0
0
mirror of https://github.com/mpv-player/mpv.git synced 2024-09-20 12:02:23 +02:00
mpv/video/mp_image_pool.c
Philip Langdale 872b068cb7 f_hwtransfer: disable vulkan multiplane images when uploading from cuda
Although we can support vulkan multiplane images, cuda lacks any such
support, and so cannot natively import such images for interop. It's
possible that we can do separate exports for each plane in the image
and have it work, but for now, we can selectively disable multiplane
when we know that we'll be consuming cuda frames.

As a reminder, even though cuda is the frame source, interop is one way
so the vulkan images have to be imported to cuda before we copy the
frame contents over.

This logic here is slightly more complex than I'd like but you can't
just set the flag blindly, as it will cause hwframes ctx creation to
fail if the format is packed or if it's planar rgb. Oh well.
2023-05-28 15:46:05 -07:00

473 lines
14 KiB
C

/*
* This file is part of mpv.
*
* mpv is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* mpv is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with mpv. If not, see <http://www.gnu.org/licenses/>.
*/
#include "config.h"
#include <stddef.h>
#include <stdbool.h>
#include <pthread.h>
#include <assert.h>
#include <libavutil/buffer.h>
#include <libavutil/hwcontext.h>
#if HAVE_VULKAN_INTEROP
#include <libavutil/hwcontext_vulkan.h>
#endif
#include <libavutil/mem.h>
#include <libavutil/pixdesc.h>
#include "mpv_talloc.h"
#include "common/common.h"
#include "fmt-conversion.h"
#include "mp_image.h"
#include "mp_image_pool.h"
static pthread_mutex_t pool_mutex = PTHREAD_MUTEX_INITIALIZER;
#define pool_lock() pthread_mutex_lock(&pool_mutex)
#define pool_unlock() pthread_mutex_unlock(&pool_mutex)
// Thread-safety: the pool itself is not thread-safe, but pool-allocated images
// can be referenced and unreferenced from other threads. (As long as the image
// destructors are thread-safe.)
struct mp_image_pool {
struct mp_image **images;
int num_images;
int fmt, w, h;
mp_image_allocator allocator;
void *allocator_ctx;
bool use_lru;
unsigned int lru_counter;
};
// Used to gracefully handle the case when the pool is freed while image
// references allocated from the image pool are still held by someone.
struct image_flags {
// If both of these are false, the image must be freed.
bool referenced; // outside mp_image reference exists
bool pool_alive; // the mp_image_pool references this
unsigned int order; // for LRU allocation (basically a timestamp)
};
static void image_pool_destructor(void *ptr)
{
struct mp_image_pool *pool = ptr;
mp_image_pool_clear(pool);
}
// If tparent!=NULL, set it as talloc parent for the pool.
struct mp_image_pool *mp_image_pool_new(void *tparent)
{
struct mp_image_pool *pool = talloc_ptrtype(tparent, pool);
talloc_set_destructor(pool, image_pool_destructor);
*pool = (struct mp_image_pool) {0};
return pool;
}
void mp_image_pool_clear(struct mp_image_pool *pool)
{
for (int n = 0; n < pool->num_images; n++) {
struct mp_image *img = pool->images[n];
struct image_flags *it = img->priv;
bool referenced;
pool_lock();
assert(it->pool_alive);
it->pool_alive = false;
referenced = it->referenced;
pool_unlock();
if (!referenced)
talloc_free(img);
}
pool->num_images = 0;
}
// This is the only function that is allowed to run in a different thread.
