diff --git a/rpcs3/Emu/RSX/GL/GLGSRender.cpp b/rpcs3/Emu/RSX/GL/GLGSRender.cpp index 75f0148006..0c7d1cbec1 100644 --- a/rpcs3/Emu/RSX/GL/GLGSRender.cpp +++ b/rpcs3/Emu/RSX/GL/GLGSRender.cpp @@ -400,7 +400,7 @@ void GLGSRender::end() m_gl_textures[i].set_target(target); - __glcheck m_gl_textures[i].init(i, textures[i]); + __glcheck m_gl_texture_cache.upload_texture(i, textures[i], m_gl_textures[i]); glProgramUniform1i(m_program->id(), location, i); } } @@ -771,6 +771,8 @@ void GLGSRender::on_init_thread() buf = new gl::buffer(); buf->create(); } + + m_gl_texture_cache.initialize_rtt_cache(); } void GLGSRender::on_exit() @@ -1208,7 +1210,11 @@ void GLGSRender::read_buffers() u32 width = rsx::method_registers[NV4097_SET_SURFACE_CLIP_HORIZONTAL] >> 16; u32 height = rsx::method_registers[NV4097_SET_SURFACE_CLIP_VERTICAL] >> 16; - + /** + * Read color buffers is useless if write buffers is enabled. I havent encountered a case where it is necessary + * since the output is usually fed back into the pipeline as a fragment shader input texture + * It is included here for completeness + */ for (int i = index; i < index + count; ++i) { u32 offset = rsx::method_registers[mr_color_offset[i]]; @@ -1218,20 +1224,28 @@ void GLGSRender::read_buffers() if (pitch <= 64) continue; - m_draw_tex_color[i].pixel_unpack_settings().row_length(pitch / (color_format.channel_size * color_format.channel_count)); - rsx::tiled_region color_buffer = get_tiled_address(offset, location & 0xf); + u32 texaddr = (u32)((u64)color_buffer.ptr - (u64)vm::base(0)); - if (!color_buffer.tile) + bool success = m_gl_texture_cache.explicit_writeback(m_draw_tex_color[i], texaddr, pitch); + + //Fall back to slower methods if the image could not be fetched. + if (!success) { - __glcheck m_draw_tex_color[i].copy_from(color_buffer.ptr, color_format.format, color_format.type); - } - else - { - std::unique_ptr buffer(new u8[pitch * height]); - color_buffer.read(buffer.get(), width, height, pitch); + if (!color_buffer.tile) + { + m_draw_tex_color[i].copy_from(color_buffer.ptr, color_format.format, color_format.type); + } + else + { + u32 range = pitch * height; + m_gl_texture_cache.remove_in_range(texaddr, range); - __glcheck m_draw_tex_color[i].copy_from(buffer.get(), color_format.format, color_format.type); + std::unique_ptr buffer(new u8[pitch * height]); + color_buffer.read(buffer.get(), width, height, pitch); + + __glcheck m_draw_tex_color[i].copy_from(buffer.get(), color_format.format, color_format.type); + } } } }; @@ -1271,8 +1285,15 @@ void GLGSRender::read_buffers() if (pitch <= 64) return; - auto depth_format = surface_depth_format_to_gl(m_surface.depth_format); + u32 depth_address = rsx::get_address(rsx::method_registers[NV4097_SET_SURFACE_ZETA_OFFSET], rsx::method_registers[NV4097_SET_CONTEXT_DMA_ZETA]); + bool in_cache = m_gl_texture_cache.explicit_writeback(m_draw_tex_depth_stencil, depth_address, pitch); + if (in_cache) + return; + + //Read failed. Fall back to slow s/w path... + + auto depth_format = surface_depth_format_to_gl(m_surface.depth_format); int pixel_size = get_pixel_size(m_surface.depth_format); gl::buffer