diff --git a/CMakeLists.txt b/CMakeLists.txt index 048dd44a..77f1c1f4 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -5,7 +5,7 @@ if (POLICY CMP0076) cmake_policy(SET CMP0076 NEW) endif() -set(CMAKE_CXX_STANDARD 11) +set(CMAKE_CXX_STANDARD 17) set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}) project(melonDS) diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 8a561bb9..c2a2f573 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -28,6 +28,7 @@ add_library(core STATIC xxhash/xxhash.c stb/stb.cpp + xbrz/xbrz.cpp ) if (WIN32) diff --git a/src/GPU.cpp b/src/GPU.cpp index 2ecd8802..d130d225 100644 --- a/src/GPU.cpp +++ b/src/GPU.cpp @@ -602,7 +602,7 @@ void MapVRAM_FG(u32 bank, u8 cnt) VRAMPtr_ABG[base] = GetUniqueBankPtr(VRAMMap_ABG[base], base << 14); VRAMPtr_ABG[base + 2] = GetUniqueBankPtr(VRAMMap_ABG[base + 2], (base + 2) << 14); } - LCDCDirty[bank][0] = 0xFFFFFFFF; + LCDCDirty[bank][0] = 0xFFFF; break; case 2: // AOBJ @@ -613,7 +613,7 @@ void MapVRAM_FG(u32 bank, u8 cnt) VRAMPtr_AOBJ[base] = GetUniqueBankPtr(VRAMMap_AOBJ[base], base << 14); VRAMPtr_AOBJ[base + 2] = GetUniqueBankPtr(VRAMMap_AOBJ[base + 2], (base + 2) << 14); } - LCDCDirty[bank][0] = 0xFFFFFFFF; + LCDCDirty[bank][0] = 0xFFFF; break; case 3: // texture palette diff --git a/src/GPU.h b/src/GPU.h index 0a13e0b0..1e369eab 100644 --- a/src/GPU.h +++ b/src/GPU.h @@ -224,7 +224,7 @@ void WriteVRAM_LCDC(u32 addr, T val) if (VRAMMap_LCDC & (1<> 16] |= 1 << ((addr >> 10) & 0x3F); + LCDCDirty[bank][addr >> 16] |= 1ULL << ((addr >> 10) & 0x3F); } } diff --git a/src/GPU3D.h b/src/GPU3D.h index 94c45a3f..5aba8704 100644 --- a/src/GPU3D.h +++ b/src/GPU3D.h @@ -132,10 +132,6 @@ namespace TexCache typedef u64 ExternalTexHandle; -typedef u32* (*AllocTextureFunc)(ExternalTexHandle* handle, u32 width, u32 height); -typedef void (*FreeTextureFunc)(ExternalTexHandle handle, u32 width, u32 height); -typedef void (*FinaliseTextureFunc)(ExternalTexHandle handle, u32 width, u32 height); - enum { outputFmt_RGB6A5, @@ -183,6 +179,10 @@ bool Init(); void DeInit(); void Reset(); +u32* AllocateTexture(TexCache::ExternalTexHandle* handle, u32 width, u32 height); +void FreeTexture(TexCache::ExternalTexHandle handle, u32 width, u32 height); +void FinaliseTexture(TexCache::ExternalTexHandle handle, u32 width, u32 height); + void UpdateDisplaySettings(); void RenderFrame(); diff --git a/src/GPU3D_OpenGL.cpp b/src/GPU3D_OpenGL.cpp index d616e1c7..6669bdbf 100644 --- a/src/GPU3D_OpenGL.cpp +++ b/src/GPU3D_OpenGL.cpp @@ -24,6 +24,10 @@ #include "OpenGLSupport.h" #include "GPU3D_OpenGL_shaders.h" +#include "xbrz/xbrz.h" + +#include + namespace GPU3D { namespace GLRenderer @@ -80,6 +84,8 @@ typedef struct u32 RenderKey; + GLuint TextureID; + } RendererPolygon; RendererPolygon PolygonList[2048]; @@ -107,9 +113,6 @@ u32 NumVertices; GLuint VertexArrayID; u16 IndexBuffer[2048 * 40]; -GLuint TexMemID; -GLuint TexPalMemID; - int ScaleFactor; bool Antialias; int ScreenW, ScreenH; @@ -119,7 +122,108 @@ int FrontBuffer; GLuint FramebufferID[4], PixelbufferID; u32 Framebuffer[256*192]; +struct TextureAllocator +{ + u32 Length = 0; + GLuint TextureID; + u32 FreeEntriesLeft = 0; + u64 FreeEntries[8]; + u32* Pixels; +}; +TextureAllocator TextureMem[8][8]; +inline TextureAllocator& GetTextureAllocator(u32 width, u32 height) +{ + return TextureMem[__builtin_ctz(width) - 3][__builtin_ctz(height) - 3]; +} + +u32 ConversionBuffer[1024*1024]; + +u32* AllocateTexture(TexCache::ExternalTexHandle* handle, u32 width, u32 height) +{ + TextureAllocator& allocator = GetTextureAllocator(width, height); + //printf("allocating texture %d %d\n", width, height); + + u32 scaledWidth = width * 2; + u32 scaledHeight = height * 2; + + if (allocator.FreeEntriesLeft == 0) + { + u32 newLength = (allocator.Length * 3) / 2 + 8; + + if (newLength >= 64*8) + abort(); + + if (allocator.Length == 0) + { + printf("created new texture\n"); + glGenTextures(1, &allocator.TextureID); + glBindTexture(GL_TEXTURE_2D_ARRAY, allocator.TextureID); + + glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_S, GL_MIRRORED_REPEAT); + glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_T, GL_MIRRORED_REPEAT); + } + else + glBindTexture(GL_TEXTURE_2D_ARRAY, allocator.TextureID); + + glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA8, scaledWidth, scaledHeight, newLength, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL); + + u32* newPixels = new u32[scaledWidth * scaledHeight * newLength]; + if (allocator.Length) + { + // copy old data + glTexSubImage3D(GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, scaledWidth, scaledHeight, allocator.Length, GL_RGBA, GL_UNSIGNED_BYTE, allocator.Pixels); + + memcpy(newPixels, allocator.Pixels, allocator.Length * scaledWidth * scaledHeight * 4); + delete[] allocator.Pixels; + } + allocator.Pixels = newPixels; + + allocator.FreeEntriesLeft = newLength - allocator.Length; + allocator.Length = newLength; + } + + for (int i = 0; i < (allocator.Length + 0x3F & ~0x3F) >> 6; i++) + { + if (allocator.FreeEntries[i] != 0xFFFFFFFFFFFFFFFF) + { + allocator.FreeEntriesLeft--; + u64 freeIdx = __builtin_ctzll(~allocator.FreeEntries[i]); + allocator.FreeEntries[i] |= 1ULL << freeIdx; + *handle = i * 64 + freeIdx; + + //return &allocator.Pixels[*handle * scaledWidth * scaledHeight]; + return ConversionBuffer; + } + } + + // should never happen, hopefully... + abort(); +} + +void FreeTexture(TexCache::ExternalTexHandle handle, u32 width, u32 height) +{ + TextureAllocator& allocator = GetTextureAllocator(width, height); + + allocator.FreeEntriesLeft++; + allocator.FreeEntries[handle >> 6] &= ~(1 << (handle & 0x3F)); +} + +void FinaliseTexture(TexCache::ExternalTexHandle handle, u32 width, u32 height) +{ + TextureAllocator& allocator = GetTextureAllocator(width, height); + + u32 scaledWidth = width * 2; + u32 scaledHeight = height * 2; + + xbrz::scale(2, ConversionBuffer, &allocator.Pixels[handle * scaledWidth * scaledHeight], width, height, xbrz::ColorFormat::ARGB); + + // could still be improved + glBindTexture(GL_TEXTURE_2D_ARRAY, allocator.TextureID); + glTexSubImage3D(GL_TEXTURE_2D_ARRAY, 0, 0, 0, handle, scaledWidth, scaledHeight, 1, GL_RGBA, GL_UNSIGNED_BYTE, &allocator.Pixels[handle * scaledWidth * scaledHeight]); +} bool BuildRenderShader(u32 flags, const char* vs, const char* fs) { @@ -166,10 +270,8 @@ bool BuildRenderShader(u32 flags, const char* vs, const char* fs) glUseProgram(prog); - uni_id = glGetUniformLocation(prog, "TexMem"); + uni_id = glGetUniformLocation(prog, "Textures"); glUniform1i(uni_id, 0); - uni_id = glGetUniformLocation(prog, "TexPalMem"); - glUniform1i(uni_id, 1); return true; } @@ -201,7 +303,6 @@ bool Init() glDepthRange(0, 1); glClearDepth(1.0); - if (!OpenGL_BuildShaderProgram(kClearVS, kClearFS, ClearShaderPlain, "ClearShader")) return false; @@ -353,32 +454,11 @@ bool Init() glGenBuffers(1, &PixelbufferID); - glActiveTexture(GL_TEXTURE0); - glGenTextures(1, &TexMemID); - glBindTexture(GL_TEXTURE_2D, TexMemID); - 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_MIN_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); - glTexImage2D(GL_TEXTURE_2D, 0, GL_R8UI, 1024, 512, 0, GL_RED_INTEGER, GL_UNSIGNED_BYTE, NULL); - - glActiveTexture(GL_TEXTURE1); - glGenTextures(1, &TexPalMemID); - glBindTexture(GL_TEXTURE_2D, TexPalMemID); - 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_MIN_FILTER, GL_NEAREST); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); - glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB5_A1, 1024, 48, 0, GL_RGBA, GL_UNSIGNED_SHORT_1_5_5_5_REV, NULL); - return true; } void DeInit() { - glDeleteTextures(1, &TexMemID); - glDeleteTextures(1, &TexPalMemID); - glDeleteFramebuffers(4, &FramebufferID[0]); glDeleteTextures(8, &FramebufferTex[0]); @@ -539,6 +619,17 @@ void BuildPolygons(RendererPolygon* polygons, int npolys) RendererPolygon* rp = &polygons[i]; Polygon* poly = rp->PolyData; + TexCache::ExternalTexHandle texture = UINT64_MAX; + if (poly->TexParam & 0x1C000000 && !poly->IsShadowMask) + { + texture = TexCache::GetTexture(poly->TexParam, poly->TexPalette); + + u32 width = (poly->TexParam >> 20) & 0x7; + u32 height = (poly->TexParam >> 23) & 0x7; + + rp->TextureID = TextureMem[width][height].TextureID; + } + rp->Indices = iptr; rp->NumIndices = 0; @@ -599,7 +690,7 @@ void BuildPolygons(RendererPolygon* polygons, int npolys) *vptr++ = vtxattr | (zshift << 16); *vptr++ = poly->TexParam; - *vptr++ = poly->TexPalette; + *vptr++ = texture; *iptr++ = vidx; rp->NumIndices++; @@ -648,7 +739,7 @@ void BuildPolygons(RendererPolygon* polygons, int npolys) *vptr++ = vtxattr | (zshift << 16); *vptr++ = poly->TexParam; - *vptr++ = poly->TexPalette; + *vptr++ = texture; if (j >= 2) { @@ -686,6 +777,8 @@ void RenderSinglePolygon(int i) { RendererPolygon* rp = &PolygonList[i]; + glBindTexture(GL_TEXTURE_2D_ARRAY, rp->TextureID); + glDrawElements(rp->PrimType, rp->NumIndices, GL_UNSIGNED_SHORT, rp->Indices); } @@ -697,17 +790,38 @@ int RenderPolygonBatch(int i) int numpolys = 0; u32 numindices = 0; - for (int iend = i; iend < NumFinalPolys; iend++) + GLuint texture; + bool textureChange; + + do { - RendererPolygon* cur_rp = &PolygonList[iend]; - if (cur_rp->PrimType != primtype) break; - if (cur_rp->RenderKey != key) break; + textureChange = false; - numpolys++; - numindices += cur_rp->NumIndices; - } + rp = &PolygonList[i]; + texture = rp->TextureID; + + for (; i < NumFinalPolys; i++) + { + RendererPolygon* cur_rp = &PolygonList[i]; + if (cur_rp->PrimType != primtype) break; + if (cur_rp->RenderKey != key) break; + if (cur_rp->TextureID != texture) + { + textureChange = true; + break; + } + + numpolys++; + numindices += cur_rp->NumIndices; + } + + glBindTexture(GL_TEXTURE_2D_ARRAY, texture); + + glDrawElements(primtype, numindices, GL_UNSIGNED_SHORT, rp->Indices); + + numindices = 0; + } while (textureChange); - glDrawElements(primtype, numindices, GL_UNSIGNED_SHORT, rp->Indices); return numpolys; } @@ -1111,38 +1225,7 @@ void RenderFrame() if (unibuf) memcpy(unibuf, &ShaderConfig, sizeof(ShaderConfig)); glUnmapBuffer(GL_UNIFORM_BUFFER); - // SUCKY!!!!!!!!!!!!!!!!!! - // TODO: detect when VRAM blocks are modified! glActiveTexture(GL_TEXTURE0); - glBindTexture(GL_TEXTURE_2D, TexMemID); - for (int i = 0; i < 4; i++) - { - u32 mask = GPU::VRAMMap_Texture[i]; - u8* vram; - if (!mask) continue; - else if (mask & (1<<0)) vram = GPU::VRAM_A; - else if (mask & (1<<1)) vram = GPU::VRAM_B; - else if (mask & (1<<2)) vram = GPU::VRAM_C; - else if (mask & (1<<3)) vram = GPU::VRAM_D; - - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, i*128, 1024, 128, GL_RED_INTEGER, GL_UNSIGNED_BYTE, vram); - } - - glActiveTexture(GL_TEXTURE1); - glBindTexture(GL_TEXTURE_2D, TexPalMemID); - for (int i = 0; i < 6; i++) - { - // 6 x 16K chunks - u32 mask = GPU::VRAMMap_TexPal[i]; - u8* vram; - if (!mask) continue; - else if (mask & (1<<4)) vram = &GPU::VRAM_E[(i&3)*0x4000]; - else if (mask & (1<<5)) vram = GPU::VRAM_F; - else if (mask & (1<<6)) vram = GPU::VRAM_G; - - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, i*8, 1024, 8, GL_RGBA, GL_UNSIGNED_SHORT_1_5_5_5_REV, vram); - } - glDisable(GL_SCISSOR_TEST); glEnable(GL_DEPTH_TEST); glEnable(GL_STENCIL_TEST); @@ -1188,6 +1271,8 @@ void RenderFrame() glDrawArrays(GL_TRIANGLES, 0, 2*3); } + TexCache::UpdateTextures(); + if (RenderNumPolygons) { // render shit here diff --git a/src/GPU3D_OpenGL_shaders.h b/src/GPU3D_OpenGL_shaders.h index 2545ee0d..1b775c9c 100644 --- a/src/GPU3D_OpenGL_shaders.h +++ b/src/GPU3D_OpenGL_shaders.h @@ -241,9 +241,6 @@ flat out ivec3 fPolygonAttr; const char* kRenderFSCommon = R"( -uniform usampler2D TexMem; -uniform sampler2D TexPalMem; - layout(std140) uniform uConfig { vec2 uScreenSize; @@ -260,334 +257,23 @@ smooth in vec4 fColor; smooth in vec2 fTexcoord; flat in ivec3 fPolygonAttr; +uniform sampler2DArray Textures; + out vec4 oColor; out vec4 oAttr; -int TexcoordWrap(int c, int maxc, int mode) +float TexcoordWrap(float c, int mode) { if ((mode & (1<<0)) != 0) { - if ((mode & (1<<2)) != 0 && (c & maxc) != 0) - return (maxc-1) - (c & (maxc-1)); + if ((mode & (1<<2)) == 0) + return fract(c); else - return (c & (maxc-1)); + // mirrored repeat is the most complex one, so we let the hardware handle it + return c; } else - return clamp(c, 0, maxc-1); -} - -vec4 TextureFetch_A3I5(ivec2 addr, ivec4 st, int wrapmode) -{ - st.x = TexcoordWrap(st.x, st.z, wrapmode>>0); - st.y = TexcoordWrap(st.y, st.w, wrapmode>>1); - - addr.x += ((st.y * st.z) + st.x); - ivec4 pixel = ivec4(texelFetch(TexMem, ivec2(addr.x&0x3FF, addr.x>>10), 0)); - - pixel.a = (pixel.r & 0xE0); - pixel.a = (pixel.a >> 3) + (pixel.a >> 6); - pixel.r &= 0x1F; - - addr.y = (addr.y << 3) + pixel.r; - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - - return vec4(color.rgb, float(pixel.a)/31.0); -} - -vec4 TextureFetch_I2(ivec2 addr, ivec4 st, int wrapmode, float alpha0) -{ - st.x = TexcoordWrap(st.x, st.z, wrapmode>>0); - st.y = TexcoordWrap(st.y, st.w, wrapmode>>1); - - addr.x += ((st.y * st.z) + st.x) >> 2; - ivec4 pixel = ivec4(texelFetch(TexMem, ivec2(addr.x&0x3FF, addr.x>>10), 0)); - pixel.r >>= (2 * (st.x & 3)); - pixel.r &= 0x03; - - addr.y = (addr.y << 2) + pixel.r; - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - - return vec4(color.rgb, (pixel.r>0)?1:alpha0); -} - -vec4 TextureFetch_I4(ivec2 addr, ivec4 st, int wrapmode, float alpha0) -{ - st.x = TexcoordWrap(st.x, st.z, wrapmode>>0); - st.y = TexcoordWrap(st.y, st.w, wrapmode>>1); - - addr.x += ((st.y * st.z) + st.x) >> 1; - ivec4 pixel = ivec4(texelFetch(TexMem, ivec2(addr.x&0x3FF, addr.x>>10), 0)); - if ((st.x & 1) != 0) pixel.r >>= 4; - else pixel.r &= 0x0F; - - addr.y = (addr.y << 3) + pixel.r; - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - - return vec4(color.rgb, (pixel.r>0)?1:alpha0); -} - -vec4 TextureFetch_I8(ivec2 addr, ivec4 st, int wrapmode, float alpha0) -{ - st.x = TexcoordWrap(st.x, st.z, wrapmode>>0); - st.y = TexcoordWrap(st.y, st.w, wrapmode>>1); - - addr.x += ((st.y * st.z) + st.x); - ivec4 pixel = ivec4(texelFetch(TexMem, ivec2(addr.x&0x3FF, addr.x>>10), 0)); - - addr.y = (addr.y << 3) + pixel.r; - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - - return vec4(color.rgb, (pixel.r>0)?1:alpha0); -} - -vec4 TextureFetch_Compressed(ivec2 addr, ivec4 st, int wrapmode) -{ - st.x = TexcoordWrap(st.x, st.z, wrapmode>>0); - st.y = TexcoordWrap(st.y, st.w, wrapmode>>1); - - addr.x += ((st.y & 0x3FC) * (st.z>>2)) + (st.x & 0x3FC) + (st.y & 0x3); - ivec4 p = ivec4(texelFetch(TexMem, ivec2(addr.x&0x3FF, addr.x>>10), 0)); - int val = (p.r >> (2 * (st.x & 0x3))) & 0x3; - - int slot1addr = 0x20000 + ((addr.x & 0x1FFFC) >> 1); - if (addr.x >= 0x40000) slot1addr += 0x10000; - - int palinfo; - p = ivec4(texelFetch(TexMem, ivec2(slot1addr&0x3FF, slot1addr>>10), 0)); - palinfo = p.r; - slot1addr++; - p = ivec4(texelFetch(TexMem, ivec2(slot1addr&0x3FF, slot1addr>>10), 0)); - palinfo |= (p.r << 8); - - addr.y = (addr.y << 3) + ((palinfo & 0x3FFF) << 1); - palinfo >>= 14; - - if (val == 0) - { - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - return vec4(color.rgb, 1.0); - } - else if (val == 1) - { - addr.y++; - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - return vec4(color.rgb, 1.0); - } - else if (val == 2) - { - if (palinfo == 1) - { - vec4 color0 = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - addr.y++; - vec4 color1 = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - return vec4((color0.rgb + color1.rgb) / 2.0, 1.0); - } - else if (palinfo == 3) - { - vec4 color0 = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - addr.y++; - vec4 color1 = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - return vec4((color0.rgb*5.0 + color1.rgb*3.0) / 8.0, 1.0); - } - else - { - addr.y += 2; - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - return vec4(color.rgb, 1.0); - } - } - else - { - if (palinfo == 2) - { - addr.y += 3; - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - return vec4(color.rgb, 1.0); - } - else if (palinfo == 3) - { - vec4 color0 = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - addr.y++; - vec4 color1 = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - return vec4((color0.rgb*3.0 + color1.rgb*5.0) / 8.0, 1.0); - } - else - { - return vec4(0.0); - } - } -} - -vec4 TextureFetch_A5I3(ivec2 addr, ivec4 st, int wrapmode) -{ - st.x = TexcoordWrap(st.x, st.z, wrapmode>>0); - st.y = TexcoordWrap(st.y, st.w, wrapmode>>1); - - addr.x += ((st.y * st.z) + st.x); - ivec4 pixel = ivec4(texelFetch(TexMem, ivec2(addr.x&0x3FF, addr.x>>10), 0)); - - pixel.a = (pixel.r & 0xF8) >> 3; - pixel.r &= 0x07; - - addr.y = (addr.y << 3) + pixel.r; - vec4 color = texelFetch(TexPalMem, ivec2(addr.y&0x3FF, addr.y>>10), 0); - - return vec4(color.rgb, float(pixel.a)/31.0); -} - -vec4 TextureFetch_Direct(ivec2 addr, ivec4 st, int wrapmode) -{ - st.x = TexcoordWrap(st.x, st.z, wrapmode>>0); - st.y = TexcoordWrap(st.y, st.w, wrapmode>>1); - - addr.x += ((st.y * st.z) + st.x) << 1; - ivec4 pixelL = ivec4(texelFetch(TexMem, ivec2(addr.x&0x3FF, addr.x>>10), 0)); - addr.x++; - ivec4 pixelH = ivec4(texelFetch(TexMem, ivec2(addr.x&0x3FF, addr.x>>10), 0)); - - vec4 color; - color.r = float(pixelL.r & 0x1F) / 31.0; - color.g = float((pixelL.r >> 5) | ((pixelH.r & 0x03) << 3)) / 31.0; - color.b = float((pixelH.r & 0x7C) >> 2) / 31.0; - color.a = float(pixelH.r >> 7); - - return color; -} - -vec4 TextureLookup_Nearest(vec2 st) -{ - int attr = int(fPolygonAttr.y); - int paladdr = int(fPolygonAttr.z); - - float alpha0; - if ((attr & (1<<29)) != 0) alpha0 = 0.0; - else alpha0 = 1.0; - - int tw = 8 << ((attr >> 20) & 0x7); - int th = 8 << ((attr >> 23) & 0x7); - ivec4 st_full = ivec4(ivec2(st), tw, th); - - ivec2 vramaddr = ivec2((attr & 0xFFFF) << 3, paladdr); - int wrapmode = (attr >> 16); - - int type = (attr >> 26) & 0x7; - if (type == 5) return TextureFetch_Compressed(vramaddr, st_full, wrapmode); - else if (type == 2) return TextureFetch_I2 (vramaddr, st_full, wrapmode, alpha0); - else if (type == 3) return TextureFetch_I4 (vramaddr, st_full, wrapmode, alpha0); - else if (type == 4) return TextureFetch_I8 (vramaddr, st_full, wrapmode, alpha0); - else if (type == 1) return TextureFetch_A3I5 (vramaddr, st_full, wrapmode); - else if (type == 6) return TextureFetch_A5I3 (vramaddr, st_full, wrapmode); - else return TextureFetch_Direct (vramaddr, st_full, wrapmode); -} - -vec4 TextureLookup_Linear(vec2 texcoord) -{ - ivec2 intpart = ivec2(texcoord); - vec2 fracpart = fract(texcoord); - - int attr = int(fPolygonAttr.y); - int paladdr = int(fPolygonAttr.z); - - float alpha0; - if ((attr & (1<<29)) != 0) alpha0 = 0.0; - else alpha0 = 1.0; - - int tw = 8 << ((attr >> 20) & 0x7); - int th = 8 << ((attr >> 23) & 0x7); - ivec4 st_full = ivec4(intpart, tw, th); - - ivec2 vramaddr = ivec2((attr & 0xFFFF) << 3, paladdr); - int wrapmode = (attr >> 16); - - vec4 A, B, C, D; - int type = (attr >> 26) & 0x7; - if (type == 5) - { - A = TextureFetch_Compressed(vramaddr, st_full , wrapmode); - B = TextureFetch_Compressed(vramaddr, st_full + ivec4(1,0,0,0), wrapmode); - C = TextureFetch_Compressed(vramaddr, st_full + ivec4(0,1,0,0), wrapmode); - D = TextureFetch_Compressed(vramaddr, st_full + ivec4(1,1,0,0), wrapmode); - } - else if (type == 2) - { - A = TextureFetch_I2(vramaddr, st_full , wrapmode, alpha0); - B = TextureFetch_I2(vramaddr, st_full + ivec4(1,0,0,0), wrapmode, alpha0); - C = TextureFetch_I2(vramaddr, st_full + ivec4(0,1,0,0), wrapmode, alpha0); - D = TextureFetch_I2(vramaddr, st_full + ivec4(1,1,0,0), wrapmode, alpha0); - } - else if (type == 