mirrors/rpcs3
0
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-08-02 22:30:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

gl: Refactoring and fallback support for compute acceleration

Browse source
This commit is contained in:
kd-11 2019-10-02 13:39:30 +03:00 • committed by kd-11
commit d9a9766e41
2 changed files with 102 additions and 93 deletions
Download patch file Download diff file Expand all files Collapse all files

10
rpcs3/Emu/RSX/GL/GLHelpers.h
View file

@ -114,6 +114,7 @@ namespace gl
bool NV_texture_barrier_supported = false; bool NV_texture_barrier_supported = false;
bool NV_gpu_shader5_supported = false; bool NV_gpu_shader5_supported = false;
bool AMD_gpu_shader_half_float_supported = false; bool AMD_gpu_shader_half_float_supported = false;
bool ARB_compute_shader_supported = false;
bool initialized = false; bool initialized = false;
bool vendor_INTEL = false; // has broken GLSL compiler bool vendor_INTEL = false; // has broken GLSL compiler
bool vendor_AMD = false; // has broken ARB_multidraw bool vendor_AMD = false; // has broken ARB_multidraw
@ -133,7 +134,7 @@ namespace gl
void initialize() void initialize()
{ {
int find_count = 10; int find_count = 11;
int ext_count = 0; int ext_count = 0;
glGetIntegerv(GL_NUM_EXTENSIONS, &ext_count); glGetIntegerv(GL_NUM_EXTENSIONS, &ext_count);
@ -212,6 +213,13 @@ namespace gl
find_count--; find_count--;
continue; continue;
} }
if (check(ext_name, "GL_ARB_compute_shader"))
{
ARB_compute_shader_supported = true;
find_count--;
continue;
}
} }
// Workaround for intel drivers which have terrible capability reporting // Workaround for intel drivers which have terrible capability reporting