// (Consider passing an image to another thread, which frees it.)
static void unref_image(void *opaque, uint8_t *data)
{
struct mp_image *img = opaque;
struct image_flags *it = img->priv;
bool alive;
pool_lock();
assert(it->referenced);
it->referenced = false;
alive = it->pool_alive;
pool_unlock();
if (!alive)
talloc_free(img);
}
// Return a new image of given format/size. Unlike mp_image_pool_get(), this
// returns NULL if there is no free image of this format/size.
struct mp_image *mp_image_pool_get_no_alloc(struct mp_image_pool *pool, int fmt,
int w, int h)
{
struct mp_image *new = NULL;
pool_lock();
for (int n = 0; n < pool->num_images; n++) {
struct mp_image *img = pool->images[n];
struct image_flags *img_it = img->priv;
assert(img_it->pool_alive);
if (!img_it->referenced) {
if (img->imgfmt == fmt && img->w == w && img->h == h) {
if (pool->use_lru) {
struct image_flags *new_it = new ? new->priv : NULL;
if (!new_it || new_it->order > img_it->order)
new = img;
} else {
new = img;
break;
}
}
}
}
pool_unlock();
if (!new)
return NULL;
// Reference the new image. Since mp_image_pool is not declared thread-safe,
// and unreffing images from other threads does not allocate new images,
// no synchronization is required here.
for (int p = 0; p < MP_MAX_PLANES; p++)
assert(!!new->bufs[p] == !p); // only 1 AVBufferRef
struct mp_image *ref = mp_image_new_dummy_ref(new);
// This assumes the buffer is at this point exclusively owned by us: we
// can't track whether the buffer is unique otherwise.
// (av_buffer_is_writable() checks the refcount of the new buffer only.)
int flags = av_buffer_is_writable(new->bufs[0]) ? 0 : AV_BUFFER_FLAG_READONLY;
ref->bufs[0] = av_buffer_create(new->bufs[0]->data, new->bufs[0]->size,
unref_image, new, flags);
if (!ref->bufs[0]) {
talloc_free(ref);
return NULL;
}
struct image_flags *it = new->priv;
assert(!it->referenced && it->pool_alive);
it->referenced = true;
it->order = ++pool->lru_counter;
return ref;
}
void mp_image_pool_add(struct mp_image_pool *pool, struct mp_image *new)
{
struct image_flags *it = talloc_ptrtype(new, it);
*it = (struct image_flags) { .pool_alive = true };
new->priv = it;
MP_TARRAY_APPEND(pool, pool->images, pool->num_images, new);
}
// Return a new image of given format/size. The only difference to
// mp_image_alloc() is that there is a transparent mechanism to recycle image
// data allocations through this pool.
// If pool==NULL, mp_image_alloc() is called (for convenience).
// The image can be free'd with talloc_free().
// Returns NULL on OOM.
struct mp_image *mp_image_pool_get(struct mp_image_pool *pool, int fmt,
int w, int h)
{
if (!pool)
return mp_image_alloc(fmt, w, h);
struct mp_image *new = mp_image_pool_get_no_alloc(pool, fmt, w, h);
if (!new) {
if (fmt != pool->fmt || w != pool->w || h != pool->h)
mp_image_pool_clear(pool);
pool->fmt = fmt;
pool->w = w;
pool->h = h;
if (pool->allocator) {
new = pool->allocator(pool->allocator_ctx, fmt, w, h);
} else {
new = mp_image_alloc(fmt, w, h);
}
if (!new)
return NULL;
mp_image_pool_add(pool, new);
new = mp_image_pool_get_no_alloc(pool, fmt, w, h);
}
return new;
}
// Like mp_image_new_copy(), but allocate the image out of the pool.
// If pool==NULL, a plain copy is made (for convenience).
// Returns NULL on OOM.
struct mp_image *mp_image_pool_new_copy(struct mp_image_pool *pool,
struct mp_image *img)
{
struct mp_image *new = mp_image_pool_get(pool, img->imgfmt, img->w, img->h);
if (new) {
mp_image_copy(new, img);
mp_image_copy_attributes(new, img);
}
return new;
}
// Like mp_image_make_writeable(), but if a copy has to be made, allocate it
// out of the pool.
// If pool==NULL, mp_image_make_writeable() is called (for convenience).
