@ -36,10 +36,7 @@ static bool IsPassThroughTevStage(const Pica::Regs::TevStageConfig& stage) {
stage . GetAlphaMultiplier ( ) = = 1 ) ;
}
RasterizerOpenGL : : RasterizerOpenGL ( ) : cached_fb_color_addr ( 0 ) , cached_fb_depth_addr ( 0 ) { }
RasterizerOpenGL : : ~ RasterizerOpenGL ( ) { }
void RasterizerOpenGL : : InitObjects ( ) {
RasterizerOpenGL : : RasterizerOpenGL ( ) : shader_dirty ( true ) {
// Create sampler objects
for ( size_t i = 0 ; i < texture_samplers . size ( ) ; + + i ) {
texture_samplers [ i ] . Create ( ) ;
@ -61,6 +58,10 @@ void RasterizerOpenGL::InitObjects() {
uniform_block_data . dirty = true ;
for ( unsigned index = 0 ; index < lighting_luts . size ( ) ; index + + ) {
uniform_block_data . lut_dirty [ index ] = true ;
}
// Set vertex attributes
glVertexAttribPointer ( GLShader : : ATTRIBUTE_POSITION , 4 , GL_FLOAT , GL_FALSE , sizeof ( HardwareVertex ) , ( GLvoid * ) offsetof ( HardwareVertex , position ) ) ;
glEnableVertexAttribArray ( GLShader : : ATTRIBUTE_POSITION ) ;
@ -81,70 +82,24 @@ void RasterizerOpenGL::InitObjects() {
glVertexAttribPointer ( GLShader : : ATTRIBUTE_VIEW , 3 , GL_FLOAT , GL_FALSE , sizeof ( HardwareVertex ) , ( GLvoid * ) offsetof ( HardwareVertex , view ) ) ;
glEnableVertexAttribArray ( GLShader : : ATTRIBUTE_VIEW ) ;
SetShader ( ) ;
// Create textures for OGL framebuffer that will be rendered to, initially 1x1 to succeed in framebuffer creation
fb_color_texture . texture . Create ( ) ;
ReconfigureColorTexture ( fb_color_texture , Pica : : Regs : : ColorFormat : : RGBA8 , 1 , 1 ) ;
state . texture_units [ 0 ] . texture_2d = fb_color_texture . texture . handle ;
state . Apply ( ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_MAX_LEVEL , 0 ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_MAG_FILTER , GL_NEAREST ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_MIN_FILTER , GL_NEAREST ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_WRAP_S , GL_CLAMP_TO_EDGE ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_WRAP_T , GL_CLAMP_TO_EDGE ) ;
state . texture_units [ 0 ] . texture_2d = 0 ;
state . Apply ( ) ;
fb_depth_texture . texture . Create ( ) ;
ReconfigureDepthTexture ( fb_depth_texture , Pica : : Regs : : DepthFormat : : D16 , 1 , 1 ) ;
state . texture_units [ 0 ] . texture_2d = fb_depth_texture . texture . handle ;
state . Apply ( ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_MAX_LEVEL , 0 ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_MAG_FILTER , GL_NEAREST ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_MIN_FILTER , GL_NEAREST ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_WRAP_S , GL_CLAMP_TO_EDGE ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_WRAP_T , GL_CLAMP_TO_EDGE ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_COMPARE_FUNC , GL_LEQUAL ) ;
glTexParameteri ( GL_TEXTURE_2D , GL_TEXTURE_COMPARE_MODE , GL_NONE ) ;
state . texture_units [ 0 ] . texture_2d = 0 ;
state . Apply ( ) ;
// Configure OpenGL framebuffer
// Create render framebuffer
framebuffer . Create ( ) ;
state . draw . framebuffer = framebuffer . handle ;
// Allocate and bind lighting lut textures
for ( size_t i = 0 ; i < lighting_luts . size ( ) ; + + i ) {
lighting_luts [ i ] . Create ( ) ;
state . lighting_luts [ i ] . texture_1d = lighting_luts [ i ] . handle ;
}
state . Apply ( ) ;
glActiveTexture ( GL_TEXTURE0 ) ;
glFramebufferTexture2D ( GL_FRAMEBUFFER , GL_COLOR_ATTACHMENT0 , GL_TEXTURE_2D , fb_color_texture . texture . handle , 0 ) ;
glFramebufferTexture2D ( GL_FRAMEBUFFER , GL_DEPTH_ATTACHMENT , GL_TEXTURE_2D , fb_depth_texture . texture . handle , 0 ) ;
for ( size_t i = 0 ; i < lighting_lut . size ( ) ; + + i ) {
lighting_lut [ i ] . Create ( ) ;
state . lighting_lut [ i ] . texture_1d = lighting_lut [ i ] . handle ;
for ( size_t i = 0 ; i < lighting_luts . size ( ) ; + + i ) {
glActiveTexture ( GL_TEXTURE3 + i ) ;
glBindTexture ( GL_TEXTURE_1D , state . lighting_lut [ i ] . texture_1d ) ;
glTexImage1D ( GL_TEXTURE_1D , 0 , GL_RGBA32F , 256 , 0 , GL_RGBA , GL_FLOAT , nullptr ) ;
glTexParameteri ( GL_TEXTURE_1D , GL_TEXTURE_MIN_FILTER , GL_LINEAR ) ;
glTexParameteri ( GL_TEXTURE_1D , GL_TEXTURE_MAG_FILTER , GL_LINEAR ) ;
}
state . Apply ( ) ;
GLenum status = glCheckFramebufferStatus ( GL_FRAMEBUFFER ) ;
ASSERT_MSG ( status = = GL_FRAMEBUFFER_COMPLETE ,
" OpenGL rasterizer framebuffer setup failed, status %X " , status ) ;
}
void RasterizerOpenGL : : Reset ( ) {
// Sync fixed function OpenGL state
SyncCullMode ( ) ;
SyncDepthModifiers ( ) ;
SyncBlendEnabled ( ) ;
@ -156,10 +111,10 @@ void RasterizerOpenGL::Reset() {
SyncColorWriteMask ( ) ;
SyncStencilWriteMask ( ) ;
SyncDepthWriteMask ( ) ;
}
SetShader ( ) ;
RasterizerOpenGL : : ~ RasterizerOpenGL ( ) {
res_cache . InvalidateAll ( ) ;
}
/**
@ -196,47 +151,98 @@ void RasterizerOpenGL::DrawTriangles() {
if ( vertex_batch . empty ( ) )
return ;
SyncFramebuffer ( ) ;
SyncDrawState ( ) ;
const auto & regs = Pica : : g_state . regs ;
if ( state . draw . shader_dirty ) {
SetShader ( ) ;
state . draw . shader_dirty = false ;
// Sync and bind the framebuffer surfaces
CachedSurface * color_surface ;
CachedSurface * depth_surface ;
MathUtil : : Rectangle < int > rect ;
std : : tie ( color_surface , depth_surface , rect ) = res_cache . GetFramebufferSurfaces ( regs . framebuffer ) ;
state . draw . draw_framebuffer = framebuffer . handle ;
state . Apply ( ) ;