3) - { - A = TextureFetch_I4(vramaddr, st_full , wrapmode, alpha0); - B = TextureFetch_I4(vramaddr, st_full + ivec4(1,0,0,0), wrapmode, alpha0); - C = TextureFetch_I4(vramaddr, st_full + ivec4(0,1,0,0), wrapmode, alpha0); - D = TextureFetch_I4(vramaddr, st_full + ivec4(1,1,0,0), wrapmode, alpha0); - } - else if (type == 4) - { - A = TextureFetch_I8(vramaddr, st_full , wrapmode, alpha0); - B = TextureFetch_I8(vramaddr, st_full + ivec4(1,0,0,0), wrapmode, alpha0); - C = TextureFetch_I8(vramaddr, st_full + ivec4(0,1,0,0), wrapmode, alpha0); - D = TextureFetch_I8(vramaddr, st_full + ivec4(1,1,0,0), wrapmode, alpha0); - } - else if (type == 1) - { - A = TextureFetch_A3I5(vramaddr, st_full , wrapmode); - B = TextureFetch_A3I5(vramaddr, st_full + ivec4(1,0,0,0), wrapmode); - C = TextureFetch_A3I5(vramaddr, st_full + ivec4(0,1,0,0), wrapmode); - D = TextureFetch_A3I5(vramaddr, st_full + ivec4(1,1,0,0), wrapmode); - } - else if (type == 6) - { - A = TextureFetch_A5I3(vramaddr, st_full , wrapmode); - B = TextureFetch_A5I3(vramaddr, st_full + ivec4(1,0,0,0), wrapmode); - C = TextureFetch_A5I3(vramaddr, st_full + ivec4(0,1,0,0), wrapmode); - D = TextureFetch_A5I3(vramaddr, st_full + ivec4(1,1,0,0), wrapmode); - } - else - { - A = TextureFetch_Direct(vramaddr, st_full , wrapmode); - B = TextureFetch_Direct(vramaddr, st_full + ivec4(1,0,0,0), wrapmode); - C = TextureFetch_Direct(vramaddr, st_full + ivec4(0,1,0,0), wrapmode); - D = TextureFetch_Direct(vramaddr, st_full + ivec4(1,1,0,0), wrapmode); - } - - float fx = fracpart.x; - vec4 AB; - if (A.a < (0.5/31.0) && B.a < (0.5/31.0)) - AB = vec4(0); - else - { - //if (A.a < (0.5/31.0) || B.a < (0.5/31.0)) - // fx = step(0.5, fx); - - AB = mix(A, B, fx); - } - - fx = fracpart.x; - vec4 CD; - if (C.a < (0.5/31.0) && D.a < (0.5/31.0)) - CD = vec4(0); - else - { - //if (C.a < (0.5/31.0) || D.a < (0.5/31.0)) - // fx = step(0.5, fx); - - CD = mix(C, D, fx); - } - - fx = fracpart.y; - vec4 ret; - if (AB.a < (0.5/31.0) && CD.a < (0.5/31.0)) - ret = vec4(0); - else - { - //if (AB.a < (0.5/31.0) || CD.a < (0.5/31.0)) - // fx = step(0.5, fx); - - ret = mix(AB, CD, fx); - } - - return ret; + return clamp(c, 0, 1); } vec4 FinalColor() @@ -618,8 +304,12 @@ vec4 FinalColor() } else { - vec4 tcol = TextureLookup_Nearest(fTexcoord); - //vec4 tcol = TextureLookup_Linear(fTexcoord); + int wrapmode = (fPolygonAttr.y >> 16); + + vec2 st = vec2( + TexcoordWrap(fTexcoord.x, wrapmode>>0), + TexcoordWrap(fTexcoord.y, wrapmode>>1)); + vec4 tcol = texture(Textures, vec3(st, float(fPolygonAttr.z))); if ((blendmode & 1) != 0) { @@ -662,7 +352,10 @@ void main() fpos.xyz *= fpos.w; fColor = vec4(vColor) / vec4(255.0,255.0,255.0,31.0); - fTexcoord = vec2(vTexcoord) / 16.0; + vec2 texSize = vec2( + float(8 << ((vPolygonAttr.y >> 20) & 0x7)), + float(8 << ((vPolygonAttr.y >> 23) & 0x7))); + fTexcoord = vec2(vTexcoord) / (texSize * 16.0); fPolygonAttr = vPolygonAttr; gl_Position = fpos; @@ -685,7 +378,11 @@ void main() fpos.xy *= fpos.w; fColor = vec4(vColor) / vec4(255.0,255.0,255.0,31.0); - fTexcoord = vec2(vTexcoord) / 16.0; + + vec2 texSize = vec2( + float(8 << ((vPolygonAttr.y >> 20) & 0x7)), + float(8 << ((vPolygonAttr.y >> 23) & 0x7))); + fTexcoord = vec2(vTexcoord) / (texSize * 16.0); fPolygonAttr = vPolygonAttr; gl_Position = fpos; diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp index 151de492..571f76c9 100644 --- a/src/GPU3D_Soft.cpp +++ b/src/GPU3D_Soft.cpp @@ -1138,7 +1138,7 @@ void SetupPolygonRightEdge(RendererPolygon* rp, s32 y) void SetupPolygon(RendererPolygon* rp, Polygon* polygon, RendererPolygon* lastRp) { - if (polygon->TexParam & 0x1C000000) + if (polygon->TexParam & 0x1C000000 && !polygon->IsShadowMask) { if (lastRp && lastRp->PolyData->TexParam == polygon->TexParam && lastRp->PolyData->TexPalette == polygon->TexPalette) diff --git a/src/GPU3D_TexCache.cpp b/src/GPU3D_TexCache.cpp index 13d3388b..25e4478e 100644 --- a/src/GPU3D_TexCache.cpp +++ b/src/GPU3D_TexCache.cpp @@ -300,13 +300,13 @@ void EnsurePaletteCoherent(u64* mask) { updatePalette = true; int idx = __builtin_ctzll(updateField); - u32 map = GPU::VRAMMap_TexPal[idx >> 4 + i * 4]; + u32 map = GPU::VRAMMap_TexPal[(idx >> 4) + i * 4]; if (map && (map & (map - 1)) == 0) { u32 bank = __builtin_ctz(map); memcpy( PaletteCache + i * 0x10000 + idx * 0x400, - GPU::VRAM[bank] + ((idx * 0x400) & GPU::VRAMMask[bank]), + GPU::VRAM[bank] + ((idx * 0x400) & GPU::VRAMMask[bank]), 0x400); } else @@ -363,20 +363,20 @@ void UpdateTextures() else { // E - if (PaletteMap[i] & (1<<3)) + if (PaletteMap[i] & (1<<4)) { - PaletteDirty[i >> 2] |= GPU::LCDCDirty[3][0]; - PaletteCacheStatus[i >> 2] &= ~GPU::LCDCDirty[3][0]; - GPU::LCDCDirty[3][0] = 0; + PaletteDirty[i >> 2] |= GPU::LCDCDirty[4][0]; + PaletteCacheStatus[i >> 2] &= ~GPU::LCDCDirty[4][0]; + GPU::LCDCDirty[4][0] = 0; } // FG for (int j = 0; j < 2; j++) { - if (PaletteMap[i] & (1<<(4+j))) + if (PaletteMap[i] & (1<<(5+j))) { - PaletteDirty[i >> 2] |= GPU::LCDCDirty[4+j][0] << (i & 0x3) * 16; - PaletteCacheStatus[i >> 2] &= ~(GPU::LCDCDirty[4+j][0] << (i & 0x3) * 16); - GPU::LCDCDirty[4+j][0] = 0; + PaletteDirty[i >> 2] |= GPU::LCDCDirty[5+j][0] << (i & 0x3) * 16; + PaletteCacheStatus[i >> 2] &= ~((GPU::LCDCDirty[5+j][0] & 0xFFFF) << (i & 0x3) * 16); + GPU::LCDCDirty[5+j][0] = 0; } } } @@ -405,6 +405,8 @@ void UpdateTextures() if (GPU3D::Renderer == 0) SoftRenderer::FreeTexture(it->second.Handle, width, height); + else + GLRenderer::FreeTexture(it->second.Handle, width, height); it = TextureCache.erase(it); } @@ -462,7 +464,7 @@ ExternalTexHandle GetTexture(u32 texParam, u32 palBase) u32* data = GPU3D::Renderer == 0 ? SoftRenderer::AllocateTexture(&texture.Handle, width, height) - : NULL; + : GLRenderer::AllocateTexture(&texture.Handle, width, height); memset(texture.TextureMask, 0, 8*8); memset(texture.PaletteMask, 0, 8*2); @@ -525,7 +527,7 @@ ExternalTexHandle GetTexture(u32 texParam, u32 palBase) } if (GPU3D::Renderer == 1) - {} + GLRenderer::FinaliseTexture(texture.Handle, width, height); } return texture.Handle; diff --git a/src/OpenGLSupport.h b/src/OpenGLSupport.h index a08bcdf1..86dfb63d 100644 --- a/src/OpenGLSupport.h +++ b/src/OpenGLSupport.h @@ -63,6 +63,9 @@ #define DO_PROCLIST(func) \ DO_PROCLIST_1_3(func) \ \ + func(GLTEXIMAGE3D, glTexImage3D); \ + func(GLTEXSUBIMAGE3D, glTexSubImage3D); \ + \ func(GLGENFRAMEBUFFERS, glGenFramebuffers); \ func(GLDELETEFRAMEBUFFERS, glDeleteFramebuffers); \ func(GLBINDFRAMEBUFFER, glBindFramebuffer); \ diff --git a/src/xbrz/xbrz.cpp b/src/xbrz/xbrz.cpp new file mode 100644 index 00000000..8c90fc80 --- /dev/null +++ b/src/xbrz/xbrz.cpp @@ -0,0 +1,1356 @@ +// **************************************************************************** +// * This file is part of the xBRZ project. It is distributed under * +// * GNU General Public License: https://www.gnu.org/licenses/gpl-3.0 * +// * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved * +// * * +// * Additionally and as a special exception, the author gives permission * +// * to link the code of this program with the following libraries * +// * (or with modified versions that use the same licenses), and distribute * +// * linked combinations including the two: MAME, FreeFileSync, Snes9x, ePSXe * +// * You must obey the GNU General Public License in all respects for all of * +// * the code used other than MAME, FreeFileSync, Snes9x, ePSXe. * +// * If you modify this file, you may extend this exception to your version * +// * of the file, but you are not obligated to do so. If you do not wish to * +// * do so, delete this exception statement from your version. * +// **************************************************************************** + +#include "xbrz.h" +#include +#include +#include +#include //std::sqrt +#include "xbrz_tools.h" + +using namespace xbrz; + + +namespace +{ +template inline +uint32_t gradientRGB(uint32_t pixFront, uint32_t pixBack) //blend front color with opacity M / N over opaque background: https://en.wikipedia.org/wiki/Alpha_compositing#Alpha_blending +{ + static_assert(0 < M && M < N && N <= 1000); + + auto calcColor = [](unsigned char colFront, unsigned char colBack) -> unsigned char { return (colFront * M + colBack * (N - M)) / N; }; + + return makePixel(calcColor(getRed (pixFront), getRed (pixBack)), + calcColor(getGreen(pixFront), getGreen(pixBack)), + calcColor(getBlue (pixFront), getBlue (pixBack))); +} + + +template inline +uint32_t gradientARGB(uint32_t pixFront, uint32_t pixBack) //find intermediate color between two colors with alpha channels (=> NO alpha blending!!!) +{ + static_assert(0 < M && M < N && N <= 1000); + + const unsigned int weightFront = getAlpha(pixFront) * M; + const unsigned int weightBack = getAlpha(pixBack) * (N - M); + const unsigned int weightSum = weightFront + weightBack; + if (weightSum == 0) + return 0; + + auto calcColor = [=](unsigned char colFront, unsigned char colBack) + { + return static_cast((colFront * weightFront + colBack * weightBack) / weightSum); + }; + + return makePixel(static_cast(weightSum / N), + calcColor(getRed (pixFront), getRed (pixBack)), + calcColor(getGreen(pixFront), getGreen(pixBack)), + calcColor(getBlue (pixFront), getBlue (pixBack))); +} + + +//inline +//double fastSqrt(double n) +//{ +// __asm //speeds up xBRZ by about 9% compared to std::sqrt which internally uses the same assembler instructions but adds some "fluff" +// { +// fld n +// fsqrt +// } +//} +// + + +#ifdef _MSC_VER + #define FORCE_INLINE __forceinline +#elif defined __GNUC__ + #define FORCE_INLINE __attribute__((always_inline)) inline +#else + #define FORCE_INLINE inline +#endif + + +enum RotationDegree //clock-wise +{ + ROT_0, + ROT_90, + ROT_180, + ROT_270 +}; + +//calculate input matrix coordinates after rotation at compile time +template +struct