pbo_depth; @@ -1312,9 +1333,6 @@ void GLGSRender::write_buffers() if (rpcs3::state.config.rsx.opengl.write_color_buffers) { - //gl::buffer pbo_color; - //__glcheck pbo_color.create(m_draw_tex_color[0].width() * m_draw_tex_color[0].height() * 4); - auto color_format = surface_color_format_to_gl(m_surface.color_format); auto write_color_buffers = [&](int index, int count) @@ -1324,23 +1342,6 @@ void GLGSRender::write_buffers() for (int i = index; i < index + count; ++i) { - //TODO: swizzle - //__glcheck m_draw_tex_color[i].copy_to(pbo_color, color_format.format, color_format.type); - - //pbo_color.map([&](GLubyte* pixels) - //{ - // u32 color_address = rsx::get_address(rsx::method_registers[mr_color_offset[i]], rsx::method_registers[mr_color_dma[i]]); - // //u32 depth_address = rsx::get_address(rsx::method_registers[NV4097_SET_SURFACE_ZETA_OFFSET], rsx::method_registers[NV4097_SET_CONTEXT_DMA_ZETA]); - - // const u32 *src = (const u32*)pixels; - // be_t* dst = vm::ps3::_ptr(color_address); - // for (int i = 0, end = m_draw_tex_color[i].width() * m_draw_tex_color[i].height(); i < end; ++i) - // { - // dst[i] = src[i]; - // } - - //}, gl::buffer::access::read); - u32 offset = rsx::method_registers[mr_color_offset[i]]; u32 location = rsx::method_registers[mr_color_dma[i]]; u32 pitch = rsx::method_registers[mr_color_pitch[i]]; @@ -1348,22 +1349,15 @@ void GLGSRender::write_buffers() if (pitch <= 64) continue; - m_draw_tex_color[i].pixel_pack_settings().row_length(pitch / (color_format.channel_size * color_format.channel_count)); - rsx::tiled_region color_buffer = get_tiled_address(offset, location & 0xf); + u32 texaddr = (u32)((u64)color_buffer.ptr - (u64)vm::base(0)); + u32 range = pitch * height; - if (!color_buffer.tile) - { - __glcheck m_draw_tex_color[i].copy_to(color_buffer.ptr, color_format.format, color_format.type); - } - else - { - std::unique_ptr buffer(new u8[pitch * height]); - - __glcheck m_draw_tex_color[i].copy_to(buffer.get(), color_format.format, color_format.type); - - color_buffer.write(buffer.get(), width, height, pitch); - } + /**Even tiles are loaded as whole textures during read_buffers from testing. + * Need further evaluation to determine correct behavior. Separate paths for both show no difference, + * but using the GPU to perform the caching is many times faster. + */ + __glcheck m_gl_texture_cache.save_render_target(texaddr, range, m_draw_tex_color[i]); } }; @@ -1403,38 +1397,12 @@ void GLGSRender::write_buffers() return; auto depth_format = surface_depth_format_to_gl(m_surface.depth_format); + u32 depth_address = rsx::get_address(rsx::method_registers[NV4097_SET_SURFACE_ZETA_OFFSET], rsx::method_registers[NV4097_SET_CONTEXT_DMA_ZETA]); + u32 range = m_draw_tex_depth_stencil.width() * m_draw_tex_depth_stencil.height() * 2; - gl::buffer pbo_depth; + if (m_surface.depth_format != rsx::surface_depth_format::z16) range *= 2; - int pixel_size = get_pixel_size(m_surface.depth_format); - - __glcheck pbo_depth.create(m_surface.width * m_surface.height * pixel_size); - __glcheck m_draw_tex_depth_stencil.copy_to(pbo_depth, depth_format.second, depth_format.first); - - __glcheck