure options. Many of the hooks for RTCD on ARM were already there, but a lot of the code had bit-rotted, and a good deal of the ARM-specific code is not integrated into the RTCD structs at all. I did not try to resolve the latter, merely to add the minimal amount of protection around them to allow RTCD to work. Those functions that were called based on an ifdef at the calling site were expanded to check the RTCD flags at that site, but they should be added to an RTCD struct somewhere in the future. The functions invoked with global function pointers still are, but these should be moved into an RTCD struct for thread safety (I believe every platform currently supported has atomic pointer stores, but this is not guaranteed). The encoder's boolhuff functions did not even have _c and armv7 suffixes, and the correct version was resolved at link time. The token packing functions did have appropriate suffixes, but the version was selected with a define, with no associated RTCD struct. However, for both of these, the only armv7 instruction they actually used was rbit, and this was completely superfluous, so I reworked them to avoid it. The only non-ARMv4 instruction remaining in them is clz, which is ARMv5 (not even ARMv5TE is required). Considering that there are no ARM-specific configs which are not at least ARMv5TE, I did not try to detect these at runtime, and simply enable them for ARMv5 and above. Finally, the NEON register saving code was completely non-reentrant, since it saved the registers to a global, static variable. I moved the storage for this onto the stack. A single binary built with this code was tested on an ARM11 (ARMv6) and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder, and produced identical output, while using the correct accelerated functions on each. I did not test on any earlier processors. Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
2010-10-21 02:39:11 +04:00
#if HAVE_ARMV7
#if CONFIG_RUNTIME_CPU_DETECT
if (cm->rtcd.flags & HAS_NEON)
#endif
{
vp8_pop_neon(dx_store_reg);
}
#endif
pbi->common.error.error_code = VPX_CODEC_ERROR;
pbi->common.error.setjmp = 0;
return -1;
}
} else
2010-05-18 19:58:33 +04:00
{
if (swap_frame_buffers (cm))
{
Add runtime CPU detection support for ARM. The primary goal is to allow a binary to be built which supports NEON, but can fall back to non-NEON routines, since some Android devices do not have NEON, even if they are otherwise ARMv7 (e.g., Tegra). The configure-generated flags HAVE_ARMV7, etc., are used to decide which versions of each function to build, and when CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen at run time. In order for this to work, the CFLAGS must be set to something appropriate (e.g., without -mfpu=neon for ARMv7, and with appropriate -march and -mcpu for even earlier configurations), or the native C code will not be able to run. The ASFLAGS must remain set for the most advanced instruction set required at build time, since the ARM assembler will refuse to emit them otherwise. I have not attempted to make any changes to configure to do this automatically. Doing so will probably require the addition of new configure options. Many of the hooks for RTCD on ARM were already there, but a lot of the code had bit-rotted, and a good deal of the ARM-specific code is not integrated into the RTCD structs at all. I did not try to resolve the latter, merely to add the minimal amount of protection around them to allow RTCD to work. Those functions that were called based on an ifdef at the calling site were expanded to check the RTCD flags at that site, but they should be added to an RTCD struct somewhere in the future. The functions invoked with global function pointers still are, but these should be moved into an RTCD struct for thread safety (I believe every platform currently supported has atomic pointer stores, but this is not guaranteed). The encoder's boolhuff functions did not even have _c and armv7 suffixes, and the correct version was resolved at link time. The token packing functions did have appropriate suffixes, but the version was selected with a define, with no associated RTCD struct. However, for both of these, the only armv7 instruction they actually used was rbit, and this was completely superfluous, so I reworked them to avoid it. The only non-ARMv4 instruction remaining in them is clz, which is ARMv5 (not even ARMv5TE is required). Considering that there are no ARM-specific configs which are not at least ARMv5TE, I did not try to detect these at runtime, and simply enable them for ARMv5 and above. Finally, the NEON register saving code was completely non-reentrant, since it saved the registers to a global, static variable. I moved the storage for this onto the stack. A single binary built with this code was tested on an ARM11 (ARMv6) and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder, and produced identical output, while using the correct accelerated functions on each. I did not test on any earlier processors. Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
2010-10-21 02:39:11 +04:00
#if HAVE_ARMV7
#if CONFIG_RUNTIME_CPU_DETECT
if (cm->rtcd.flags & HAS_NEON)
#endif
{
vp8_pop_neon(dx_store_reg);
}
#endif
pbi->common.error.error_code = VPX_CODEC_ERROR;
pbi->common.error.setjmp = 0;
return -1;
}
if(pbi->common.filter_level)
{
struct vpx_usec_timer lpftimer;
vpx_usec_timer_start(&lpftimer);
/* Apply the loop filter if appropriate. */
2010-05-18 19:58:33 +04:00
vp8_loop_filter_frame(cm, &pbi->mb, cm->filter_level);
vpx_usec_timer_mark(&lpftimer);