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_COLOR_ATTACHMENT0 , GL_TEXTURE_2D , color_surface ! = nullptr ? color_surface - > texture . handle : 0 , 0 ) ;
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_DEPTH_ATTACHMENT , GL_TEXTURE_2D , depth_surface ! = nullptr ? depth_surface - > texture . handle : 0 , 0 ) ;
bool has_stencil = regs . framebuffer . depth_format = = Pica : : Regs : : DepthFormat : : D24S8 ;
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_STENCIL_ATTACHMENT , GL_TEXTURE_2D , ( has_stencil & & depth_surface ! = nullptr ) ? depth_surface - > texture . handle : 0 , 0 ) ;
if ( OpenGLState : : CheckFBStatus ( GL_DRAW_FRAMEBUFFER ) ! = GL_FRAMEBUFFER_COMPLETE ) {
return ;
}
for ( unsigned index = 0 ; index < lighting_lut . size ( ) ; index + + ) {
// Sync the viewport
// These registers hold half-width and half-height, so must be multiplied by 2
GLsizei viewport_width = ( GLsizei ) Pica : : float24 : : FromRaw ( regs . viewport_size_x ) . ToFloat32 ( ) * 2 ;
GLsizei viewport_height = ( GLsizei ) Pica : : float24 : : FromRaw ( regs . viewport_size_y ) . ToFloat32 ( ) * 2 ;
glViewport ( ( GLint ) ( rect . left + regs . viewport_corner . x * color_surface - > res_scale_width ) ,
( GLint ) ( rect . bottom + regs . viewport_corner . y * color_surface - > res_scale_height ) ,
( GLsizei ) ( viewport_width * color_surface - > res_scale_width ) , ( GLsizei ) ( viewport_height * color_surface - > res_scale_height ) ) ;
// Sync and bind the texture surfaces
const auto pica_textures = regs . GetTextures ( ) ;
for ( unsigned texture_index = 0 ; texture_index < pica_textures . size ( ) ; + + texture_index ) {
const auto & texture = pica_textures [ texture_index ] ;
if ( texture . enabled ) {
texture_samplers [ texture_index ] . SyncWithConfig ( texture . config ) ;
CachedSurface * surface = res_cache . GetTextureSurface ( texture ) ;
if ( surface ! = nullptr ) {
state . texture_units [ texture_index ] . texture_2d = surface - > texture . handle ;
} else {
// Can occur when texture addr is null or its memory is unmapped/invalid
state . texture_units [ texture_index ] . texture_2d = 0 ;
}
} else {
state . texture_units [ texture_index ] . texture_2d = 0 ;
}
}
// Sync and bind the shader
if ( shader_dirty ) {
SetShader ( ) ;
shader_dirty = false ;
}
// Sync the lighting luts
for ( unsigned index = 0 ; index < lighting_luts . size ( ) ; index + + ) {
if ( uniform_block_data . lut_dirty [ index ] ) {
SyncLightingLUT ( index ) ;
uniform_block_data . lut_dirty [ index ] = false ;
}
}
// Sync the uniform data
if ( uniform_block_data . dirty ) {
glBufferData ( GL_UNIFORM_BUFFER , sizeof ( UniformData ) , & uniform_block_data . data , GL_STATIC_DRAW ) ;
uniform_block_data . dirty = false ;
}
state . Apply ( ) ;
// Draw the vertex batch
glBufferData ( GL_ARRAY_BUFFER , vertex_batch . size ( ) * sizeof ( HardwareVertex ) , vertex_batch . data ( ) , GL_STREAM_DRAW ) ;
glDrawArrays ( GL_TRIANGLES , 0 , ( GLsizei ) vertex_batch . size ( ) ) ;
// Mark framebuffer surfaces as dirty
// TODO: Restrict invalidation area to the viewport
if ( color_surface ! = nullptr ) {
color_surface - > dirty = true ;
res_cache . FlushRegion ( color_surface - > addr , color_surface - > size , color_surface , true ) ;
}
if ( depth_surface ! = nullptr ) {
depth_surface - > dirty = true ;
res_cache . FlushRegion ( depth_surface - > addr , depth_surface - > size , depth_surface , true ) ;
}
vertex_batch . clear ( ) ;
// Flush the resource cache at the current depth and color framebuffer addresses for render-to-texture
const auto & regs = Pica : : g_state . regs ;
u32 cached_fb_color_size = Pica : : Regs : : BytesPerColorPixel ( fb_color_texture . format )
* fb_color_texture . width * fb_color_texture . height ;
u32 cached_fb_depth_size = Pica : : Regs : : BytesPerDepthPixel ( fb_depth_texture . format )
* fb_depth_texture . width * fb_depth_texture . height ;
res_cache . InvalidateInRange ( cached_fb_color_addr , cached_fb_color_size , true ) ;
res_cache . InvalidateInRange ( cached_fb_depth_addr , cached_fb_depth_size , true ) ;
}
void RasterizerOpenGL : : FlushFramebuffer ( ) {
CommitColorBuffer ( ) ;
CommitDepthBuffer ( ) ;
// Unbind textures for potential future use as framebuffer attachments
for ( unsigned texture_index = 0 ; texture_index < pica_textures . size ( ) ; + + texture_index ) {
state . texture_units [ texture_index ] . texture_2d = 0 ;
}
state . Apply ( ) ;
}
void RasterizerOpenGL : : NotifyPicaRegisterChanged ( u32 id ) {
@ -268,7 +274,7 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
// Alpha test
case PICA_REG_INDEX ( output_merger . alpha_test ) :
SyncAlphaTest ( ) ;
s tate. draw . s hader_dirty = true ;
s hader_dirty = true ;
break ;
// Sync GL stencil test + stencil write mask
@ -334,7 +340,7 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
case PICA_REG_INDEX ( tev_stage5 . color_op ) :
case PICA_REG_INDEX ( tev_stage5 . color_scale ) :
case PICA_REG_INDEX ( tev_combiner_buffer_input ) :
s tate. draw . s hader_dirty = true ;
s hader_dirty = true ;
break ;
case PICA_REG_INDEX ( tev_stage0 . const_r ) :
SyncTevConstColor ( 0 , regs . tev_stage0 ) ;
@ -521,41 +527,257 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
}
}
void RasterizerOpenGL : : FlushRegion ( PAddr addr , u32 size ) {
const auto & regs = Pica : : g_state . regs ;
u32 cached_fb_color_size = Pica : : Regs : : BytesPerColorPixel ( fb_color_texture . format )
* fb_color_texture . width * fb_color_texture . height ;