MatrixRotation; + +template +struct MatrixRotation +{ + static const size_t I_old = I; + static const size_t J_old = J; +}; + +template //(i, j) = (row, col) indices, N = size of (square) matrix +struct MatrixRotation +{ + static const size_t I_old = N - 1 - MatrixRotation(rotDeg - 1), I, J, N>::J_old; //old coordinates before rotation! + static const size_t J_old = MatrixRotation(rotDeg - 1), I, J, N>::I_old; // +}; + + +template +class OutputMatrix +{ +public: + OutputMatrix(uint32_t* out, int outWidth) : //access matrix area, top-left at position "out" for image with given width + out_(out), + outWidth_(outWidth) {} + + template + uint32_t& ref() const + { + static const size_t I_old = MatrixRotation::I_old; + static const size_t J_old = MatrixRotation::J_old; + return *(out_ + J_old + I_old * outWidth_); + } + +private: + uint32_t* out_; + const int outWidth_; +}; + + +template inline +T square(T value) { return value * value; } + + +#if 0 +inline +double distRGB(uint32_t pix1, uint32_t pix2) +{ + const double r_diff = static_cast(getRed (pix1)) - getRed (pix2); + const double g_diff = static_cast(getGreen(pix1)) - getGreen(pix2); + const double b_diff = static_cast(getBlue (pix1)) - getBlue (pix2); + + //euklidean RGB distance + return std::sqrt(square(r_diff) + square(g_diff) + square(b_diff)); +} +#endif + + +inline +double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight) +{ + //https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion + //YCbCr conversion is a matrix multiplication => take advantage of linearity by subtracting first! + const int r_diff = static_cast(getRed (pix1)) - getRed (pix2); //we may delay division by 255 to after matrix multiplication + const int g_diff = static_cast(getGreen(pix1)) - getGreen(pix2); // + const int b_diff = static_cast(getBlue (pix1)) - getBlue (pix2); //substraction for int is noticeable faster than for double! + + //const double k_b = 0.0722; //ITU-R BT.709 conversion + //const double k_r = 0.2126; // + const double k_b = 0.0593; //ITU-R BT.2020 conversion + const double k_r = 0.2627; // + const double k_g = 1 - k_b - k_r; + + const double scale_b = 0.5 / (1 - k_b); + const double scale_r = 0.5 / (1 - k_r); + + const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr! + const double c_b = scale_b * (b_diff - y); + const double c_r = scale_r * (r_diff - y); + + //we skip division by 255 to have similar range like other distance functions + return std::sqrt(square(lumaWeight * y) + square(c_b) + square(c_r)); +} + + +inline +double distYCbCrBuffered(uint32_t pix1, uint32_t pix2) +{ + //30% perf boost compared to plain distYCbCr()! + //consumes 64 MB memory; using double is only 2% faster, but takes 128 MB + static const std::vector diffToDist = [] + { + std::vector tmp; + + for (uint32_t i = 0; i < 256 * 256 * 256; ++i) //startup time: 114 ms on Intel Core i5 (four cores) + { + const int r_diff = static_cast(getByte<2>(i)) * 2; + const int g_diff = static_cast(getByte<1>(i)) * 2; + const int b_diff = static_cast(getByte<0>(i)) * 2; + + const double k_b = 0.0593; //ITU-R BT.2020 conversion + const double k_r = 0.2627; // + const double k_g = 1 - k_b - k_r; + + const double scale_b = 0.5 / (1 - k_b); + const double scale_r = 0.5 / (1 - k_r); + + const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr! + const double c_b = scale_b * (b_diff - y); + const double c_r = scale_r * (r_diff - y); + + tmp.push_back(static_cast(std::sqrt(square(y) + square(c_b) + square(c_r)))); + } + return tmp; + }(); + + //if (pix1 == pix2) -> 8% perf degradation! + // return 0; + //if (pix1 < pix2) + // std::swap(pix1, pix2); -> 30% perf degradation!!! + + const int r_diff = static_cast(getRed (pix1)) - getRed (pix2); + const int g_diff = static_cast(getGreen(pix1)) - getGreen(pix2); + const int b_diff = static_cast(getBlue (pix1)) - getBlue (pix2); + + const size_t index = (static_cast(r_diff / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte + (static_cast(g_diff / 2) << 8) | + (static_cast(b_diff / 2)); + +#if 0 //attention: the following calculation creates an asymmetric color distance!!! (e.g. r_diff=46 will be unpacked as 45, but r_diff=-46 unpacks to -47 + const size_t index = (((r_diff + 0xFF) / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte + (((g_diff + 0xFF) / 2) << 8) | + (( b_diff + 0xFF) / 2); +#endif + return diffToDist[index]; +} + + +#if defined _MSC_VER && !defined NDEBUG + const int debugPixelX = -1; + const int debugPixelY = 58; + + thread_local bool breakIntoDebugger = false; +#endif + + +enum BlendType +{ + BLEND_NONE = 0, + BLEND_NORMAL, //a normal indication to blend + BLEND_DOMINANT, //a strong indication to blend + //attention: BlendType must fit into the value range of 2 bit!!! +}; + +struct BlendResult +{ + BlendType + /**/blend_f, blend_g, + /**/blend_j, blend_k; +}; + + +struct Kernel_3x3 +{ + uint32_t + a, b, c, + d, e, f, + g, h, i; +}; + +struct Kernel_4x4 //kernel for preprocessing step +{ + uint32_t + a, b, c, // + e, f, g, // support reinterpret_cast from Kernel_4x4 => Kernel_3x3 + i, j, k, // + m, n, o, + d, h, l, p; +}; + +/* input kernel area naming convention: +----------------- +| A | B | C | D | +|---|---|---|---| +| E | F | G | H | evaluate the four corners between F, G, J, K +|---|---|---|---| input pixel is at position F +| I | J | K | L | +|---|---|---|---| +| M | N | O | P | +----------------- +*/ +template +FORCE_INLINE //detect blend direction +BlendResult preProcessCorners(const Kernel_4x4& ker, const xbrz::ScalerCfg& cfg) //result: F, G, J, K corners of "GradientType" +{ +#if defined _MSC_VER && !defined NDEBUG + if (breakIntoDebugger) + __debugbreak(); //__asm int 3; +#endif + + BlendResult result = {}; + + if ((ker.f == ker.g && + ker.j == ker.k) || + (ker.f == ker.j && + ker.g == ker.k)) + return result; + + auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight); }; + + double jg = dist(ker.i, ker.f) + dist(ker.f, ker.c) + dist(ker.n, ker.k) + dist(ker.k, ker.h) + cfg.centerDirectionBias * dist(ker.j, ker.g); + double fk = dist(ker.e, ker.j) + dist(ker.j, ker.o) + dist(ker.b, ker.g) + dist(ker.g, ker.l) + cfg.centerDirectionBias * dist(ker.f, ker.k); + + if (jg < fk) //test sample: 70% of values max(jg, fk) / min(jg, fk) are between 1.1 and 3.7 with median being 1.8 + { + const bool dominantGradient = cfg.dominantDirectionThreshold * jg < fk; + if (ker.f != ker.g && ker.f != ker.j) + result.blend_f = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL; + + if (ker.k != ker.j && ker.k != ker.g) + result.blend_k = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL; + } + else if (fk < jg) + { + const bool dominantGradient = cfg.dominantDirectionThreshold * fk < jg; + if (ker.j != ker.f && ker.j != ker.k) + result.blend_j = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL; + + if (ker.g != ker.f && ker.g != ker.k) + result.blend_g = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL; + } + return result; +} + +#define DEF_GETTER(x) template uint32_t inline get_##x(const Kernel_3x3& ker) { return ker.x; } +//we cannot and NEED NOT write "ker.##x" since ## concatenates preprocessor tokens but "." is not a token +DEF_GETTER(a) DEF_GETTER(b) DEF_GETTER(c) +DEF_GETTER(d) DEF_GETTER(e) DEF_GETTER(f) +DEF_GETTER(g) DEF_GETTER(h) DEF_GETTER(i) +#undef DEF_GETTER + +#define DEF_GETTER(x, y) template <> inline uint32_t get_##x(const Kernel_3x3& ker) { return ker.y; } +DEF_GETTER(a, g) DEF_GETTER(b, d) DEF_GETTER(c, a) +DEF_GETTER(d, h) DEF_GETTER(e, e) DEF_GETTER(f, b) +DEF_GETTER(g, i) DEF_GETTER(h, f) DEF_GETTER(i, c) +#undef DEF_GETTER + +#define DEF_GETTER(x, y) template <> inline uint32_t get_##x(const Kernel_3x3& ker) { return ker.y; } +DEF_GETTER(a, i) DEF_GETTER(b, h) DEF_GETTER(c, g) +DEF_GETTER(d, f) DEF_GETTER(e, e) DEF_GETTER(f, d) +DEF_GETTER(g, c) DEF_GETTER(h, b) DEF_GETTER(i, a) +#undef DEF_GETTER + +#define DEF_GETTER(x, y) template <> inline uint32_t get_##x(const Kernel_3x3& ker) { return ker.y; } +DEF_GETTER(a, c) DEF_GETTER(b, f) DEF_GETTER(c, i) +DEF_GETTER(d, b) DEF_GETTER(e, e) DEF_GETTER(f, h) +DEF_GETTER(g, a) DEF_GETTER(h, d) DEF_GETTER(i, g) +#undef DEF_GETTER + + +//compress four blend types into a single byte +//inline BlendType getTopL (unsigned char b) { return static_cast(0x3 & b); } +inline BlendType getTopR (unsigned char b) { return static_cast(0x3 & (b >> 2)); } +inline BlendType getBottomR(unsigned char b) { return static_cast(0x3 & (b >> 4)); } +inline BlendType getBottomL(unsigned char b) { return static_cast(0x3 & (b >> 6)); } + +inline void clearAddTopL(unsigned char& b, BlendType bt) { b = static_cast(bt); } +inline void addTopR (unsigned char& b, BlendType bt) { b |= (bt << 2); } //buffer is assumed to be initialized before preprocessing! +inline void addBottomR (unsigned char& b, BlendType bt) { b |= (bt << 4); } //e.g. via clearAddTopL() +inline void addBottomL (unsigned char& b, BlendType bt) { b |= (bt << 6); } // + +inline bool blendingNeeded(unsigned char b) +{ + static_assert(BLEND_NONE == 0); + return b != 0; +} + +template inline +unsigned char rotateBlendInfo(unsigned char b) { return b; } +template <> inline unsigned char rotateBlendInfo(unsigned char b) { return ((b << 2) | (b >> 6)) & 0xff; } +template <> inline unsigned char rotateBlendInfo(unsigned char b) { return ((b << 4) | (b >> 4)) & 0xff; } +template <> inline unsigned char rotateBlendInfo(unsigned char b) { return ((b << 6) | (b >> 2)) & 0xff; } + + +/* input kernel area naming convention: +------------- +| A | B | C | +|---|---|---| +| D | E | F | input pixel is at position E +|---|---|---| +| G | H | I | +------------- +*/ +template +FORCE_INLINE //perf: quite worth it! +void blendPixel(const Kernel_3x3& ker, + uint32_t* target, int trgWidth, + unsigned char blendInfo, //result of preprocessing