pbi->time_loop_filtering += vpx_usec_timer_elapsed(&lpftimer);
cm->last_frame_type = cm->frame_type;
cm->last_filter_type = cm->filter_type;
cm->last_sharpness_level = cm->sharpness_level;
}
vp8_yv12_extend_frame_borders_ptr(cm->frame_to_show);
}
2010-05-18 19:58:33 +04:00
#if 0
/* DEBUG code */
/*vp8_recon_write_yuv_frame("recon.yuv", cm->frame_to_show);*/
2010-05-18 19:58:33 +04:00
if (cm->current_video_frame <= 5)
write_dx_frame_to_file(cm->frame_to_show, cm->current_video_frame);
#endif
vp8_clear_system_state();
vpx_usec_timer_mark(&timer);
pbi->decode_microseconds = vpx_usec_timer_elapsed(&timer);
pbi->time_decoding += pbi->decode_microseconds;
/*vp8_print_modes_and_motion_vectors( cm->mi, cm->mb_rows,cm->mb_cols, cm->current_video_frame);*/
2010-05-18 19:58:33 +04:00
if (cm->show_frame)
cm->current_video_frame++;
pbi->ready_for_new_data = 0;
pbi->last_time_stamp = time_stamp;
#if 0
{
int i;
INT64 earliest_time = pbi->dr[0].time_stamp;
INT64 latest_time = pbi->dr[0].time_stamp;
INT64 time_diff = 0;
int bytes = 0;
pbi->dr[pbi->common.current_video_frame&0xf].size = pbi->bc.pos + pbi->bc2.pos + 4;;
pbi->dr[pbi->common.current_video_frame&0xf].time_stamp = time_stamp;
for (i = 0; i < 16; i++)
{
bytes += pbi->dr[i].size;
if (pbi->dr[i].time_stamp < earliest_time)
earliest_time = pbi->dr[i].time_stamp;
if (pbi->dr[i].time_stamp > latest_time)
latest_time = pbi->dr[i].time_stamp;
}
time_diff = latest_time - earliest_time;
if (time_diff > 0)
{
pbi->common.bitrate = 80000.00 * bytes / time_diff ;
pbi->common.framerate = 160000000.00 / time_diff ;
}
}
#endif
#if HAVE_ARMV7
Add runtime CPU detection support for ARM. The primary goal is to allow a binary to be built which supports NEON, but can fall back to non-NEON routines, since some Android devices do not have NEON, even if they are otherwise ARMv7 (e.g., Tegra). The configure-generated flags HAVE_ARMV7, etc., are used to decide which versions of each function to build, and when CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen at run time. In order for this to work, the CFLAGS must be set to something appropriate (e.g., without -mfpu=neon for ARMv7, and with appropriate -march and -mcpu for even earlier configurations), or the native C code will not be able to run. The ASFLAGS must remain set for the most advanced instruction set required at build time, since the ARM assembler will refuse to emit them otherwise. I have not attempted to make any changes to configure to do this automatically. Doing so will probably require the addition of new configure options. Many of the hooks for RTCD on ARM were already there, but a lot of the code had bit-rotted, and a good deal of the ARM-specific code is not integrated into the RTCD structs at all. I did not try to resolve the latter, merely to add the minimal amount of protection around them to allow RTCD to work. Those functions that were called based on an ifdef at the calling site were expanded to check the RTCD flags at that site, but they should be added to an RTCD struct somewhere in the future. The functions invoked with global function pointers still are, but these should be moved into an RTCD struct for thread safety (I believe every platform currently supported has atomic pointer stores, but this is not guaranteed). The encoder's boolhuff functions did not even have _c and armv7 suffixes, and the correct version was resolved at link time. The token packing functions did have appropriate suffixes, but the version was selected with a define, with no associated RTCD struct. However, for both of these, the only armv7 instruction they actually used was rbit, and this was completely superfluous, so I reworked them to avoid it. The only non-ARMv4 instruction remaining in them is clz, which is ARMv5 (not even ARMv5TE is required). Considering that there are no ARM-specific configs which are not at least ARMv5TE, I did not try to detect these at runtime, and simply enable them for ARMv5 and above. Finally, the NEON register saving code was completely non-reentrant, since it saved the registers to a global, static variable. I moved the storage for this onto the stack. A single binary built with this code was tested on an ARM11 (ARMv6) and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder, and produced identical output, while using the correct accelerated functions on each. I did not test on any earlier processors. Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
2010-10-21 02:39:11 +04:00
#if CONFIG_RUNTIME_CPU_DETECT
if (cm->rtcd.flags & HAS_NEON)
#endif
{
vp8_pop_neon(dx_store_reg);
}
2010-05-18 19:58:33 +04:00
#endif
pbi->common.error.setjmp = 0;
return retcode;
}
int vp8dx_get_raw_frame(VP8D_PTR ptr, YV12_BUFFER_CONFIG *sd, INT64 *time_stamp, INT64 *time_end_stamp, int deblock_level, int noise_level, int flags)
{
int ret = -1;
VP8D_COMP *pbi = (VP8D_COMP *) ptr;
if (pbi->ready_for_new_data == 1)
return ret;
/* ie no raw frame to show!!! */
2010-05-18 19:58:33 +04:00
if (pbi->common.show_frame == 0)
return ret;
pbi->ready_for_new_data = 1;
*time_stamp = pbi->last_time_stamp;
*time_end_stamp = 0;
sd->clrtype = pbi->common.clr_type;
#if CONFIG_POSTPROC
ret = vp8_post_proc_frame(&pbi->common, sd, deblock_level, noise_level, flags);
#else
if (pbi->common.frame_to_show)
{
*sd = *pbi->common.frame_to_show;
sd->y_width = pbi->common.Width;
sd->y_height = pbi->common.Height;
sd->uv_height = pbi->common.Height / 2;
ret = 0;
}
else
{
ret = -1;
}
#endif /*!CONFIG_POSTPROC*/
2010-05-18 19:58:33 +04:00
vp8_clear_system_state();
return ret;
}