u32 cached_fb_depth_size = Pica : : Regs : : BytesPerDepthPixel ( fb_depth_texture . format )
* fb_depth_texture . width * fb_depth_texture . height ;
// If source memory region overlaps 3DS framebuffers, commit them before the copy happens
if ( MathUtil : : IntervalsIntersect ( addr , size , cached_fb_color_addr , cached_fb_color_size ) )
CommitColorBuffer ( ) ;
if ( MathUtil : : IntervalsIntersect ( addr , size , cached_fb_depth_addr , cached_fb_depth_size ) )
CommitDepthBuffer ( ) ;
void RasterizerOpenGL : : FlushAll ( ) {
res_cache . FlushAll ( ) ;
}
void RasterizerOpenGL : : InvalidateRegion ( PAddr addr , u32 size ) {
const auto & regs = Pica : : g_state . regs ;
void RasterizerOpenGL : : FlushRegion ( PAddr addr , u32 size ) {
res_cache . FlushRegion ( addr , size , nullptr , false ) ;
}
u32 cached_fb_color_size = Pica : : Regs : : BytesPerColorPixel ( fb_color_texture . format )
* fb_color_texture . width * fb_color_texture . height ;
void RasterizerOpenGL : : FlushAndInvalidateRegion ( PAddr addr , u32 size ) {
res_cache . FlushRegion ( addr , size , nullptr , true ) ;
}
u32 cached_fb_depth_size = Pica : : Regs : : BytesPerDepthPixel ( fb_depth_texture . format )
* fb_depth_texture . width * fb_depth_texture . height ;
bool RasterizerOpenGL : : AccelerateDisplayTransfer ( const GPU : : Regs : : DisplayTransferConfig & config ) {
using PixelFormat = CachedSurface : : PixelFormat ;
using SurfaceType = CachedSurface : : SurfaceType ;
// If modified memory region overlaps 3DS framebuffers, reload their contents into OpenGL
if ( MathUtil : : IntervalsIntersect ( addr , size , cached_fb_color_addr , cached_fb_color_size ) )
ReloadColorBuffer ( ) ;
if ( config . is_texture_copy ) {
// TODO(tfarley): Try to hardware accelerate this
return false ;
}
if ( MathUtil : : IntervalsIntersect ( addr , size , cached_fb_depth_addr , cached_fb_depth_size ) )
ReloadDepthBuffer ( ) ;
CachedSurface src_params ;
src_params . addr = config . GetPhysicalInputAddress ( ) ;
src_params . width = config . output_width ;
src_params . height = config . output_height ;
src_params . is_tiled = ! config . input_linear ;
src_params . pixel_format = CachedSurface : : PixelFormatFromGPUPixelFormat ( config . input_format ) ;
// Notify cache of flush in case the region touches a cached resource
res_cache . InvalidateInRange ( addr , size ) ;
CachedSurface dst_params ;
dst_params . addr = config . GetPhysicalOutputAddress ( ) ;
dst_params . width = config . scaling ! = config . NoScale ? config . output_width / 2 : config . output_width . Value ( ) ;
dst_params . height = config . scaling = = config . ScaleXY ? config . output_height / 2 : config . output_height . Value ( ) ;
dst_params . is_tiled = config . input_linear ! = config . dont_swizzle ;
dst_params . pixel_format = CachedSurface : : PixelFormatFromGPUPixelFormat ( config . output_format ) ;
MathUtil : : Rectangle < int > src_rect ;
CachedSurface * src_surface = res_cache . GetSurfaceRect ( src_params , false , true , src_rect ) ;
if ( src_surface = = nullptr ) {
return false ;
}
// Require destination surface to have same resolution scale as source to preserve scaling
dst_params . res_scale_width = src_surface - > res_scale_width ;
dst_params . res_scale_height = src_surface - > res_scale_height ;
MathUtil : : Rectangle < int > dst_rect ;
CachedSurface * dst_surface = res_cache . GetSurfaceRect ( dst_params , true , false , dst_rect ) ;
if ( dst_surface = = nullptr ) {
return false ;
}
// Don't accelerate if the src and dst surfaces are the same
if ( src_surface = = dst_surface ) {
return false ;
}
if ( config . flip_vertically ) {
std : : swap ( dst_rect . top , dst_rect . bottom ) ;
}
if ( ! res_cache . TryBlitSurfaces ( src_surface , src_rect , dst_surface , dst_rect ) ) {
return false ;
}
u32 dst_size = dst_params . width * dst_params . height * CachedSurface : : GetFormatBpp ( dst_params . pixel_format ) / 8 ;
dst_surface - > dirty = true ;
res_cache . FlushRegion ( config . GetPhysicalOutputAddress ( ) , dst_size , dst_surface , true ) ;
return true ;
}
bool RasterizerOpenGL : : AccelerateFill ( const GPU : : Regs : : MemoryFillConfig & config ) {
using PixelFormat = CachedSurface : : PixelFormat ;
using SurfaceType = CachedSurface : : SurfaceType ;
CachedSurface * dst_surface = res_cache . TryGetFillSurface ( config ) ;
if ( dst_surface = = nullptr ) {
return false ;
}
OpenGLState cur_state = OpenGLState : : GetCurState ( ) ;
SurfaceType dst_type = CachedSurface : : GetFormatType ( dst_surface - > pixel_format ) ;
GLuint old_fb = cur_state . draw . draw_framebuffer ;
cur_state . draw . draw_framebuffer = framebuffer . handle ;
// TODO: When scissor test is implemented, need to disable scissor test in cur_state here so Clear call isn't affected
cur_state . Apply ( ) ;
if ( dst_type = = SurfaceType : : Color | | dst_type = = SurfaceType : : Texture ) {
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_COLOR_ATTACHMENT0 , GL_TEXTURE_2D , dst_surface - > texture . handle , 0 ) ;
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_DEPTH_STENCIL_ATTACHMENT , GL_TEXTURE_2D , 0 , 0 ) ;
if ( OpenGLState : : CheckFBStatus ( GL_DRAW_FRAMEBUFFER ) ! = GL_FRAMEBUFFER_COMPLETE ) {
return false ;
}
GLfloat color_values [ 4 ] = { 0.0f , 0.0f , 0.0f , 0.0f } ;
// TODO: Handle additional pixel format and fill value size combinations to accelerate more cases
// For instance, checking if fill value's bytes/bits repeat to allow filling I8/A8/I4/A4/...