all four corners of pixel "e" + const xbrz::ScalerCfg& cfg) +{ + //#define a get_a(ker) +#define b get_b(ker) +#define c get_c(ker) +#define d get_d(ker) +#define e get_e(ker) +#define f get_f(ker) +#define g get_g(ker) +#define h get_h(ker) +#define i get_i(ker) + +#if defined _MSC_VER && !defined NDEBUG + if (breakIntoDebugger) + __debugbreak(); //__asm int 3; +#endif + + const unsigned char blend = rotateBlendInfo(blendInfo); + + if (getBottomR(blend) >= BLEND_NORMAL) + { + auto eq = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight) < cfg.equalColorTolerance; }; + auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight); }; + + const bool doLineBlend = [&]() -> bool + { + if (getBottomR(blend) >= BLEND_DOMINANT) + return true; + + //make sure there is no second blending in an adjacent rotation for this pixel: handles insular pixels, mario eyes + if (getTopR(blend) != BLEND_NONE && !eq(e, g)) //but support double-blending for 90� corners + return false; + if (getBottomL(blend) != BLEND_NONE && !eq(e, c)) + return false; + + //no full blending for L-shapes; blend corner only (handles "mario mushroom eyes") + if (!eq(e, i) && eq(g, h) && eq(h, i) && eq(i, f) && eq(f, c)) + return false; + + return true; + }(); + + const uint32_t px = dist(e, f) <= dist(e, h) ? f : h; //choose most similar color + + OutputMatrix out(target, trgWidth); + + if (doLineBlend) + { + const double fg = dist(f, g); //test sample: 70% of values max(fg, hc) / min(fg, hc) are between 1.1 and 3.7 with median being 1.9 + const double hc = dist(h, c); // + + const bool haveShallowLine = cfg.steepDirectionThreshold * fg <= hc && e != g && d != g; + const bool haveSteepLine = cfg.steepDirectionThreshold * hc <= fg && e != c && b != c; + + if (haveShallowLine) + { + if (haveSteepLine) + Scaler::blendLineSteepAndShallow(px, out); + else + Scaler::blendLineShallow(px, out); + } + else + { + if (haveSteepLine) + Scaler::blendLineSteep(px, out); + else + Scaler::blendLineDiagonal(px, out); + } + } + else + Scaler::blendCorner(px, out); + } + + //#undef a +#undef b +#undef c +#undef d +#undef e +#undef f +#undef g +#undef h +#undef i +} + + +class OobReaderTransparent +{ +public: + OobReaderTransparent(const uint32_t* src, int srcWidth, int srcHeight, int y) : + s_m1(0 <= y - 1 && y - 1 < srcHeight ? src + srcWidth * (y - 1) : nullptr), + s_0 (0 <= y && y < srcHeight ? src + srcWidth * y : nullptr), + s_p1(0 <= y + 1 && y + 1 < srcHeight ? src + srcWidth * (y + 1) : nullptr), + s_p2(0 <= y + 2 && y + 2 < srcHeight ? src + srcWidth * (y + 2) : nullptr), + srcWidth_(srcWidth) {} + + void readDhlp(Kernel_4x4& ker, int x) const //(x, y) is at kernel position F + { + [[likely]] if (const int x_p2 = x + 2; 0 <= x_p2 && x_p2 < srcWidth_) + { + ker.d = s_m1 ? s_m1[x_p2] : 0; + ker.h = s_0 ? s_0 [x_p2] : 0; + ker.l = s_p1 ? s_p1[x_p2] : 0; + ker.p = s_p2 ? s_p2[x_p2] : 0; + } + else + { + ker.d = 0; + ker.h = 0; + ker.l = 0; + ker.p = 0; + } + } + +private: + const uint32_t* const s_m1; + const uint32_t* const s_0; + const uint32_t* const s_p1; + const uint32_t* const s_p2; + const int srcWidth_; +}; + + +class OobReaderDuplicate +{ +public: + OobReaderDuplicate(const uint32_t* src, int srcWidth, int srcHeight, int y) : + s_m1(src + srcWidth * std::clamp(y - 1, 0, srcHeight - 1)), + s_0 (src + srcWidth * std::clamp(y, 0, srcHeight - 1)), + s_p1(src + srcWidth * std::clamp(y + 1, 0, srcHeight - 1)), + s_p2(src + srcWidth * std::clamp(y + 2, 0, srcHeight - 1)), + srcWidth_(srcWidth) {} + + void readDhlp(Kernel_4x4& ker, int x) const //(x, y) is at kernel position F + { + const int x_p2 = std::clamp(x + 2, 0, srcWidth_ - 1); + ker.d = s_m1[x_p2]; + ker.h = s_0 [x_p2]; + ker.l = s_p1[x_p2]; + ker.p = s_p2[x_p2]; + } + +private: + const uint32_t* const s_m1; + const uint32_t* const s_0; + const uint32_t* const s_p1; + const uint32_t* const s_p2; + const int srcWidth_; +}; + + +template //scaler policy: see "Scaler2x" reference implementation +void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, const xbrz::ScalerCfg& cfg, int yFirst, int yLast) +{ + yFirst = std::max(yFirst, 0); + yLast = std::min(yLast, srcHeight); + if (yFirst >= yLast || srcWidth <= 0) + return; + + const int trgWidth = srcWidth * Scaler::scale; + + //(ab)use space of "sizeof(uint32_t) * srcWidth * Scaler::scale" at the end of the image as temporary + //buffer for "on the fly preprocessing" without risk of accidental overwriting before accessing + unsigned char* const preProcBuf = reinterpret_cast(trg + yLast * Scaler::scale * trgWidth) - srcWidth; + + //initialize preprocessing buffer for first row of current stripe: detect upper left and right corner blending + //this cannot be optimized for adjacent processing stripes; we must not allow for a memory race condition! + { + const OobReader oobReader(src, srcWidth, srcHeight, yFirst - 1); + + //initialize at position x = -1 + Kernel_4x4 ker4 = {}; + oobReader.readDhlp(ker4, -4); //hack: read a, e, i, m at x = -1 + ker4.a = ker4.d; + ker4.e = ker4.h; + ker4.i = ker4.l; + ker4.m = ker4.p; + + oobReader.readDhlp(ker4, -3); + ker4.b = ker4.d; + ker4.f = ker4.h; + ker4.j = ker4.l; + ker4.n = ker4.p; + + oobReader.readDhlp(ker4, -2); + ker4.c = ker4.d; + ker4.g = ker4.h; + ker4.k = ker4.l; + ker4.o = ker4.p; + + oobReader.readDhlp(ker4, -1); + + { + const BlendResult res = preProcessCorners(ker4, cfg); + clearAddTopL(preProcBuf[0], res.blend_k); //set 1st known corner for (0, yFirst) + } + + for (int x = 0; x < srcWidth; ++x) + { + ker4.a = ker4.b; //shift previous kernel to the left + ker4.e = ker4.f; // ----------------- + ker4.i = ker4.j; // | A | B | C | D | + ker4.m = ker4.n; // |---|---|---|---| + /**/ // | E | F | G | H | (x, yFirst - 1) is at position F + ker4.b = ker4.c; // |---|---|---|---| + ker4.f = ker4.g; // | I | J | K | L | + ker4.j = ker4.k; // |---|---|---|---| + ker4.n = ker4.o; // | M | N | O | P | + /**/ // ----------------- + ker4.c = ker4.d; + ker4.g = ker4.h; + ker4.k = ker4.l; + ker4.o = ker4.p; + + oobReader.readDhlp(ker4, x); + + /* preprocessing blend result: + --------- + | F | G | evaluate corner between F, G, J, K + |---+---| current input pixel is at position F + | J | K | + --------- */ + const BlendResult res = preProcessCorners(ker4, cfg); + addTopR(preProcBuf[x], res.blend_j); //set 2nd known corner for (x, yFirst) + + if (x + 1 < srcWidth) + clearAddTopL(preProcBuf[x + 1], res.blend_k); //set 1st known corner for (x + 1, yFirst) + } + } + //------------------------------------------------------------------------------------ + + for (int y = yFirst; y < yLast; ++y) + { + uint32_t* out = trg + Scaler::scale * y * trgWidth; //consider MT "striped" access + + const OobReader oobReader(src, srcWidth, srcHeight, y); + + //initialize at position x = -1 + Kernel_4x4 ker4 = {}; + oobReader.readDhlp(ker4, -4); //hack: read a, e, i, m at x = -1 + ker4.a = ker4.d; + ker4.e = ker4.h; + ker4.i = ker4.l; + ker4.m = ker4.p; + + oobReader.readDhlp(ker4, -3); + ker4.b = ker4.d; + ker4.f = ker4.h; + ker4.j = ker4.l; + ker4.n = ker4.p; + + oobReader.readDhlp(ker4, -2); + ker4.c = ker4.d; + ker4.g = ker4.h; + ker4.k = ker4.l; + ker4.o = ker4.p; + + oobReader.readDhlp(ker4, -1); + + unsigned char blend_xy1 = 0; //corner blending for current (x, y + 1) position + { + const BlendResult res = preProcessCorners(ker4, cfg); + clearAddTopL(blend_xy1, res.blend_k); //set 1st known corner for (0, y + 1) and buffer for use on next column + + addBottomL(preProcBuf[0], res.blend_g); //set 3rd known corner for (0, y) + } + + for (int x = 0; x < srcWidth; ++x, out += Scaler::scale) + { +#if defined _MSC_VER && !defined NDEBUG + breakIntoDebugger = debugPixelX == x && debugPixelY == y; +#endif + ker4.a = ker4.b; //shift previous kernel to the left + ker4.e = ker4.f; // ----------------- + ker4.i = ker4.j; // | A | B | C | D | + ker4.m = ker4.n; // |---|---|---|---| + /**/ // | E | F | G | H | (x, y) is at position F + ker4.b = ker4.c; // |---|---|---|---| + ker4.f = ker4.g; // | I | J | K | L | + ker4.j = ker4.k; // |---|---|---|---| + ker4.n = ker4.o; // | M | N | O | P | + /**/ // ----------------- + ker4.c = ker4.d; + ker4.g = ker4.h; + ker4.k = ker4.l; + ker4.o = ker4.p; + + oobReader.readDhlp(ker4, x); + + //evaluate the four corners on bottom-right of current pixel + unsigned char blend_xy = preProcBuf[x]; //for current (x, y) position + { + /* preprocessing blend result: + --------- + | F | G | evaluate corner between F, G, J, K + |---+---| current input pixel is at position F + | J | K | + --------- */ + const BlendResult res = preProcessCorners(ker4, cfg); + addBottomR(blend_xy, res.blend_f); //all four corners of (x, y) have been determined at this point due to processing sequence! + + addTopR(blend_xy1, res.blend_j); //set 2nd known corner for (x, y + 1) + preProcBuf[x] = blend_xy1; //store on current buffer position for use on next row + + [[likely]] if (x + 1 < srcWidth) + { + //blend_xy1 -> blend_x1y1 + clearAddTopL(blend_xy1, res.blend_k); //set 1st known corner for (x + 1, y + 1) and buffer for use on next column + + addBottomL(preProcBuf[x + 1], res.blend_g); //set 3rd known corner for (x + 1, y) + } + } + + //fill block of size scale * scale with the given color + fillBlock(out, trgWidth * sizeof(uint32_t), ker4.f, Scaler::scale, Scaler::scale); + //place *after* preprocessing step, to not overwrite the results while processing the last pixel! + + //blend all four corners of current pixel + if (blendingNeeded(blend_xy)) + { + const auto& ker3 = reinterpret_cast(ker4); //"The Things We Do for Perf" + blendPixel(ker3, out, trgWidth, blend_xy, cfg); + blendPixel(ker3, out, trgWidth, blend_xy, cfg); + blendPixel(ker3, out, trgWidth, blend_xy, cfg); + blendPixel(ker3, out, trgWidth, blend_xy, cfg); + } + } + } +} + +//------------------------------------------------------------------------------------ + +template +struct Scaler2x : public ColorGradient +{ + static const int scale = 2; + + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + + template + static void blendLineShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + } + + template + static void blendLineSteep(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + } + + template + static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<1, 0>(), col); + alphaGrad<1, 4>(out.template ref<0, 1>(), col); + alphaGrad<5, 6>(out.template ref<1, 1>(), col); //[!] fixes 7/8 used in xBR + } + + template + static void blendLineDiagonal(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 2>(out.template ref<1, 1>(), col); + } + + template + static void blendCorner(uint32_t col, OutputMatrix& out) + { + //model a round corner + alphaGrad<21, 100>(out.template ref<1, 1>(), col); //exact: 1 - pi/4 = 0.2146018366 + } +}; + + +template +struct Scaler3x : public ColorGradient +{ + static const int scale = 3; + + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + + template + static void blendLineShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + + alphaGrad<3, 4>(out.template ref(), col); + out.template ref() = col; + } + + template + static void blendLineSteep(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + out.template ref<2, scale - 1>() = col; + } + + template + static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<2, 0>(), col); + alphaGrad<1, 4>(out.template ref<0, 2>(), col); + alphaGrad<3, 4>(out.template ref<2, 1>(), col); + alphaGrad<3, 4>(out.template ref<1, 2>(), col); + out.template ref<2, 2>() = col; + } + + template + static void blendLineDiagonal(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 8>(out.template ref<1, 2>(), col); //conflict with other rotations for this odd scale + alphaGrad<1, 8>(out.template ref<2, 1>(), col); + alphaGrad<7, 8>(out.template ref<2, 2>(), col); // + } + + template + static void blendCorner(uint32_t col, OutputMatrix& out) + { + //model a round corner + alphaGrad<45, 100>(out.template ref<2, 2>(), col); //exact: 0.4545939598 + //alphaGrad<7, 256>(out.template ref<2, 1>(), col); //0.02826017254 -> negligible + avoid conflicts with other rotations for this odd scale + //alphaGrad<7, 256>(out.template ref<1, 2>(), col); //0.02826017254 + } +}; + + +template +struct Scaler4x : public ColorGradient +{ + static const int scale = 4; + + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + + template + static void blendLineShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendLineSteep(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col); + + out.template ref<2, scale - 1>() = col; + out.template ref<3, scale - 1>() = col; + } + + template + static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<3, 4>(out.template ref<3, 1>(), col); + alphaGrad<3, 4>(out.template ref<1, 3>(), col); + alphaGrad<1, 4>(out.template ref<3, 0>(), col); + alphaGrad<1, 4>(out.template ref<0, 3>(), col); + + alphaGrad<1, 3>(out.template ref<2, 2>(), col); //[!] fixes 1/4 used in xBR + + out.template ref<3, 3>() = col; + out.template ref<3, 2>() = col; + out.template ref<2, 3>() = col; + } + + template + static void blendLineDiagonal(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 2>(out.template ref(), col); + alphaGrad<1, 2>(out.template ref(), col); + out.template ref() = col; + } + + template + static void blendCorner(uint32_t col, OutputMatrix& out) + { + //model a round corner + alphaGrad<68, 100>(out.template ref<3, 3>(), col); //exact: 0.6848532563 + alphaGrad< 9, 100>(out.template ref<3, 2>(), col); //0.08677704501 + alphaGrad< 9, 100>(out.template ref<2, 3>(), col); //0.08677704501 + } +}; + + +template +struct Scaler5x : public ColorGradient +{ + static const int scale = 5; + + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + + template + static void blendLineShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + + out.template ref() = col; + out.template ref() = col; + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendLineSteep(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<1, 4>(out.template ref<4, scale - 3>(), col); + + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col); + + out.template ref<2, scale - 1>() = col; + out.template ref<3, scale - 1>() = col; + out.template ref<4, scale - 1>() = col; + out.template ref<4, scale - 2>() = col; + } + + template + static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + + alphaGrad<2, 3>(out.template ref<3, 3>(), col); + + out.template ref<2, scale - 1>() = col; + out.template ref<3, scale - 1>() = col; + out.template ref<4, scale - 1>() = col; + + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendLineDiagonal(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 8>(out.template ref(), col); //conflict with other rotations for this odd scale + alphaGrad<1, 8>(out.template ref(), col); + alphaGrad<1, 8>(out.template ref(), col); // + + alphaGrad<7, 8>(out.template ref<4, 3>(), col); + alphaGrad<7, 8>(out.template ref<3, 4>(), col); + + out.template ref<4, 4>() = col; + } + + template + static void blendCorner(uint32_t col, OutputMatrix& out) + { + //model a round corner + alphaGrad<86, 100>(out.template ref<4, 4>(), col); //exact: 0.8631434088 + alphaGrad<23, 100>(out.template ref<4, 3>(), col); //0.2306749731 + alphaGrad<23, 100>(out.template ref<3, 4>(), col); //0.2306749731 + //alphaGrad<1, 64>(out.template ref<4, 2>(), col); //0.01676812367 -> negligible + avoid conflicts with other rotations for this odd scale + //alphaGrad<1, 64>(out.template ref<2, 4>(), col); //0.01676812367 + } +}; + + +template +struct Scaler6x : public ColorGradient +{ + static const int scale = 6; + + template //bring template function into scope for GCC + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) { ColorGradient::template alphaGrad(pixBack, pixFront); } + + + template + static void blendLineShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + + out.template ref() = col; + out.template ref() = col; + out.template ref() = col; + out.template ref() = col; + + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendLineSteep(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<1, 4>(out.template ref<4, scale - 3>(), col); + + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col); + alphaGrad<3, 4>(out.template ref<5, scale - 3>(), col); + + out.template ref<2, scale - 1>() = col; + out.template ref<3, scale - 1>() = col; + out.template ref<4, scale - 1>() = col; + out.template ref<5, scale - 1>() = col; + + out.template ref<4, scale - 2>() = col; + out.template ref<5, scale - 2>() = col; + } + + template + static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 4>(out.template ref<0, scale - 1>(), col); + alphaGrad<1, 4>(out.template ref<2, scale - 2>(), col); + alphaGrad<3, 4>(out.template ref<1, scale - 1>(), col); + alphaGrad<3, 4>(out.template ref<3, scale - 2>(), col); + + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<1, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + alphaGrad<3, 4>(out.template ref(), col); + + out.template ref<2, scale - 1>() = col; + out.template ref<3, scale - 1>() = col; + out.template ref<4, scale - 1>() = col; + out.template ref<5, scale - 1>() = col; + + out.template ref<4, scale - 2>() = col; + out.template ref<5, scale - 2>() = col; + + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendLineDiagonal(uint32_t col, OutputMatrix& out) + { + alphaGrad<1, 2>(out.template ref(), col); + alphaGrad<1, 2>(out.template ref(), col); + alphaGrad<1, 2>(out.template ref(), col); + + out.template ref() = col; + out.template ref() = col; + out.template ref() = col; + } + + template + static void blendCorner(uint32_t col, OutputMatrix& out) + { + //model a round corner + alphaGrad<97, 100>(out.template ref<5, 5>(), col); //exact: 0.9711013910 + alphaGrad<42, 100>(out.template ref<4, 5>(), col); //0.4236372243 + alphaGrad<42, 100>(out.template ref<5, 4>(), col); //0.4236372243 + alphaGrad< 6, 100>(out.template ref<5, 3>(), col); //0.05652034508 + alphaGrad< 6, 100>(out.template ref<3, 5>(), col); //0.05652034508 + } +}; + +//------------------------------------------------------------------------------------ + +struct ColorDistanceRGB +{ + static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight) + { + return distYCbCrBuffered(pix1, pix2); + + //if (pix1 == pix2) //about 4% perf boost + // return 0; + //return distYCbCr(pix1, pix2, luminanceWeight); + } +}; + +struct ColorDistanceARGB +{ + static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight) + { + const double a1 = getAlpha(pix1) / 255.0 ; + const double a2 = getAlpha(pix2) / 255.0 ; + /* + Requirements for a color distance handling alpha channel: with a1, a2 in [0, 1] + + 1. if a1 = a2, distance should be: a1 * distYCbCr() + 2. if a1 = 0, distance should be: a2 * distYCbCr(black, white) = a2 * 255 + 3. if a1 = 1, ??? maybe: 255 * (1 - a2) + a2 * distYCbCr() + */ + + //return std::min(a1, a2) * distYCbCrBuffered(pix1, pix2) + 255 * abs(a1 - a2); + //=> following code is 15% faster: + const double d = distYCbCrBuffered(pix1, pix2); + if (a1 < a2) + return a1 * d + 255 * (a2 - a1); + else + return a2 * d + 255 * (a1 - a2); + + //alternative? return std::sqrt(a1 * a2 * square(distYCbCrBuffered(pix1, pix2)) + square(255 * (a1 - a2))); + } +}; + + +struct ColorDistanceUnbufferedARGB +{ + static double dist(uint32_t pix1, uint32_t pix2, double luminanceWeight) + { + const double a1 = getAlpha(pix1) / 255.0 ; + const double a2 = getAlpha(pix2) / 255.0 ; + + const double d = distYCbCr(pix1, pix2, luminanceWeight); + if (a1 < a2) + return a1 * d + 255 * (a2 - a1); + else + return a2 * d + 255 * (a1 - a2); + } +}; + + +struct ColorGradientRGB +{ + template + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) + { + pixBack = gradientRGB(pixFront, pixBack); + } +}; + +struct ColorGradientARGB +{ + template + static void alphaGrad(uint32_t& pixBack, uint32_t pixFront) + { + pixBack = gradientARGB(pixFront, pixBack); + } +}; +} + + +void xbrz::scale(size_t factor, const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, ColorFormat colFmt, const xbrz::ScalerCfg& cfg, int yFirst, int yLast) +{ + if (factor == 1) + { + std::copy(src + yFirst * srcWidth, src + yLast * srcWidth, trg); + return; + } + + static_assert(SCALE_FACTOR_MAX == 6); + switch (colFmt) + { + case ColorFormat::RGB: + switch (factor) + { + case 2: + return scaleImage, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 3: + return scaleImage, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 4: + return scaleImage, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 5: + return scaleImage, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 6: + return scaleImage, ColorDistanceRGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + } + break; + + case ColorFormat::ARGB: + switch (factor) + { + case 2: + return scaleImage, ColorDistanceARGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 3: + return scaleImage, ColorDistanceARGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 4: + return scaleImage, ColorDistanceARGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 5: + return scaleImage, ColorDistanceARGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 6: + return scaleImage, ColorDistanceARGB, OobReaderDuplicate>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + } + break; + + case ColorFormat::ARGB_UNBUFFERED: + switch (factor) + { + case 2: + return scaleImage, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 3: + return scaleImage, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 4: + return scaleImage, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 5: + return scaleImage, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + case 6: + return scaleImage, ColorDistanceUnbufferedARGB, OobReaderTransparent>(src, trg, srcWidth, srcHeight, cfg, yFirst, yLast); + } + break; + } + assert(false); +} + + +bool xbrz::equalColorTest(uint32_t col1, uint32_t col2, ColorFormat colFmt, double luminanceWeight, double equalColorTolerance) +{ + switch (colFmt) + { + case ColorFormat::RGB: + return ColorDistanceRGB::dist(col1, col2, luminanceWeight) < equalColorTolerance; + case ColorFormat::ARGB: + return ColorDistanceARGB::dist(col1, col2, luminanceWeight) < equalColorTolerance; + case ColorFormat::ARGB_UNBUFFERED: + return ColorDistanceUnbufferedARGB::dist(col1, col2, luminanceWeight) < equalColorTolerance; + } + assert(false); + return false; +} + + +void xbrz::bilinearScale(const uint32_t* src, int srcWidth, int srcHeight, + /**/ uint32_t* trg, int trgWidth, int trgHeight) +{ + bilinearScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t), + trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t), + 0, trgHeight, [](uint32_t pix) { return pix; }); +} + + +void xbrz::nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight, + /**/ uint32_t* trg, int trgWidth, int trgHeight) +{ + nearestNeighborScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t), + trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t), + 0, trgHeight, [](uint32_t pix) { return pix; }); +} + + +#if 0 +//#include +void bilinearScaleCpu(const uint32_t* src, int srcWidth, int srcHeight, + /**/ uint32_t* trg, int trgWidth, int trgHeight) +{ + const int TASK_GRANULARITY = 16; + + concurrency::task_group tg; + + for (int i = 0; i < trgHeight; i += TASK_GRANULARITY) + tg.run([=] + { + const int iLast = std::min(i + TASK_GRANULARITY, trgHeight); + xbrz::bilinearScale(src, srcWidth, srcHeight, srcWidth * sizeof(uint32_t), + trg, trgWidth, trgHeight, trgWidth * sizeof(uint32_t), + i, iLast, [](uint32_t pix) { return pix; }); + }); + tg.wait(); +} + + +//Perf: AMP vs CPU: merely ~10% shorter runtime (scaling 1280x800 -> 1920x1080) +//#include +void bilinearScaleAmp(const uint32_t* src, int srcWidth, int srcHeight, //throw concurrency::runtime_exception + /**/ uint32_t* trg, int trgWidth, int trgHeight) +{ + //C++ AMP reference: https://msdn.microsoft.com/en-us/library/hh289390.aspx + //introduction to C++ AMP: https://msdn.microsoft.com/en-us/magazine/hh882446.aspx + using namespace concurrency; + //TODO: pitch + + if (srcHeight <= 0 || srcWidth <= 0) return; + + const float scaleX = static_cast(trgWidth ) / srcWidth; + const float scaleY = static_cast(trgHeight) / srcHeight; + + array_view srcView(srcHeight, srcWidth, src); + array_view< uint32_t, 2> trgView(trgHeight, trgWidth, trg); + trgView.discard_data(); + + parallel_for_each(trgView.extent, [=](index<2> idx) restrict(amp) //throw ? + { + const int y = idx[0]; + const int x = idx[1]; + //Perf notes: + // -> float-based calculation is (almost) 2x as fas as double! + // -> no noticeable improvement via tiling: https://msdn.microsoft.com/en-us/magazine/hh882447.aspx + // -> no noticeable improvement with restrict(amp,cpu) + // -> iterating over y-axis only is significantly slower! + // -> pre-calculating x,y-dependent variables in a buffer + array_view<> is ~ 20 % slower! + const int y1 = srcHeight * y / trgHeight; + int y2 = y1 + 1; + if (y2 == srcHeight) --y2; + + const float yy1 = y / scaleY - y1; + const float y2y = 1 - yy1; + //------------------------------------- + const int x1 = srcWidth * x / trgWidth; + int x2 = x1 + 1; + if (x2 == srcWidth) --x2; + + const float xx1 = x / scaleX - x1; + const float x2x = 1 - xx1; + //------------------------------------- + const float x2xy2y = x2x * y2y; + const float xx1y2y = xx1 * y2y; + const float x2xyy1 = x2x * yy1; + const float xx1yy1 = xx1 * yy1; + + auto interpolate = [=](int offset) + { + /* + https://en.wikipedia.org/wiki/Bilinear_interpolation + (c11(x2 - x) + c21(x - x1)) * (y2 - y ) + + (c12(x2 - x) + c22(x - x1)) * (y - y1) + */ + const auto c11 = (srcView(y1, x1) >> (8 * offset)) & 0xff; + const auto c21 = (srcView(y1, x2) >> (8 * offset)) & 0xff; + const auto c12 = (srcView(y2, x1) >> (8 * offset)) & 0xff; + const auto c22 = (srcView(y2, x2) >> (8 * offset)) & 0xff; + + return c11 * x2xy2y + c21 * xx1y2y + + c12 * x2xyy1 + c22 * xx1yy1; + }; + + const float bi = interpolate(0); + const float gi = interpolate(1); + const float ri = interpolate(2); + const float ai = interpolate(3); + + const auto b = static_cast(bi + 0.5f); + const auto g = static_cast(gi + 0.5f); + const auto r = static_cast(ri + 0.5f); + const auto a = static_cast(ai + 0.5f); + + trgView(y, x) = (a << 24) | (r << 16) | (g << 8) | b; + }); + trgView.synchronize(); //throw ? +} +#endif \ No newline at end of file diff --git a/src/xbrz/xbrz.h b/src/xbrz/xbrz.h new file mode 100644 index 00000000..492fb43a --- /dev/null +++ b/src/xbrz/xbrz.h @@ -0,0 +1,79 @@ +// **************************************************************************** +// * This file is part of the xBRZ project. It is distributed under * +// * GNU General Public License: https://www.gnu.org/licenses/gpl-3.0 * +// * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved * +// * * +// * Additionally and as a special exception, the author gives permission * +// * to link the code of this program with the following libraries * +// * (or with modified versions that use the same licenses), and distribute * +// * linked combinations including the two: MAME, FreeFileSync, Snes9x, ePSXe * +// * You must obey the GNU General Public License in all respects for all of * +// * the code used other than MAME, FreeFileSync, Snes9x, ePSXe. * +// * If you modify this file, you may extend this exception to your version * +// * of the file, but you are not obligated to do so. If you do not wish to * +// * do so, delete this exception statement from your version. * +// **************************************************************************** + +#ifndef XBRZ_HEADER_3847894708239054 +#define XBRZ_HEADER_3847894708239054 + +#include //size_t +#include //uint32_t +#include +#include "xbrz_config.h" + + +namespace xbrz +{ +/* +------------------------------------------------------------------------- +| xBRZ: "Scale by rules" - high quality image upscaling filter by Zenju | +------------------------------------------------------------------------- +using a modified approach of xBR: +http://board.byuu.org/viewtopic.php?f=10&t=2248 +- new rule set preserving small image features +- highly optimized for performance +- support alpha channel +- support multithreading +- support 64-bit architectures +- support processing image slices +- support scaling up to 6xBRZ +*/ + +enum class ColorFormat //from high bits -> low bits, 8 bit per channel +{ + RGB, //8 bit for each red, green, blue, upper 8 bits unused + ARGB, //including alpha channel, BGRA byte order on little-endian machines + ARGB_UNBUFFERED, //like ARGB, but without the one-time buffer creation overhead (ca. 100 - 300 ms) at the expense of a slightly slower scaling time +}; + +const int SCALE_FACTOR_MAX = 6; + +/* +-> map source (srcWidth * srcHeight) to target (scale * width x scale * height) image, optionally processing a half-open slice of rows [yFirst, yLast) only +-> if your emulator changes only a few image slices during each cycle (e.g. DOSBox) then there's no need to run xBRZ on the complete image: + Just make sure you enlarge the source image slice by 2 rows on top and 2 on bottom (this is the additional range the xBRZ algorithm is using during analysis) + CAVEAT: If there are multiple changed slices, make sure they do not overlap after adding these additional rows in order to avoid a memory race condition + in the target image data if you are using multiple threads for processing each enlarged slice! + +THREAD-SAFETY: - parts