pbo_depth.map([&](GLubyte* pixels) - { - u32 depth_address = rsx::get_address(rsx::method_registers[NV4097_SET_SURFACE_ZETA_OFFSET], rsx::method_registers[NV4097_SET_CONTEXT_DMA_ZETA]); - - if (m_surface.depth_format == rsx::surface_depth_format::z16) - { - const u16 *src = (const u16*)pixels; - be_t* dst = vm::ps3::_ptr(depth_address); - for (int i = 0, end = m_draw_tex_depth_stencil.width() * m_draw_tex_depth_stencil.height(); i < end; ++i) - { - dst[i] = src[i]; - } - } - else - { - const u32 *src = (const u32*)pixels; - be_t* dst = vm::ps3::_ptr(depth_address); - for (int i = 0, end = m_draw_tex_depth_stencil.width() * m_draw_tex_depth_stencil.height(); i < end; ++i) - { - dst[i] = src[i]; - } - } - - }, gl::buffer::access::read); + m_gl_texture_cache.save_render_target(depth_address, range, m_draw_tex_depth_stencil); } } @@ -1449,7 +1417,10 @@ void GLGSRender::flip(int buffer) bool skip_read = false; - if (draw_fbo && !rpcs3::state.config.rsx.opengl.write_color_buffers) + /** + * Calling read_buffers will overwrite cached content + */ + if (draw_fbo) { skip_read = true; /* @@ -1558,3 +1529,9 @@ u64 GLGSRender::timestamp() const glGetInteger64v(GL_TIMESTAMP, &result); return result; } + +bool GLGSRender::on_access_violation(u32 address, bool is_writing) +{ + if (is_writing) return m_gl_texture_cache.mark_as_dirty(address); + return false; +} \ No newline at end of file diff --git a/rpcs3/Emu/RSX/GL/GLGSRender.h b/rpcs3/Emu/RSX/GL/GLGSRender.h index bcc77b5300..4a7c150657 100644 --- a/rpcs3/Emu/RSX/GL/GLGSRender.h +++ b/rpcs3/Emu/RSX/GL/GLGSRender.h @@ -2,6 +2,7 @@ #include "Emu/RSX/GSRender.h" #include "gl_helpers.h" #include "rsx_gl_texture.h" +#include "gl_texture_cache.h" #define RSX_DEBUG 1 @@ -29,6 +30,8 @@ private: } m_gl_attrib_buffers[rsx::limits::vertex_count]; + gl::gl_texture_cache m_gl_texture_cache; + public: gl::fbo draw_fbo; @@ -73,4 +76,6 @@ protected: bool do_method(u32 id, u32 arg) override; void flip(int buffer) override; u64 timestamp() const override; + + bool on_access_violation(u32 address, bool is_writing) override; }; diff --git a/rpcs3/Emu/RSX/GL/GLProcTable.h b/rpcs3/Emu/RSX/GL/GLProcTable.h index 89f971baee..54c06d4a4b 100644 --- a/rpcs3/Emu/RSX/GL/GLProcTable.h +++ b/rpcs3/Emu/RSX/GL/GLProcTable.h @@ -170,6 +170,9 @@ OPENGL_PROC(PFNGLBINDBUFFERBASEPROC, BindBufferBase); OPENGL_PROC(PFNGLTEXBUFFERPROC, TexBuffer); OPENGL_PROC(PFNGLTEXTUREBUFFERRANGEEXTPROC, TextureBufferRangeEXT); +//ARB_Copy_Image +OPENGL_PROC(PFNGLCOPYIMAGESUBDATAPROC, CopyImageSubData); + //KHR_debug OPENGL_PROC(PFNGLDEBUGMESSAGECONTROLARBPROC, DebugMessageControlARB); OPENGL_PROC(PFNGLDEBUGMESSAGEINSERTARBPROC, DebugMessageInsertARB); diff --git a/rpcs3/Emu/RSX/GL/gl_texture_cache.h b/rpcs3/Emu/RSX/GL/gl_texture_cache.h new file mode 100644 index 0000000000..fa1ea9f8d2 --- /dev/null +++ b/rpcs3/Emu/RSX/GL/gl_texture_cache.h @@ -0,0 +1,659 @@ +#pragma once + +#include "stdafx.h" + +#include +#include +#include +#include +#include +#include + +#include "GLGSRender.h" +#include "../Common/TextureUtils.h" +#include + +namespace gl +{ + class gl_texture_cache + { + public: + + struct gl_cached_texture + { + u32 gl_id; + u32 w; + u32 h; + u64 data_addr; + u32 block_sz; + u32 frame_ctr; + u32 protected_block_start; + u32 protected_block_sz; + u16 mipmap; + bool deleted; + bool locked; + }; + + struct invalid_cache_area + { + u32 block_base; + u32 block_sz; + }; + + struct cached_rtt + { + u32 copy_glid; + u32 data_addr; + u32 block_sz; + + bool is_dirty; + bool is_depth; + bool valid; + + u32 current_width; + u32 current_height; + + bool locked; + cached_rtt() : valid(false) {} + }; + + private: + std::vector texture_cache; + std::vector rtt_cache; + u32 frame_ctr; + + bool lock_memory_region(u32 start, u32 size) + { + static const u32 memory_page_size = 4096; + start = start & ~(memory_page_size - 1); + size = (u32)align(size, memory_page_size); + + return vm::page_protect(start, size, 0, 0, vm::page_writable); + } + + bool unlock_memory_region(u32 start, u32 size) + { + static const u32 memory_page_size = 4096; + start = start & ~(memory_page_size - 1); + size = (u32)align(size, memory_page_size); + + return vm::page_protect(start, size, 0, vm::page_writable, 0); + } + + void lock_gl_object(gl_cached_texture &obj) + { + static const u32 memory_page_size = 4096; + obj.protected_block_start = obj.data_addr & ~(memory_page_size - 1); + obj.protected_block_sz = (u32)align(obj.block_sz, memory_page_size); + + if (!lock_memory_region(obj.protected_block_start, obj.protected_block_sz)) + LOG_ERROR(RSX, "lock_gl_object failed!"); + else + obj.locked = true; + } + + void unlock_gl_object(gl_cached_texture &obj) + { + if (!unlock_memory_region(obj.protected_block_start, obj.protected_block_sz)) + LOG_ERROR(RSX, "unlock_gl_object failed! Will probably crash soon..."); + else + obj.locked = false; + } + + gl_cached_texture *find_obj_for_params(u64 texaddr, u32 w, u32 h, u16 mipmap) + { + for (gl_cached_texture &tex: texture_cache) + { + if (tex.gl_id && tex.data_addr == texaddr) + { + if (w && h && mipmap && (tex.h != h || tex.w != w || tex.mipmap != mipmap)) + { + LOG_ERROR(RSX, "Texture params are invalid for block starting 0x%X!", tex.data_addr); + LOG_ERROR(RSX, "Params passed w=%d, h=%d, mip=%d, found w=%d, h=%d, mip=%d", w, h, mipmap, tex.w, tex.h, tex.mipmap); + + continue; + } + + tex.frame_ctr = frame_ctr; + return &tex; + } + } + + return nullptr; + } + + gl_cached_texture& create_obj_for_params(u32 gl_id, u64 texaddr, u32 w, u32 h, u16 mipmap) + { + gl_cached_texture obj = { 0 }; + + obj.gl_id = gl_id; + obj.data_addr = texaddr; + obj.w = w; + obj.h = h; + obj.mipmap = mipmap; + obj.deleted = false; + obj.locked = false; + + for (gl_cached_texture &tex : texture_cache) + { + if (tex.gl_id == 0 || (tex.deleted && (frame_ctr - tex.frame_ctr) > 32768)) + { + if (tex.gl_id) + { + LOG_NOTICE(RSX, "Reclaiming GL texture %d, cache_size=%d, master_ctr=%d, ctr=%d", tex.gl_id, texture_cache.size(), frame_ctr, tex.frame_ctr); + __glcheck glDeleteTextures(1, &tex.gl_id); + unlock_gl_object(tex); + tex.gl_id = 0; + } + + tex = obj; + return tex; + } + } + + texture_cache.push_back(obj); + return texture_cache[texture_cache.size()-1]; + } + + void remove_obj(gl_cached_texture &tex) + { + if (tex.locked) + unlock_gl_object(tex); + + tex.deleted = true; + } + + void remove_obj_for_glid(u32 gl_id) + { + for (gl_cached_texture &tex : texture_cache) + { + if (tex.gl_id == gl_id) + remove_obj(tex); + } + } + + void clear_obj_cache() + { + for (gl_cached_texture &tex : texture_cache) + { + if (tex.locked) + unlock_gl_object(tex); + + if (tex.gl_id) + { + LOG_NOTICE(RSX, "Deleting texture %d", tex.gl_id); + glDeleteTextures(1, &tex.gl_id); + } + + tex.deleted = true; + tex.gl_id = 0; + } + + texture_cache.resize(0); + destroy_rtt_cache(); + } + + bool region_overlaps(u32 base1, u32 limit1, u32 base2, u32 limit2) + { + //Check for memory area overlap. unlock page(s) if needed and add this index to array. + //Axis separation test + const u32 &block_start = base1; + const u32 block_end = limit1; + + if (limit2 < block_start) return false; + if (base2 > block_end) return false; + + u32 min_separation = (limit2 - base2) + (limit1 - base1); + u32 range_limit = (block_end > limit2) ? block_end : limit2; + u32 range_base = (block_start < base2) ? block_start : base2; + + u32 actual_separation = (range_limit - range_base); + + if (actual_separation < min_separation) + return true; + + return false; + } + + cached_rtt* find_cached_rtt(u32 base, u32 size) + { + for (cached_rtt &rtt : rtt_cache) + { + if (region_overlaps(base, base+size, rtt.data_addr, rtt.data_addr+rtt.block_sz)) + { + return &rtt; + } + } + + return nullptr; + } + + void invalidate_rtts_in_range(u32 base, u32 size) + { + for (cached_rtt &rtt : rtt_cache) + { + if (!rtt.data_addr || rtt.is_dirty) continue; + + u32 rtt_aligned_base = ((u32)(rtt.data_addr)) & ~(4096 - 1); + u32 rtt_block_sz = align(rtt.block_sz, 4096); + + if (region_overlaps(rtt_aligned_base, (rtt_aligned_base + rtt_block_sz), base, base+size)) + { + LOG_NOTICE(RSX, "Dirty RTT FOUND addr=0x%X", base); + rtt.is_dirty = true; + if (rtt.locked) + { + rtt.locked = false; + unlock_memory_region((u32)rtt.data_addr, rtt.block_sz); + } + } + } + } + + void prep_rtt(cached_rtt &rtt, u32 width, u32 height, u32 gl_pixel_format_internal) + { + int binding = 0; + bool is_depth = false; + + if (gl_pixel_format_internal == GL_DEPTH24_STENCIL8 || + gl_pixel_format_internal == GL_DEPTH_COMPONENT24 || + gl_pixel_format_internal == GL_DEPTH_COMPONENT16 || + gl_pixel_format_internal == GL_DEPTH_COMPONENT32) + { + is_depth = true; + } + + glGetIntegerv(GL_TEXTURE_2D_BINDING_EXT, &binding); + glBindTexture(GL_TEXTURE_2D, rtt.copy_glid); + + rtt.current_width = width; + rtt.current_height = height; + + if (!is_depth) + { + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + + __glcheck glTexImage2D(GL_TEXTURE_2D, 0, gl_pixel_format_internal, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr); + } + else + { + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + + u32 ex_format = GL_UNSIGNED_SHORT; + u32 in_format = GL_DEPTH_COMPONENT16; + + switch (gl_pixel_format_internal) + { + case GL_DEPTH24_STENCIL8: + { + ex_format = GL_UNSIGNED_INT_24_8; + in_format = GL_DEPTH_STENCIL; + break; + } + case GL_DEPTH_COMPONENT16: + break; + default: + throw