// Currently only handles formats that are multiples of the fill value size
if ( config . fill_24bit ) {
switch ( dst_surface - > pixel_format ) {
case PixelFormat : : RGB8 :
color_values [ 0 ] = config . value_24bit_r / 255.0f ;
color_values [ 1 ] = config . value_24bit_g / 255.0f ;
color_values [ 2 ] = config . value_24bit_b / 255.0f ;
break ;
default :
return false ;
}
} else if ( config . fill_32bit ) {
u32 value = config . value_32bit ;
switch ( dst_surface - > pixel_format ) {
case PixelFormat : : RGBA8 :
color_values [ 0 ] = ( value > > 24 ) / 255.0f ;
color_values [ 1 ] = ( ( value > > 16 ) & 0xFF ) / 255.0f ;
color_values [ 2 ] = ( ( value > > 8 ) & 0xFF ) / 255.0f ;
color_values [ 3 ] = ( value & 0xFF ) / 255.0f ;
break ;
default :
return false ;
}
} else {
u16 value_16bit = config . value_16bit . Value ( ) ;
Math : : Vec4 < u8 > color ;
switch ( dst_surface - > pixel_format ) {
case PixelFormat : : RGBA8 :
color_values [ 0 ] = ( value_16bit > > 8 ) / 255.0f ;
color_values [ 1 ] = ( value_16bit & 0xFF ) / 255.0f ;
color_values [ 2 ] = color_values [ 0 ] ;
color_values [ 3 ] = color_values [ 1 ] ;
break ;
case PixelFormat : : RGB5A1 :
color = Color : : DecodeRGB5A1 ( ( const u8 * ) & value_16bit ) ;
color_values [ 0 ] = color [ 0 ] / 31.0f ;
color_values [ 1 ] = color [ 1 ] / 31.0f ;
color_values [ 2 ] = color [ 2 ] / 31.0f ;
color_values [ 3 ] = color [ 3 ] ;
break ;
case PixelFormat : : RGB565 :
color = Color : : DecodeRGB565 ( ( const u8 * ) & value_16bit ) ;
color_values [ 0 ] = color [ 0 ] / 31.0f ;
color_values [ 1 ] = color [ 1 ] / 63.0f ;
color_values [ 2 ] = color [ 2 ] / 31.0f ;
break ;
case PixelFormat : : RGBA4 :
color = Color : : DecodeRGBA4 ( ( const u8 * ) & value_16bit ) ;
color_values [ 0 ] = color [ 0 ] / 15.0f ;
color_values [ 1 ] = color [ 1 ] / 15.0f ;
color_values [ 2 ] = color [ 2 ] / 15.0f ;
color_values [ 3 ] = color [ 3 ] / 15.0f ;
break ;
case PixelFormat : : IA8 :
case PixelFormat : : RG8 :
color_values [ 0 ] = ( value_16bit > > 8 ) / 255.0f ;
color_values [ 1 ] = ( value_16bit & 0xFF ) / 255.0f ;
break ;
default :
return false ;
}
}
cur_state . color_mask . red_enabled = true ;
cur_state . color_mask . green_enabled = true ;
cur_state . color_mask . blue_enabled = true ;
cur_state . color_mask . alpha_enabled = true ;
cur_state . Apply ( ) ;
glClearBufferfv ( GL_COLOR , 0 , color_values ) ;
} else if ( dst_type = = SurfaceType : : Depth ) {
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_COLOR_ATTACHMENT0 , GL_TEXTURE_2D , 0 , 0 ) ;
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_DEPTH_ATTACHMENT , GL_TEXTURE_2D , dst_surface - > texture . handle , 0 ) ;
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_STENCIL_ATTACHMENT , GL_TEXTURE_2D , 0 , 0 ) ;
if ( OpenGLState : : CheckFBStatus ( GL_DRAW_FRAMEBUFFER ) ! = GL_FRAMEBUFFER_COMPLETE ) {
return false ;
}
GLfloat value_float ;
if ( dst_surface - > pixel_format = = CachedSurface : : PixelFormat : : D16 ) {
value_float = config . value_32bit / 65535.0f ; // 2^16 - 1
} else if ( dst_surface - > pixel_format = = CachedSurface : : PixelFormat : : D24 ) {
value_float = config . value_32bit / 16777215.0f ; // 2^24 - 1
}
cur_state . depth . write_mask = true ;
cur_state . Apply ( ) ;
glClearBufferfv ( GL_DEPTH , 0 , & value_float ) ;
} else if ( dst_type = = SurfaceType : : DepthStencil ) {
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_COLOR_ATTACHMENT0 , GL_TEXTURE_2D , 0 , 0 ) ;
glFramebufferTexture2D ( GL_DRAW_FRAMEBUFFER , GL_DEPTH_STENCIL_ATTACHMENT , GL_TEXTURE_2D , dst_surface - > texture . handle , 0 ) ;
if ( OpenGLState : : CheckFBStatus ( GL_DRAW_FRAMEBUFFER ) ! = GL_FRAMEBUFFER_COMPLETE ) {
return false ;
}
GLfloat value_float = ( config . value_32bit & 0xFFFFFF ) / 16777215.0f ; // 2^24 - 1
GLint value_int = ( config . value_32bit > > 24 ) ;
cur_state . depth . write_mask = true ;
cur_state . stencil . write_mask = true ;
cur_state . Apply ( ) ;
glClearBufferfi ( GL_DEPTH_STENCIL , 0 , value_float , value_int ) ;
}
cur_state . draw . draw_framebuffer = old_fb ;
// TODO: Return scissor test to previous value when scissor test is implemented
cur_state . Apply ( ) ;
dst_surface - > dirty = true ;
res_cache . FlushRegion ( dst_surface - > addr , dst_surface - > size , dst_surface , true ) ;
return true ;
}
bool RasterizerOpenGL : : AccelerateDisplay ( const GPU : : Regs : : FramebufferConfig & config , PAddr framebuffer_addr , u32 pixel_stride , ScreenInfo & screen_info ) {
if ( framebuffer_addr = = 0 ) {
return false ;
}
CachedSurface src_params ;
src_params . addr = framebuffer_addr ;