// Returns false on failure (see mp_image_make_writeable()).
bool mp_image_pool_make_writeable(struct mp_image_pool *pool,
struct mp_image *img)
{
if (mp_image_is_writeable(img))
return true;
struct mp_image *new = mp_image_pool_new_copy(pool, img);
if (!new)
return false;
mp_image_steal_data(img, new);
assert(mp_image_is_writeable(img));
return true;
}
// Call cb(cb_data, fmt, w, h) to allocate an image. Note that the resulting
// image must use only 1 AVBufferRef. The returned image must also be owned
// exclusively by the image pool, otherwise mp_image_is_writeable() will not
// work due to FFmpeg restrictions.
void mp_image_pool_set_allocator(struct mp_image_pool *pool,
mp_image_allocator cb, void *cb_data)
{
pool->allocator = cb;
pool->allocator_ctx = cb_data;
}
// Put into LRU mode. (Likely better for hwaccel surfaces, but worse for memory.)
void mp_image_pool_set_lru(struct mp_image_pool *pool)
{
pool->use_lru = true;
}
// Return the sw image format mp_image_hw_download() would use. This can be
// different from src->params.hw_subfmt in obscure cases.
int mp_image_hw_download_get_sw_format(struct mp_image *src)
{
if (!src->hwctx)
return 0;
// Try to find the first format which we can apparently use.
int imgfmt = 0;
enum AVPixelFormat *fmts;
if (av_hwframe_transfer_get_formats(src->hwctx,
AV_HWFRAME_TRANSFER_DIRECTION_FROM, &fmts, 0) < 0)
return 0;
for (int n = 0; fmts[n] != AV_PIX_FMT_NONE; n++) {
imgfmt = pixfmt2imgfmt(fmts[n]);
if (imgfmt)
break;
}
av_free(fmts);
return imgfmt;
}
// Copies the contents of the HW surface src to system memory and returns it.
// If swpool is not NULL, it's used to allocate the target image.
// src must be a hw surface with a AVHWFramesContext attached.
// The returned image is cropped as needed.
// Returns NULL on failure.
struct mp_image *mp_image_hw_download(struct mp_image *src,
struct mp_image_pool *swpool)
{
int imgfmt = mp_image_hw_download_get_sw_format(src);
if (!imgfmt)
return NULL;
assert(src->hwctx);
AVHWFramesContext *fctx = (void *)src->hwctx->data;
struct mp_image *dst =
mp_image_pool_get(swpool, imgfmt, fctx->width, fctx->height);
if (!dst)
return NULL;
// Target image must be writable, so unref it.
AVFrame *dstav = mp_image_to_av_frame_and_unref(dst);
if (!dstav)
return NULL;
AVFrame *srcav = mp_image_to_av_frame(src);
if (!srcav) {
av_frame_unref(dstav);
return NULL;
}
int res = av_hwframe_transfer_data(dstav, srcav, 0);
av_frame_free(&srcav);
dst = mp_image_from_av_frame(dstav);
av_frame_free(&dstav);
if (res >= 0 && dst) {
mp_image_set_size(dst, src->w, src->h);
mp_image_copy_attributes(dst, src);
} else {
mp_image_unrefp(&dst);
}
return dst;
}
bool mp_image_hw_upload(struct mp_image *hw_img, struct mp_image *src)
{
if (hw_img->w != src->w || hw_img->h != src->h)
return false;
if (!hw_img->hwctx)
return false;
bool ok = false;
AVFrame *dstav = NULL;
AVFrame *srcav = NULL;
// This means the destination image will not be "writable", which would be
// a pain if Libav enforced this - fortunately it doesn't care. We can
// transfer data to it even if there are multiple refs.