EXCEPTION("Unsupported depth format!"); + } + + __glcheck glTexImage2D(GL_TEXTURE_2D, 0, gl_pixel_format_internal, width, height, 0, in_format, ex_format, nullptr); + } + + glBindTexture(GL_TEXTURE_2D, binding); + rtt.is_depth = is_depth; + } + + void save_rtt(u32 base, u32 size, u32 width, u32 height, u32 gl_pixel_format_internal, gl::texture &source) + { + cached_rtt *region = find_cached_rtt(base, size); + + if (!region) + { + for (cached_rtt &rtt : rtt_cache) + { + if (rtt.valid && rtt.data_addr == 0) + { + prep_rtt(rtt, width, height, gl_pixel_format_internal); + + rtt.block_sz = size; + rtt.data_addr = base; + rtt.is_dirty = true; + + LOG_NOTICE(RSX, "New RTT created for block 0x%X + 0x%X", (u32)rtt.data_addr, rtt.block_sz); + + lock_memory_region((u32)rtt.data_addr, rtt.block_sz); + rtt.locked = true; + + region = &rtt; + break; + } + } + + if (!region) throw EXCEPTION("No region created!!"); + } + + if (width != region->current_width || + height != region->current_height) + { + prep_rtt(*region, width, height, gl_pixel_format_internal); + + if (region->locked && region->block_sz != size) + { + LOG_NOTICE(RSX, "Unlocking RTT since size has changed!"); + unlock_memory_region((u32)region->data_addr, region->block_sz); + + LOG_NOTICE(RSX, "Locking down RTT after size change!"); + region->block_sz = size; + lock_memory_region((u32)region->data_addr, region->block_sz); + region->locked = true; + } + } + + __glcheck glCopyImageSubData(source.id(), GL_TEXTURE_2D, 0, 0, 0, 0, + region->copy_glid, GL_TEXTURE_2D, 0, 0, 0, 0, + width, height, 1); + + region->is_dirty = false; + + if (!region->locked) + { + LOG_WARNING(RSX, "Locking down RTT, was unlocked!"); + lock_memory_region((u32)region->data_addr, region->block_sz); + region->locked = true; + } + } + + void write_rtt(u32 base, u32 size, u32 texaddr) + { + //Actually download the data, since it seems that cell is writing to it manually + throw; + } + + void destroy_rtt_cache() + { + for (cached_rtt &rtt : rtt_cache) + { + rtt.valid = false; + rtt.is_dirty = false; + rtt.block_sz = 0; + rtt.data_addr = 0; + + glDeleteTextures(1, &rtt.copy_glid); + rtt.copy_glid = 0; + } + + rtt_cache.resize(0); + } + + public: + + gl_texture_cache() + : frame_ctr(0) + { + } + + ~gl_texture_cache() + { + clear_obj_cache(); + } + + void update_frame_ctr() + { + frame_ctr++; + } + + void initialize_rtt_cache() + { + if (rtt_cache.size()) throw EXCEPTION("Initialize RTT cache while cache already exists! Leaking objects??"); + + for (int i = 0; i < 64; ++i) + { + cached_rtt rtt; + + glGenTextures(1, &rtt.copy_glid); + rtt.is_dirty = true; + rtt.valid = true; + rtt.block_sz = 0; + rtt.data_addr = 0; + rtt.locked = false; + + rtt_cache.push_back(rtt); + } + } + + void upload_texture(int index, rsx::texture &tex, rsx::gl::texture &gl_texture) + { + const u32 texaddr = rsx::get_address(tex.offset(), tex.location()); + const u32 range = (u32)get_texture_size(tex); + + cached_rtt *rtt = find_cached_rtt(texaddr, range); + + if (rtt && !rtt->is_dirty) + { + if (!rtt->is_depth) + { + u32 real_id = gl_texture.id(); + + glActiveTexture(GL_TEXTURE0 + index); + gl_texture.set_id(rtt->copy_glid); + gl_texture.bind(); + + gl_texture.set_id(real_id); + } + else + { + LOG_NOTICE(RSX, "Depth RTT found from 0x%X, Trying to upload width dims: %d x %d, Saved as %d x %d", rtt->data_addr, tex.width(), tex.height(), rtt->current_width, rtt->current_height); + //The texture should have already been loaded through the writeback interface call + //Bind it directly + u32 real_id = gl_texture.id(); + + glActiveTexture(GL_TEXTURE0 + index); + gl_texture.set_id(rtt->copy_glid); + gl_texture.bind(); + + gl_texture.set_id(real_id); + } + return; + } + else if (rtt) + LOG_NOTICE(RSX, "RTT texture for address 0x%X is dirty!", texaddr); + + gl_cached_texture *obj = nullptr; + + if (!rtt) + obj = find_obj_for_params(texaddr, tex.width(), tex.height(), tex.mipmap()); + + if (obj && !obj->deleted) + { + u32 real_id = gl_texture.id(); + + glActiveTexture(GL_TEXTURE0 + index); + gl_texture.set_id(obj->gl_id); + gl_texture.bind(); + + gl_texture.set_id(real_id); + } + else + { + if (!obj) gl_texture.set_id(0); + else + { + //Reuse this GLid + gl_texture.set_id(obj->gl_id); + + //Empty this slot for another one. A new holder will be created below anyway... + if (obj->locked) unlock_gl_object(*obj); + obj->gl_id = 0; + } + + __glcheck gl_texture.init(index, tex); + gl_cached_texture &_obj = create_obj_for_params(gl_texture.id(), texaddr, tex.width(), tex.height(), tex.mipmap()); + + _obj.block_sz = (u32)get_texture_size(tex); + lock_gl_object(_obj); + } + } + + bool mark_as_dirty(u32 address) + { + bool response = false; + + for (gl_cached_texture &tex: texture_cache) + { + if (!tex.locked) continue; + + if (tex.protected_block_start <= address && + tex.protected_block_sz >(address - tex.protected_block_start)) + { + LOG_NOTICE(RSX, "Texture object is dirty! %d", tex.gl_id); + unlock_gl_object(tex); + + invalidate_rtts_in_range((u32)tex.data_addr, tex.block_sz); + + tex.deleted = true; + response = true; + } + } + + if (response) return true; + + for (cached_rtt &rtt: rtt_cache) + { + if (!rtt.data_addr || rtt.is_dirty) continue; + + u32 rtt_aligned_base = ((u32)(rtt.data_addr)) & ~(4096 - 1); + u32 rtt_block_sz = align(rtt.block_sz, 4096); + + if (rtt.locked && (u64)address >= rtt_aligned_base) + { + u32 offset = address - rtt_aligned_base; + if (offset >= rtt_block_sz) continue; + + LOG_NOTICE(RSX, "Dirty non-texture RTT FOUND! addr=0x%X", rtt.data_addr); + rtt.is_dirty = true; + + unlock_memory_region(rtt_aligned_base, rtt_block_sz); + rtt.locked = false; + + response = true; + } + } + + return response; + } + + void save_render_target(u32 texaddr, u32 range, gl::texture &gl_texture) + { + save_rtt(texaddr, range, gl_texture.width(), gl_texture.height(), (GLenum)gl_texture.get_internal_format(), gl_texture); + } + + std::vector find_and_invalidate_in_range(u32 base, u32 limit) + { + /** + * Sometimes buffers can share physical pages. + * Return objects if we really encroach on texture + */ + + std::vector result; + + for (gl_cached_texture &obj : texture_cache) + { + //Check for memory area overlap. unlock page(s) if needed and add this index to array. + //Axis separation test + const u32 &block_start = obj.protected_block_start; + const u32 block_end = block_start + obj.protected_block_sz; + + if (limit < block_start) continue; + if (base > block_end) continue; + + u32 min_separation = (limit - base) + obj.protected_block_sz; + u32 range_limit = (block_end > limit) ? block_end : limit; + u32 range_base = (block_start < base) ? block_start : base; + + u32 actual_separation = (range_limit - range_base); + + if (actual_separation < min_separation) + { + const u32 texture_start = (u32)obj.data_addr; + const u32 texture_end = texture_start + obj.block_sz; + + min_separation = (limit - base) + obj.block_sz; + range_limit = (texture_end > limit) ? texture_end : limit; + range_base = (texture_start < base) ? texture_start : base; + + actual_separation = (range_limit - range_base); + + if (actual_separation < min_separation) + { + //Texture area is invalidated! + unlock_gl_object(obj); + obj.deleted = true; + + continue; + } + + //Overlap in this case will be at most 1 page... + invalid_cache_area invalid = { 0 }; + if (base < obj.data_addr) + invalid.block_base = obj.protected_block_start; + else + invalid.block_base = obj.protected_block_start + obj.protected_block_sz - 4096; + + invalid.block_sz = 4096; + unlock_memory_region(invalid.block_base, invalid.block_sz); + result.push_back(invalid); + } + } + + return result; + } + + void lock_invalidated_ranges(std::vector invalid) + { + for (invalid_cache_area area : invalid) + { + lock_memory_region(area.block_base, area.block_sz); + } + } + + void remove_in_range(u32 texaddr, u32 range) + { + //Seems that the rsx only 'reads' full texture objects.. + //This simplifies this function to simply check for matches + for (gl_cached_texture &cached : texture_cache) + { + if (cached.data_addr == texaddr && + cached.block_sz == range) + remove_obj(cached); + } + } + + bool explicit_writeback(gl::texture &tex, const u32 address, const u32 pitch) + { + const u32 range = tex.height() * pitch; + cached_rtt *rtt = find_cached_rtt(address, range); + + if (rtt && !rtt->is_dirty) + { + u32 min_w = rtt->current_width; + u32 min_h = rtt->current_height; + + if ((u32)tex.width() < min_w) min_w = (u32)tex.width(); + if ((u32)tex.height() < min_h) min_h = (u32)tex.height(); + + //TODO: Image reinterpretation e.g read back rgba data as depth texture and vice-versa + + __glcheck glCopyImageSubData(rtt->copy_glid, GL_TEXTURE_2D, 0, 0, 0, 0, + tex.id(), GL_TEXTURE_2D, 0, 0, 0, 0, + min_w, min_h, 1); + + return true; + } + + //No valid object found in cache + return false; + } + }; +} \ No newline at end of file diff --git a/rpcs3/Emu/RSX/GL/rsx_gl_texture.h b/rpcs3/Emu/RSX/GL/rsx_gl_texture.h index b9167d6e71..c16da5854b 100644 --- a/rpcs3/Emu/RSX/GL/rsx_gl_texture.h +++ b/rpcs3/Emu/RSX/GL/rsx_gl_texture.h @@ -56,6 +56,7 @@ namespace rsx void remove(); void set_target(u32 target) { m_target = target; } + void set_id(u32 id) { m_id = id; } u32 id() const; }; } diff --git a/rpcs3/GLGSRender.vcxproj b/rpcs3/GLGSRender.vcxproj index 7e0ba6601a..604cffd106 100644 --- a/rpcs3/GLGSRender.vcxproj +++ b/rpcs3/GLGSRender.vcxproj @@ -79,6 +79,7 @@ +