src_params . width = config . width ;
src_params . height = config . height ;
src_params . stride = pixel_stride ;
src_params . is_tiled = false ;
src_params . pixel_format = CachedSurface : : PixelFormatFromGPUPixelFormat ( config . color_format ) ;
MathUtil : : Rectangle < int > src_rect ;
CachedSurface * src_surface = res_cache . GetSurfaceRect ( src_params , false , true , src_rect ) ;
if ( src_surface = = nullptr ) {
return false ;
}
u32 scaled_width = src_surface - > GetScaledWidth ( ) ;
u32 scaled_height = src_surface - > GetScaledHeight ( ) ;
screen_info . display_texcoords = MathUtil : : Rectangle < float > ( ( float ) src_rect . top / ( float ) scaled_height ,
( float ) src_rect . left / ( float ) scaled_width ,
( float ) src_rect . bottom / ( float ) scaled_height ,
( float ) src_rect . right / ( float ) scaled_width ) ;
screen_info . display_texture = src_surface - > texture . handle ;
return true ;
}
void RasterizerOpenGL : : SamplerInfo : : Create ( ) {
@ -597,108 +819,6 @@ void RasterizerOpenGL::SamplerInfo::SyncWithConfig(const Pica::Regs::TextureConf
}
}
void RasterizerOpenGL : : ReconfigureColorTexture ( TextureInfo & texture , Pica : : Regs : : ColorFormat format , u32 width , u32 height ) {
GLint internal_format ;
texture . format = format ;
texture . width = width ;
texture . height = height ;
switch ( format ) {
case Pica : : Regs : : ColorFormat : : RGBA8 :
internal_format = GL_RGBA ;
texture . gl_format = GL_RGBA ;
texture . gl_type = GL_UNSIGNED_INT_8_8_8_8 ;
break ;
case Pica : : Regs : : ColorFormat : : RGB8 :
// This pixel format uses BGR since GL_UNSIGNED_BYTE specifies byte-order, unlike every
// specific OpenGL type used in this function using native-endian (that is, little-endian
// mostly everywhere) for words or half-words.
// TODO: check how those behave on big-endian processors.
internal_format = GL_RGB ;
texture . gl_format = GL_BGR ;
texture . gl_type = GL_UNSIGNED_BYTE ;
break ;
case Pica : : Regs : : ColorFormat : : RGB5A1 :
internal_format = GL_RGBA ;
texture . gl_format = GL_RGBA ;
texture . gl_type = GL_UNSIGNED_SHORT_5_5_5_1 ;
break ;
case Pica : : Regs : : ColorFormat : : RGB565 :
internal_format = GL_RGB ;
texture . gl_format = GL_RGB ;
texture . gl_type = GL_UNSIGNED_SHORT_5_6_5 ;
break ;
case Pica : : Regs : : ColorFormat : : RGBA4 :
internal_format = GL_RGBA ;
texture . gl_format = GL_RGBA ;
texture . gl_type = GL_UNSIGNED_SHORT_4_4_4_4 ;
break ;
default :
LOG_CRITICAL ( Render_OpenGL , " Unknown framebuffer texture color format %x " , format ) ;
UNIMPLEMENTED ( ) ;
break ;
}
state . texture_units [ 0 ] . texture_2d = texture . texture . handle ;
state . Apply ( ) ;
glActiveTexture ( GL_TEXTURE0 ) ;
glTexImage2D ( GL_TEXTURE_2D , 0 , internal_format , texture . width , texture . height , 0 ,
texture . gl_format , texture . gl_type , nullptr ) ;
state . texture_units [ 0 ] . texture_2d = 0 ;
state . Apply ( ) ;
}
void RasterizerOpenGL : : ReconfigureDepthTexture ( DepthTextureInfo & texture , Pica : : Regs : : DepthFormat format , u32 width , u32 height ) {
GLint internal_format ;
texture . format = format ;
texture . width = width ;
texture . height = height ;
switch ( format ) {
case Pica : : Regs : : DepthFormat : : D16 :
internal_format = GL_DEPTH_COMPONENT16 ;
texture . gl_format = GL_DEPTH_COMPONENT ;
texture . gl_type = GL_UNSIGNED_SHORT ;
break ;
case Pica : : Regs : : DepthFormat : : D24 :
internal_format = GL_DEPTH_COMPONENT24 ;
texture . gl_format = GL_DEPTH_COMPONENT ;
texture . gl_type = GL_UNSIGNED_INT ;
break ;
case Pica : : Regs : : DepthFormat : : D24S8 :
internal_format = GL_DEPTH24_STENCIL8 ;
texture . gl_format = GL_DEPTH_STENCIL ;
texture . gl_type = GL_UNSIGNED_INT_24_8 ;
break ;
default :
LOG_CRITICAL ( Render_OpenGL , " Unknown framebuffer texture depth format %x " , format ) ;
UNIMPLEMENTED ( ) ;
break ;
}
state . texture_units [ 0 ] . texture_2d = texture . texture . handle ;
state . Apply ( ) ;
glActiveTexture ( GL_TEXTURE0 ) ;
glTexImage2D ( GL_TEXTURE_2D , 0 , internal_format , texture . width , texture . height , 0 ,
texture . gl_format , texture . gl_type , nullptr ) ;
state . texture_units [ 0 ] . texture_2d = 0 ;
state . Apply ( ) ;
}
void RasterizerOpenGL : : SetShader ( ) {
PicaShaderConfig config = PicaShaderConfig : : CurrentConfig ( ) ;
std : : unique_ptr < PicaShader > shader = std : : make_unique < PicaShader > ( ) ;
@ -761,83 +881,6 @@ void RasterizerOpenGL::SetShader() {
}
}
void RasterizerOpenGL : : SyncFramebuffer ( ) {
const auto & regs = Pica : : g_state . regs ;
PAddr new_fb_color_addr = regs . framebuffer . GetColorBufferPhysicalAddress ( ) ;