dstav = mp_image_to_av_frame(hw_img);
if (!dstav)
goto done;
srcav = mp_image_to_av_frame(src);
if (!srcav)
goto done;
ok = av_hwframe_transfer_data(dstav, srcav, 0) >= 0;
done:
av_frame_free(&srcav);
av_frame_free(&dstav);
if (ok)
mp_image_copy_attributes(hw_img, src);
return ok;
}
bool mp_update_av_hw_frames_pool(struct AVBufferRef **hw_frames_ctx,
struct AVBufferRef *hw_device_ctx,
int imgfmt, int sw_imgfmt, int w, int h,
bool disable_multiplane)
{
enum AVPixelFormat format = imgfmt2pixfmt(imgfmt);
enum AVPixelFormat sw_format = imgfmt2pixfmt(sw_imgfmt);
if (format == AV_PIX_FMT_NONE || sw_format == AV_PIX_FMT_NONE ||
!hw_device_ctx || w < 1 || h < 1)
{
av_buffer_unref(hw_frames_ctx);
return false;
}
if (*hw_frames_ctx) {
AVHWFramesContext *hw_frames = (void *)(*hw_frames_ctx)->data;
if (hw_frames->device_ref->data != hw_device_ctx->data ||
hw_frames->format != format || hw_frames->sw_format != sw_format ||
hw_frames->width != w || hw_frames->height != h)
av_buffer_unref(hw_frames_ctx);
}
if (!*hw_frames_ctx) {
*hw_frames_ctx = av_hwframe_ctx_alloc(hw_device_ctx);
if (!*hw_frames_ctx)
return false;
AVHWFramesContext *hw_frames = (void *)(*hw_frames_ctx)->data;
hw_frames->format = format;
hw_frames->sw_format = sw_format;
hw_frames->width = w;
hw_frames->height = h;
#if HAVE_VULKAN_INTEROP
if (format == AV_PIX_FMT_VULKAN && disable_multiplane) {
const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sw_format);
if ((desc->flags & AV_PIX_FMT_FLAG_PLANAR) &&
!(desc->flags & AV_PIX_FMT_FLAG_RGB)) {
AVVulkanFramesContext *vk_frames = hw_frames->hwctx;
vk_frames->flags = AV_VK_FRAME_FLAG_DISABLE_MULTIPLANE;
}
}
#endif
if (av_hwframe_ctx_init(*hw_frames_ctx) < 0) {
av_buffer_unref(hw_frames_ctx);
return false;
}
}
return true;
}
struct mp_image *mp_av_pool_image_hw_upload(struct AVBufferRef *hw_frames_ctx,
struct mp_image *src)
{
AVFrame *av_frame = av_frame_alloc();
if (!av_frame)
return NULL;
if (av_hwframe_get_buffer(hw_frames_ctx, av_frame, 0) < 0) {
av_frame_free(&av_frame);
return NULL;
}
struct mp_image *dst = mp_image_from_av_frame(av_frame);
av_frame_free(&av_frame);
if (!dst)
return NULL;
if (dst->w < src->w || dst->h < src->h) {
talloc_free(dst);
return NULL;
}
mp_image_set_size(dst, src->w, src->h);
if (!mp_image_hw_upload(dst, src)) {
talloc_free(dst);
return NULL;
}
mp_image_copy_attributes(dst, src);
return dst;
}
struct mp_image *mp_av_pool_image_hw_map(struct AVBufferRef *hw_frames_ctx,
struct mp_image *src)
{
AVFrame *dst_frame = av_frame_alloc();
if (!dst_frame)
return NULL;
dst_frame->format = ((AVHWFramesContext*)hw_frames_ctx->data)->format;
dst_frame->hw_frames_ctx = av_buffer_ref(hw_frames_ctx);
AVFrame *src_frame = mp_image_to_av_frame(src);
if (av_hwframe_map(dst_frame, src_frame, 0) < 0) {
av_frame_free(&src_frame);
av_frame_free(&dst_frame);
return NULL;
}
av_frame_free(&src_frame);
struct mp_image *dst = mp_image_from_av_frame(dst_frame);
av_frame_free(&dst_frame);
if (!dst)
return NULL;
mp_image_copy_attributes(dst, src);
return dst;
}