Merge pull request #9641 from lynlevenick/ash/texture-cache-opt

Remove spurious memory allocations in TextureCacheBase::SerializeTexture and DeserializeTexture
This commit is contained in:
Léo Lam 2021-05-07 15:01:23 +02:00 committed by GitHub
commit 2b632f6d5d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 48 additions and 18 deletions

View File

@ -200,6 +200,16 @@ public:
DoArray(arr, static_cast<u32>(N)); DoArray(arr, static_cast<u32>(N));
} }
// The caller is required to inspect the mode of this PointerWrap
// and deal with the pointer returned from this function themself.
[[nodiscard]] u8* DoExternal(u32& count)
{
Do(count);
u8* current = *ptr;
*ptr += count;
return current;
}
void Do(Common::Flag& flag) void Do(Common::Flag& flag)
{ {
bool s = flag.IsSet(); bool s = flag.IsSet();

View File

@ -438,12 +438,33 @@ void TextureCacheBase::SerializeTexture(AbstractTexture* tex, const TextureConfi
const bool skip_readback = p.GetMode() == PointerWrap::MODE_MEASURE; const bool skip_readback = p.GetMode() == PointerWrap::MODE_MEASURE;
p.DoPOD(config); p.DoPOD(config);
std::vector<u8> texture_data;
if (skip_readback || CheckReadbackTexture(config.width, config.height, config.format)) if (skip_readback || CheckReadbackTexture(config.width, config.height, config.format))
{ {
// Save out each layer of the texture to the staging texture, and then // First, measure the amount of memory needed.
// append it onto the end of the vector. This gives us all the sub-images u32 total_size = 0;
// in one single buffer which can be written out to the save state. for (u32 layer = 0; layer < config.layers; layer++)
{
for (u32 level = 0; level < config.levels; level++)
{
u32 level_width = std::max(config.width >> level, 1u);
u32 level_height = std::max(config.height >> level, 1u);
u32 stride = AbstractTexture::CalculateStrideForFormat(config.format, level_width);
u32 size = stride * level_height;
total_size += size;
}
}
// Set aside total_size bytes of space for the textures.
// When measuring, this will be set aside and not written to,
// but when writing we'll use this pointer directly to avoid
// needing to allocate/free an extra buffer.
u8* texture_data = p.DoExternal(total_size);
if (!skip_readback)
{
// Save out each layer of the texture to the pointer.
for (u32 layer = 0; layer < config.layers; layer++) for (u32 layer = 0; layer < config.layers; layer++)
{ {
for (u32 level = 0; level < config.levels; level++) for (u32 level = 0; level < config.levels; level++)
@ -451,15 +472,14 @@ void TextureCacheBase::SerializeTexture(AbstractTexture* tex, const TextureConfi
u32 level_width = std::max(config.width >> level, 1u); u32 level_width = std::max(config.width >> level, 1u);
u32 level_height = std::max(config.height >> level, 1u); u32 level_height = std::max(config.height >> level, 1u);
auto rect = tex->GetConfig().GetMipRect(level); auto rect = tex->GetConfig().GetMipRect(level);
if (!skip_readback)
m_readback_texture->CopyFromTexture(tex, rect, layer, level, rect); m_readback_texture->CopyFromTexture(tex, rect, layer, level, rect);
size_t stride = AbstractTexture::CalculateStrideForFormat(config.format, level_width); u32 stride = AbstractTexture::CalculateStrideForFormat(config.format, level_width);
size_t size = stride * level_height; u32 size = stride * level_height;
size_t start = texture_data.size(); m_readback_texture->ReadTexels(rect, texture_data, stride);
texture_data.resize(texture_data.size() + size);
if (!skip_readback) texture_data += size;
m_readback_texture->ReadTexels(rect, &texture_data[start], static_cast<u32>(stride)); }
} }
} }
} }
@ -467,8 +487,6 @@ void TextureCacheBase::SerializeTexture(AbstractTexture* tex, const TextureConfi
{ {
PanicAlertFmt("Failed to create staging texture for serialization"); PanicAlertFmt("Failed to create staging texture for serialization");
} }
p.Do(texture_data);
} }
std::optional<TextureCacheBase::TexPoolEntry> TextureCacheBase::DeserializeTexture(PointerWrap& p) std::optional<TextureCacheBase::TexPoolEntry> TextureCacheBase::DeserializeTexture(PointerWrap& p)
@ -476,10 +494,12 @@ std::optional<TextureCacheBase::TexPoolEntry> TextureCacheBase::DeserializeTextu
TextureConfig config; TextureConfig config;
p.Do(config); p.Do(config);
std::vector<u8> texture_data; // Read in the size from the save state, then texture data will point to
p.Do(texture_data); // a region of size total_size where textures are stored.
u32 total_size = 0;
u8* texture_data = p.DoExternal(total_size);
if (p.GetMode() != PointerWrap::MODE_READ || texture_data.empty()) if (p.GetMode() != PointerWrap::MODE_READ || total_size == 0)
return std::nullopt; return std::nullopt;
auto tex = AllocateTexture(config); auto tex = AllocateTexture(config);
@ -498,7 +518,7 @@ std::optional<TextureCacheBase::TexPoolEntry> TextureCacheBase::DeserializeTextu
const u32 level_height = std::max(config.height >> level, 1u); const u32 level_height = std::max(config.height >> level, 1u);
const size_t stride = AbstractTexture::CalculateStrideForFormat(config.format, level_width); const size_t stride = AbstractTexture::CalculateStrideForFormat(config.format, level_width);
const size_t size = stride * level_height; const size_t size = stride * level_height;
if ((start + size) > texture_data.size()) if ((start + size) > total_size)
{ {
ERROR_LOG_FMT(VIDEO, "Insufficient texture data for layer {} level {}", layer, level); ERROR_LOG_FMT(VIDEO, "Insufficient texture data for layer {} level {}", layer, level);
return tex; return tex;