Pica : : Regs : : ColorFormat new_fb_color_format = regs . framebuffer . color_format ;
PAddr new_fb_depth_addr = regs . framebuffer . GetDepthBufferPhysicalAddress ( ) ;
Pica : : Regs : : DepthFormat new_fb_depth_format = regs . framebuffer . depth_format ;
bool fb_size_changed = fb_color_texture . width ! = static_cast < GLsizei > ( regs . framebuffer . GetWidth ( ) ) | |
fb_color_texture . height ! = static_cast < GLsizei > ( regs . framebuffer . GetHeight ( ) ) ;
bool color_fb_prop_changed = fb_color_texture . format ! = new_fb_color_format | |
fb_size_changed ;
bool depth_fb_prop_changed = fb_depth_texture . format ! = new_fb_depth_format | |
fb_size_changed ;
bool color_fb_modified = cached_fb_color_addr ! = new_fb_color_addr | |
color_fb_prop_changed ;
bool depth_fb_modified = cached_fb_depth_addr ! = new_fb_depth_addr | |
depth_fb_prop_changed ;
// Commit if framebuffer modified in any way
if ( color_fb_modified )
CommitColorBuffer ( ) ;
if ( depth_fb_modified )
CommitDepthBuffer ( ) ;
// Reconfigure framebuffer textures if any property has changed
if ( color_fb_prop_changed ) {
ReconfigureColorTexture ( fb_color_texture , new_fb_color_format ,
regs . framebuffer . GetWidth ( ) , regs . framebuffer . GetHeight ( ) ) ;
}
if ( depth_fb_prop_changed ) {
ReconfigureDepthTexture ( fb_depth_texture , new_fb_depth_format ,
regs . framebuffer . GetWidth ( ) , regs . framebuffer . GetHeight ( ) ) ;
// Only attach depth buffer as stencil if it supports stencil
switch ( new_fb_depth_format ) {
case Pica : : Regs : : DepthFormat : : D16 :
case Pica : : Regs : : DepthFormat : : D24 :
glFramebufferTexture2D ( GL_FRAMEBUFFER , GL_STENCIL_ATTACHMENT , GL_TEXTURE_2D , 0 , 0 ) ;
break ;
case Pica : : Regs : : DepthFormat : : D24S8 :
glFramebufferTexture2D ( GL_FRAMEBUFFER , GL_STENCIL_ATTACHMENT , GL_TEXTURE_2D , fb_depth_texture . texture . handle , 0 ) ;
break ;
default :
LOG_CRITICAL ( Render_OpenGL , " Unknown framebuffer depth format %x " , new_fb_depth_format ) ;
UNIMPLEMENTED ( ) ;
break ;
}
}
// Load buffer data again if fb modified in any way
if ( color_fb_modified ) {
cached_fb_color_addr = new_fb_color_addr ;
ReloadColorBuffer ( ) ;
}
if ( depth_fb_modified ) {
cached_fb_depth_addr = new_fb_depth_addr ;
ReloadDepthBuffer ( ) ;
}
GLenum status = glCheckFramebufferStatus ( GL_FRAMEBUFFER ) ;
ASSERT_MSG ( status = = GL_FRAMEBUFFER_COMPLETE ,
" OpenGL rasterizer framebuffer setup failed, status %X " , status ) ;
}
void RasterizerOpenGL : : SyncCullMode ( ) {
const auto & regs = Pica : : g_state . regs ;
@ -1034,229 +1077,3 @@ void RasterizerOpenGL::SyncLightPosition(int light_index) {
uniform_block_data . dirty = true ;
}
}
void RasterizerOpenGL : : SyncDrawState ( ) {
const auto & regs = Pica : : g_state . regs ;
// Sync the viewport
GLsizei viewport_width = ( GLsizei ) Pica : : float24 : : FromRaw ( regs . viewport_size_x ) . ToFloat32 ( ) * 2 ;
GLsizei viewport_height = ( GLsizei ) Pica : : float24 : : FromRaw ( regs . viewport_size_y ) . ToFloat32 ( ) * 2 ;
// OpenGL uses different y coordinates, so negate corner offset and flip origin
// TODO: Ensure viewport_corner.x should not be negated or origin flipped
// TODO: Use floating-point viewports for accuracy if supported
glViewport ( ( GLsizei ) regs . viewport_corner . x ,
( GLsizei ) regs . viewport_corner . y ,
viewport_width , viewport_height ) ;
// Sync bound texture(s), upload if not cached
const auto pica_textures = regs . GetTextures ( ) ;
for ( unsigned texture_index = 0 ; texture_index < pica_textures . size ( ) ; + + texture_index ) {
const auto & texture = pica_textures [ texture_index ] ;
if ( texture . enabled ) {
texture_samplers [ texture_index ] . SyncWithConfig ( texture . config ) ;
res_cache . LoadAndBindTexture ( state , texture_index , texture ) ;
} else {
state . texture_units [ texture_index ] . texture_2d = 0 ;
}
}
state . draw . uniform_buffer = uniform_buffer . handle ;
state . Apply ( ) ;
}
MICROPROFILE_DEFINE ( OpenGL_FramebufferReload , " OpenGL " , " FB Reload " , MP_RGB ( 70 , 70 , 200 ) ) ;
void RasterizerOpenGL : : ReloadColorBuffer ( ) {
u8 * color_buffer = Memory : : GetPhysicalPointer ( cached_fb_color_addr ) ;
if ( color_buffer = = nullptr )
return ;
MICROPROFILE_SCOPE ( OpenGL_FramebufferReload ) ;
u32 bytes_per_pixel = Pica : : Regs : : BytesPerColorPixel ( fb_color_texture . format ) ;
std : : unique_ptr < u8 [ ] > temp_fb_color_buffer ( new u8 [ fb_color_texture . width * fb_color_texture . height * bytes_per_pixel ] ) ;
// Directly copy pixels. Internal OpenGL color formats are consistent so no conversion is necessary.