185
rpcs3/Emu/RSX/GL/GLTexture.cpp
View file

@ -738,116 +738,117 @@ namespace gl
return false; return false;
} }
cs_shuffle_base* get_pixel_transform_job(const pixel_buffer_layout& pack_info)
{
const bool is_depth_stencil = (pack_info.type == GL_UNSIGNED_INT_24_8);
if (LIKELY(!is_depth_stencil))
{
if (!pack_info.swap_bytes)
{
return nullptr;
}
switch (pack_info.size)
{
case 1:
return nullptr;
case 2:
return gl::get_compute_task<gl::cs_shuffle_16>();
break;
case 4:
return gl::get_compute_task<gl::cs_shuffle_32>();
break;
default:
fmt::throw_exception("Unsupported format");
}
}
else
{
if (pack_info.swap_bytes)
{
return gl::get_compute_task<gl::cs_shuffle_d24x8_to_x8d24<true>>();
}
else
{
return gl::get_compute_task<gl::cs_shuffle_d24x8_to_x8d24<false>>();
}
}
}
void copy_typeless(texture * dst, const texture * src) void copy_typeless(texture * dst, const texture * src)
{ {
GLsizeiptr src_mem = src->width() * src->height(); GLsizeiptr src_mem = src->pitch() * src->height();
GLsizeiptr dst_mem = dst->width() * dst->height(); GLsizeiptr dst_mem = dst->pitch() * dst->height();
auto max_mem = std::max(src_mem, dst_mem) * 16; auto max_mem = std::max(src_mem, dst_mem);
if (!g_typeless_transfer_buffer || max_mem > g_typeless_transfer_buffer.size()) if (!g_typeless_transfer_buffer || max_mem > g_typeless_transfer_buffer.size())
{ {
if (g_typeless_transfer_buffer) g_typeless_transfer_buffer.remove(); if (g_typeless_transfer_buffer) g_typeless_transfer_buffer.remove();
g_typeless_transfer_buffer.create(buffer::target::pixel_pack, max_mem, nullptr, buffer::memory_type::local, GL_STATIC_COPY); g_typeless_transfer_buffer.create(buffer::target::pixel_pack, max_mem, nullptr, buffer::memory_type::local, GL_STATIC_COPY);
} }
const auto& caps = gl::get_driver_caps();
const auto pack_info = get_format_type(src->get_internal_format()); const auto pack_info = get_format_type(src->get_internal_format());
const auto unpack_info = get_format_type(dst->get_internal_format()); const auto unpack_info = get_format_type(dst->get_internal_format());
pixel_pack_settings pack_settings{}; // Start pack operation
g_typeless_transfer_buffer.bind(buffer::target::pixel_pack); g_typeless_transfer_buffer.bind(buffer::target::pixel_pack);
src->copy_to(nullptr, (texture::format)pack_info.format, (texture::type)pack_info.type, pack_settings);
glBindBuffer(GL_PIXEL_PACK_BUFFER, GL_NONE);
const bool src_is_ds = !!(src->aspect() & gl::image_aspect::stencil); if (LIKELY(caps.ARB_compute_shader_supported))
const bool dst_is_ds = !!(src->aspect() & gl::image_aspect::stencil);
if (pack_info.swap_bytes || unpack_info.swap_bytes || src_is_ds || dst_is_ds)
{ {
gl::cs_shuffle_base *src_transform = nullptr, *dst_transform = nullptr; // Raw copy
src->copy_to(nullptr, (texture::format)pack_info.format, (texture::type)pack_info.type);
if (src_is_ds) }
{ else
if (pack_info.swap_bytes) {
{ pixel_pack_settings pack_settings{};
src_transform = gl::get_compute_task<gl::cs_shuffle_d24x8_to_x8d24<true>>(); pack_settings.swap_bytes(pack_info.swap_bytes);
} src->copy_to(nullptr, (texture::format)pack_info.format, (texture::type)pack_info.type, pack_settings);
else
{
src_transform = gl::get_compute_task<gl::cs_shuffle_d24x8_to_x8d24<false>>();
}
}
else if (pack_info.swap_bytes)
{
switch (pack_info.size)
{
case 1:
break;
case 2:
src_transform = gl::get_compute_task<gl::cs_shuffle_16>();
break;
case 4:
src_transform = gl::get_compute_task<gl::cs_shuffle_32>();
break;
default:
fmt::throw_exception("Unsupported format");
}
}
if (dst_is_ds)
{
if (unpack_info.swap_bytes)
{
dst_transform = gl::get_compute_task<gl::cs_shuffle_x8d24_to_d24x8<true>>();
}
else
{
dst_transform = gl::get_compute_task<gl::cs_shuffle_x8d24_to_d24x8<false>>();
}
}
else if (unpack_info.swap_bytes)
{
switch (unpack_info.size)
{
case 1:
break;
case 2:
dst_transform = gl::get_compute_task<gl::cs_shuffle_16>();
break;
case 4:
dst_transform = gl::get_compute_task<gl::cs_shuffle_32>();
break;
default:
fmt::throw_exception("Unsupported format");
}
if (!src_is_ds)
{
if (src_transform == dst_transform)
{
src_transform = dst_transform = nullptr;
}
else if (src_transform)
{
src_transform = gl::get_compute_task<gl::cs_shuffle_32_16>();
dst_transform = nullptr;
}
}
if (src_transform)
{
const auto image_size = src->pitch() * src->height();
src_transform->run(&g_typeless_transfer_buffer, image_size);
}
if (dst_transform)
{
const auto image_size = dst->pitch() * dst->height();
dst_transform->run(&g_typeless_transfer_buffer, image_size);
}
}
} }
glBindBuffer(GL_PIXEL_PACK_BUFFER, GL_NONE);
// Start unpack operation
pixel_unpack_settings unpack_settings{}; pixel_unpack_settings unpack_settings{};
if (LIKELY(caps.ARB_compute_shader_supported))
{
auto src_transform = get_pixel_transform_job(pack_info);
auto dst_transform = get_pixel_transform_job(unpack_info);
if (src->aspect() == gl::image_aspect::color && dst->aspect() == gl::image_aspect::color)
{
if (src_transform == dst_transform)
{
src_transform = dst_transform = nullptr;
}
else if (src_transform && dst_transform)
{
src_transform = gl::get_compute_task<cs_shuffle_32_16>();
dst_transform = nullptr;
}
}
const auto job_length = std::min(src_mem, dst_mem);
if (src_transform)
{
src_transform->run(&g_typeless_transfer_buffer, job_length);
}
if (dst_transform)
{
dst_transform->run(&g_typeless_transfer_buffer, job_length);
}
// NOTE: glBindBufferRange also binds the buffer to the old-school target.
// Unbind it to avoid glitching later
glBindBuffer(GL_SHADER_STORAGE_BUFFER, GL_NONE);
}
else
{
unpack_settings.swap_bytes(unpack_info.swap_bytes);
}
g_typeless_transfer_buffer.bind(buffer::target::pixel_unpack); g_typeless_transfer_buffer.bind(buffer::target::pixel_unpack);
dst->copy_from(nullptr, (texture::format)unpack_info.format, (texture::type)unpack_info.type, unpack_settings); dst->copy_from(nullptr, (texture::format)unpack_info.format, (texture::type)unpack_info.type, unpack_settings);
glBindBuffer(GL_PIXEL_UNPACK_BUFFER, GL_NONE); glBindBuffer(GL_PIXEL_UNPACK_BUFFER, GL_NONE);