for ( int y = 0 ; y < fb_color_texture . height ; + + y ) {
for ( int x = 0 ; x < fb_color_texture . width ; + + x ) {
const u32 coarse_y = y & ~ 7 ;
u32 dst_offset = VideoCore : : GetMortonOffset ( x , y , bytes_per_pixel ) + coarse_y * fb_color_texture . width * bytes_per_pixel ;
u32 gl_pixel_index = ( x + ( fb_color_texture . height - 1 - y ) * fb_color_texture . width ) * bytes_per_pixel ;
u8 * pixel = color_buffer + dst_offset ;
memcpy ( & temp_fb_color_buffer [ gl_pixel_index ] , pixel , bytes_per_pixel ) ;
}
}
state . texture_units [ 0 ] . texture_2d = fb_color_texture . texture . handle ;
state . Apply ( ) ;
glActiveTexture ( GL_TEXTURE0 ) ;
glTexSubImage2D ( GL_TEXTURE_2D , 0 , 0 , 0 , fb_color_texture . width , fb_color_texture . height ,
fb_color_texture . gl_format , fb_color_texture . gl_type , temp_fb_color_buffer . get ( ) ) ;
state . texture_units [ 0 ] . texture_2d = 0 ;
state . Apply ( ) ;
}
void RasterizerOpenGL : : ReloadDepthBuffer ( ) {
if ( cached_fb_depth_addr = = 0 )
return ;
// TODO: Appears to work, but double-check endianness of depth values and order of depth-stencil
u8 * depth_buffer = Memory : : GetPhysicalPointer ( cached_fb_depth_addr ) ;
if ( depth_buffer = = nullptr )
return ;
MICROPROFILE_SCOPE ( OpenGL_FramebufferReload ) ;
u32 bytes_per_pixel = Pica : : Regs : : BytesPerDepthPixel ( fb_depth_texture . format ) ;
// OpenGL needs 4 bpp alignment for D24
u32 gl_bpp = bytes_per_pixel = = 3 ? 4 : bytes_per_pixel ;
std : : unique_ptr < u8 [ ] > temp_fb_depth_buffer ( new u8 [ fb_depth_texture . width * fb_depth_texture . height * gl_bpp ] ) ;
u8 * temp_fb_depth_data = bytes_per_pixel = = 3 ? ( temp_fb_depth_buffer . get ( ) + 1 ) : temp_fb_depth_buffer . get ( ) ;
if ( fb_depth_texture . format = = Pica : : Regs : : DepthFormat : : D24S8 ) {
for ( int y = 0 ; y < fb_depth_texture . height ; + + y ) {
for ( int x = 0 ; x < fb_depth_texture . width ; + + x ) {
const u32 coarse_y = y & ~ 7 ;
u32 dst_offset = VideoCore : : GetMortonOffset ( x , y , bytes_per_pixel ) + coarse_y * fb_depth_texture . width * bytes_per_pixel ;
u32 gl_pixel_index = ( x + ( fb_depth_texture . height - 1 - y ) * fb_depth_texture . width ) ;
u8 * pixel = depth_buffer + dst_offset ;
u32 depth_stencil = * ( u32 * ) pixel ;
( ( u32 * ) temp_fb_depth_data ) [ gl_pixel_index ] = ( depth_stencil < < 8 ) | ( depth_stencil > > 24 ) ;
}
}
} else {
for ( int y = 0 ; y < fb_depth_texture . height ; + + y ) {
for ( int x = 0 ; x < fb_depth_texture . width ; + + x ) {
const u32 coarse_y = y & ~ 7 ;
u32 dst_offset = VideoCore : : GetMortonOffset ( x , y , bytes_per_pixel ) + coarse_y * fb_depth_texture . width * bytes_per_pixel ;
u32 gl_pixel_index = ( x + ( fb_depth_texture . height - 1 - y ) * fb_depth_texture . width ) * gl_bpp ;
u8 * pixel = depth_buffer + dst_offset ;
memcpy ( & temp_fb_depth_data [ gl_pixel_index ] , pixel , bytes_per_pixel ) ;
}
}
}
state . texture_units [ 0 ] . texture_2d = fb_depth_texture . texture . handle ;
state . Apply ( ) ;
glActiveTexture ( GL_TEXTURE0 ) ;
if ( fb_depth_texture . format = = Pica : : Regs : : DepthFormat : : D24S8 ) {
// TODO(Subv): There is a bug with Intel Windows drivers that makes glTexSubImage2D not change the stencil buffer.
// The bug has been reported to Intel (https://communities.intel.com/message/324464)
glTexImage2D ( GL_TEXTURE_2D , 0 , GL_DEPTH24_STENCIL8 , fb_depth_texture . width , fb_depth_texture . height , 0 ,
GL_DEPTH_STENCIL , GL_UNSIGNED_INT_24_8 , temp_fb_depth_buffer . get ( ) ) ;
} else {
glTexSubImage2D ( GL_TEXTURE_2D , 0 , 0 , 0 , fb_depth_texture . width , fb_depth_texture . height ,
fb_depth_texture . gl_format , fb_depth_texture . gl_type , temp_fb_depth_buffer . get ( ) ) ;
}
state . texture_units [ 0 ] . texture_2d = 0 ;
state . Apply ( ) ;
}
Common : : Profiling : : TimingCategory buffer_commit_category ( " Framebuffer Commit " ) ;
MICROPROFILE_DEFINE ( OpenGL_FramebufferCommit , " OpenGL " , " FB Commit " , MP_RGB ( 70 , 70 , 200 ) ) ;
void RasterizerOpenGL : : CommitColorBuffer ( ) {
if ( cached_fb_color_addr ! = 0 ) {
u8 * color_buffer = Memory : : GetPhysicalPointer ( cached_fb_color_addr ) ;
if ( color_buffer ! = nullptr ) {
Common : : Profiling : : ScopeTimer timer ( buffer_commit_category ) ;
MICROPROFILE_SCOPE ( OpenGL_FramebufferCommit ) ;
u32 bytes_per_pixel = Pica : : Regs : : BytesPerColorPixel ( fb_color_texture . format ) ;
std : : unique_ptr < u8 [ ] > temp_gl_color_buffer ( new u8 [ fb_color_texture . width * fb_color_texture . height * bytes_per_pixel ] ) ;
state . texture_units [ 0 ] . texture_2d = fb_color_texture . texture . handle ;
state . Apply ( ) ;
glActiveTexture ( GL_TEXTURE0 ) ;
glGetTexImage ( GL_TEXTURE_2D , 0 , fb_color_texture . gl_format , fb_color_texture . gl_type , temp_gl_color_buffer . get ( ) ) ;
state . texture_units [ 0 ] . texture_2d = 0 ;
state . Apply ( ) ;
// Directly copy pixels. Internal OpenGL color formats are consistent so no conversion is necessary.
for ( int y = 0 ; y < fb_color_texture . height ; + + y ) {
for ( int x = 0 ; x < fb_color_texture . width ; + + x ) {
const u32 coarse_y = y & ~ 7 ;
u32 dst_offset = VideoCore : : GetMortonOffset ( x , y , bytes_per_pixel ) + coarse_y * fb_color_texture . width * bytes_per_pixel ;
u32 gl_pixel_index = x * bytes_per_pixel + ( fb_color_texture . height - 1 - y ) * fb_color_texture . width * bytes_per_pixel ;
u8 * pixel = color_buffer + dst_offset ;
memcpy ( pixel , & temp_gl_color_buffer [ gl_pixel_index ] , bytes_per_pixel ) ;
}
}
}
}
}
void RasterizerOpenGL : : CommitDepthBuffer ( ) {
if ( cached_fb_depth_addr ! = 0 ) {
// TODO: Output seems correct visually, but doesn't quite match sw renderer output. One of them is wrong.
u8 * depth_buffer = Memory : : GetPhysicalPointer ( cached_fb_depth_addr ) ;
if ( depth_buffer ! = nullptr ) {
Common : : Profiling : : ScopeTimer timer ( buffer_commit_category ) ;
MICROPROFILE_SCOPE ( OpenGL_FramebufferCommit ) ;
u32 bytes_per_pixel = Pica : : Regs : : BytesPerDepthPixel ( fb_depth_texture . format ) ;
// OpenGL needs 4 bpp alignment for D24
u32 gl_bpp = bytes_per_pixel = = 3 ? 4 : bytes_per_pixel ;
std : : unique_ptr < u8 [ ] > temp_gl_depth_buffer ( new u8 [ fb_depth_texture . width * fb_depth_texture . height * gl_bpp ] ) ;
state . texture_units [ 0 ] . texture_2d = fb_depth_texture . texture . handle ;
state . Apply ( ) ;
glActiveTexture ( GL_TEXTURE0 ) ;
glGetTexImage ( GL_TEXTURE_2D , 0 , fb_depth_texture . gl_format , fb_depth_texture . gl_type , temp_gl_depth_buffer . get ( ) ) ;
state . texture_units [ 0 ] . texture_2d = 0 ;
state . Apply ( ) ;
u8 * temp_gl_depth_data = bytes_per_pixel = = 3 ? ( temp_gl_depth_buffer . get ( ) + 1 ) : temp_gl_depth_buffer . get ( ) ;
if ( fb_depth_texture . format = = Pica : : Regs : : DepthFormat : : D24S8 ) {
for ( int y = 0 ; y < fb_depth_texture . height ; + + y ) {
for ( int x = 0 ; x < fb_depth_texture . width ; + + x ) {
const u32 coarse_y = y & ~ 7 ;
u32 dst_offset = VideoCore : : GetMortonOffset ( x , y , bytes_per_pixel ) + coarse_y * fb_depth_texture . width * bytes_per_pixel ;
u32 gl_pixel_index = ( x + ( fb_depth_texture . height - 1 - y ) * fb_depth_texture . width ) ;
u8 * pixel = depth_buffer + dst_offset ;
u32 depth_stencil = ( ( u32 * ) temp_gl_depth_data ) [ gl_pixel_index ] ;
* ( u32 * ) pixel = ( depth_stencil > > 8 ) | ( depth_stencil < < 24 ) ;
}
}
} else {
for ( int y = 0 ; y < fb_depth_texture . height ; + + y ) {
for ( int x = 0 ; x < fb_depth_texture . width ; + + x ) {
const u32 coarse_y = y & ~ 7 ;
u32 dst_offset = VideoCore : : GetMortonOffset ( x , y , bytes_per_pixel ) + coarse_y * fb_depth_texture . width * bytes_per_pixel ;
u32 gl_pixel_index = ( x + ( fb_depth_texture . height - 1 - y ) * fb_depth_texture . width ) * gl_bpp ;
u8 * pixel = depth_buffer + dst_offset ;
memcpy ( pixel , & temp_gl_depth_data [ gl_pixel_index ] , bytes_per_pixel ) ;
}
}
}
}
}
}
tfarley