of the same image may be scaled by multiple threads as long as the [yFirst, yLast) ranges do not overlap! + - there is a minor inefficiency for the first row of a slice, so avoid processing single rows only; suggestion: process at least 8-16 rows +*/ +void scale(size_t factor, //valid range: 2 - SCALE_FACTOR_MAX + const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, + ColorFormat colFmt, + const ScalerCfg& cfg = ScalerCfg(), + int yFirst = 0, int yLast = std::numeric_limits::max()); //slice of source image + +void bilinearScale(const uint32_t* src, int srcWidth, int srcHeight, + /**/ uint32_t* trg, int trgWidth, int trgHeight); + +void nearestNeighborScale(const uint32_t* src, int srcWidth, int srcHeight, + /**/ uint32_t* trg, int trgWidth, int trgHeight); + + +//parameter tuning +bool equalColorTest(uint32_t col1, uint32_t col2, ColorFormat colFmt, double luminanceWeight, double equalColorTolerance); +} + +#endif diff --git a/src/xbrz/xbrz_config.h b/src/xbrz/xbrz_config.h new file mode 100644 index 00000000..fcfda99a --- /dev/null +++ b/src/xbrz/xbrz_config.h @@ -0,0 +1,35 @@ +// **************************************************************************** +// * This file is part of the xBRZ project. It is distributed under * +// * GNU General Public License: https://www.gnu.org/licenses/gpl-3.0 * +// * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved * +// * * +// * Additionally and as a special exception, the author gives permission * +// * to link the code of this program with the following libraries * +// * (or with modified versions that use the same licenses), and distribute * +// * linked combinations including the two: MAME, FreeFileSync, Snes9x, ePSXe * +// * You must obey the GNU General Public License in all respects for all of * +// * the code used other than MAME, FreeFileSync, Snes9x, ePSXe. * +// * If you modify this file, you may extend this exception to your version * +// * of the file, but you are not obligated to do so. If you do not wish to * +// * do so, delete this exception statement from your version. * +// **************************************************************************** + +#ifndef XBRZ_CONFIG_HEADER_284578425345 +#define XBRZ_CONFIG_HEADER_284578425345 + +//do NOT include any headers here! used by xBRZ_dll!!! + +namespace xbrz +{ +struct ScalerCfg +{ + double luminanceWeight = 1; + double equalColorTolerance = 30; + double centerDirectionBias = 4; + double dominantDirectionThreshold = 3.6; + double steepDirectionThreshold = 2.2; + double newTestAttribute = 0; //unused; test new parameters +}; +} + +#endif diff --git a/src/xbrz/xbrz_tools.h b/src/xbrz/xbrz_tools.h new file mode 100644 index 00000000..b8bb8aa0 --- /dev/null +++ b/src/xbrz/xbrz_tools.h @@ -0,0 +1,266 @@ +// **************************************************************************** +// * This file is part of the xBRZ project. It is distributed under * +// * GNU General Public License: https://www.gnu.org/licenses/gpl-3.0 * +// * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved * +// * * +// * Additionally and as a special exception, the author gives permission * +// * to link the code of this program with the following libraries * +// * (or with modified versions that use the same licenses), and distribute * +// * linked combinations including the two: MAME, FreeFileSync, Snes9x, ePSXe * +// * You must obey the GNU General Public License in all respects for all of * +// * the code used other than MAME, FreeFileSync, Snes9x, ePSXe. * +// * If you modify this file, you may extend this exception to your version * +// * of the file, but you are not obligated to do so. If you do not wish to * +// * do so, delete this exception statement from your version. * +// **************************************************************************** + +#ifndef XBRZ_TOOLS_H_825480175091875 +#define XBRZ_TOOLS_H_825480175091875 + +#include +#include +#include + + +namespace xbrz +{ +template inline +unsigned char getByte(uint32_t val) { return static_cast((val >> (8 * N)) & 0xff); } + +inline unsigned char getAlpha(uint32_t pix) { return getByte<3>(pix); } +inline unsigned char getRed (uint32_t pix) { return getByte<2>(pix); } +inline unsigned char getGreen(uint32_t pix) { return getByte<1>(pix); } +inline unsigned char getBlue (uint32_t pix) { return getByte<0>(pix); } + +inline uint32_t makePixel(unsigned char a, unsigned char r, unsigned char g, unsigned char b) { return (a << 24) | (r << 16) | (g << 8) | b; } +inline uint32_t makePixel( unsigned char r, unsigned char g, unsigned char b) { return (r << 16) | (g << 8) | b; } + +inline uint32_t rgb555to888(uint16_t pix) { return ((pix & 0x7C00) << 9) | ((pix & 0x03E0) << 6) | ((pix & 0x001F) << 3); } +inline uint32_t rgb565to888(uint16_t pix) { return ((pix & 0xF800) << 8) | ((pix & 0x07E0) << 5) | ((pix & 0x001F) << 3); } + +inline uint16_t rgb888to555(uint32_t pix) { return static_cast(((pix & 0xF80000) >> 9) | ((pix & 0x00F800) >> 6) | ((pix & 0x0000F8) >> 3)); } +inline uint16_t rgb888to565(uint32_t pix) { return static_cast(((pix & 0xF80000) >> 8) | ((pix & 0x00FC00) >> 5) | ((pix & 0x0000F8) >> 3)); } + + +template inline +Pix* byteAdvance(Pix* ptr, int bytes) +{ + using PixNonConst = typename std::remove_cv::type; + using PixByte = typename std::conditional::value, char, const char>::type; + + static_assert(std::is_integral::value, "Pix* is expected to be cast-able to char*"); + + return reinterpret_cast(reinterpret_cast(ptr) + bytes); +} + + +//fill block with the given color +template inline +void fillBlock(Pix* trg, int pitch /*[bytes]*/, Pix col, int blockWidth, int blockHeight) +{ + //for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch)) + // std::fill(trg, trg + blockWidth, col); + + for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch)) + for (int x = 0; x < blockWidth; ++x) + trg[x] = col; +} + + +//nearest-neighbor (going over target image - slow for upscaling, since source is read multiple times missing out on cache! Fast for similar image sizes!) +template +void nearestNeighborScale(const PixSrc* src, int srcWidth, int srcHeight, int srcPitch /*[bytes]*/, + /**/ PixTrg* trg, int trgWidth, int trgHeight, int trgPitch /*[bytes]*/, + int yFirst, int yLast, PixConverter pixCvrt /*convert PixSrc to PixTrg*/) +{ + static_assert(std::is_integral::value, "PixSrc* is expected to be cast-able to char*"); + static_assert(std::is_integral::value, "PixTrg* is expected to be cast-able to char*"); + + static_assert(std::is_same::value, "PixConverter returning wrong pixel format"); + + if (srcPitch < srcWidth * static_cast(sizeof(PixSrc)) || + trgPitch < trgWidth * static_cast(sizeof(PixTrg))) + { + assert(false); + return; + } + + yFirst = std::max(yFirst, 0); + yLast = std::min(yLast, trgHeight); + if (yFirst >= yLast || srcHeight <= 0 || srcWidth <= 0) return; + + for (int y = yFirst; y < yLast; ++y) + { + const int ySrc = srcHeight * y / trgHeight; + const PixSrc* const srcLine = byteAdvance(src, ySrc * srcPitch); + PixTrg* const trgLine = byteAdvance(trg, y * trgPitch); + + for (int x = 0; x < trgWidth; ++x) + { + const int xSrc = srcWidth * x / trgWidth; + trgLine[x] = pixCvrt(srcLine[xSrc]); + } + } +} + + +//nearest-neighbor (going over source image - fast for upscaling, since source is read only once +template +void nearestNeighborScaleOverSource(const PixSrc* src, int srcWidth, int srcHeight, int srcPitch /*[bytes]*/, + /**/ PixTrg* trg, int trgWidth, int trgHeight, int trgPitch /*[bytes]*/, + int yFirst, int yLast, PixConverter pixCvrt /*convert PixSrc to PixTrg*/) +{ + static_assert(std::is_integral::value, "PixSrc* is expected to be cast-able to char*"); + static_assert(std::is_integral::value, "PixTrg* is expected to be cast-able to char*"); + + static_assert(std::is_same::value, "PixConverter returning wrong pixel format"); + + if (srcPitch < srcWidth * static_cast(sizeof(PixSrc)) || + trgPitch < trgWidth * static_cast(sizeof(PixTrg))) + { + assert(false); + return; + } + + yFirst = std::max(yFirst, 0); + yLast = std::min(yLast, srcHeight); + if (yFirst >= yLast || trgWidth <= 0 || trgHeight <= 0) return; + + for (int y = yFirst; y < yLast; ++y) + { + //mathematically: ySrc = floor(srcHeight * yTrg / trgHeight) + // => search for integers in: [ySrc, ySrc + 1) * trgHeight / srcHeight + + //keep within for loop to support MT input slices! + const int yTrgFirst = ( y * trgHeight + srcHeight - 1) / srcHeight; //=ceil(y * trgHeight / srcHeight) + const int yTrgLast = ((y + 1) * trgHeight + srcHeight - 1) / srcHeight; //=ceil(((y + 1) * trgHeight) / srcHeight) + const int blockHeight = yTrgLast - yTrgFirst; + + if (blockHeight > 0) + { + const PixSrc* srcLine = byteAdvance(src, y * srcPitch); + /**/ PixTrg* trgLine = byteAdvance(trg, yTrgFirst * trgPitch); + int xTrgFirst = 0; + + for (int x = 0; x < srcWidth; ++x) + { + const int xTrgLast = ((x + 1) * trgWidth + srcWidth - 1) / srcWidth; + const int blockWidth = xTrgLast - xTrgFirst; + if (blockWidth > 0) + { + xTrgFirst = xTrgLast; + + const auto trgPix = pixCvrt(srcLine[x]); + fillBlock(trgLine, trgPitch, trgPix, blockWidth, blockHeight); + trgLine += blockWidth; + } + } + } + } +} + + +template +void bilinearScale(const uint32_t* src, int srcWidth, int srcHeight, int srcPitch, + /**/ PixTrg* trg, int trgWidth, int trgHeight, int trgPitch, + int yFirst, int yLast, PixConverter pixCvrt /*convert uint32_t to PixTrg*/) +{ + static_assert(std::is_integral::value, "PixTrg* is expected to be cast-able to char*"); + static_assert(std::is_same::value, "PixConverter returning wrong pixel format"); + + if (srcPitch < srcWidth * static_cast(sizeof(uint32_t)) || + trgPitch < trgWidth * static_cast(sizeof(PixTrg))) + { + assert(false); + return; + } + + yFirst = std::max(yFirst, 0); + yLast = std::min(yLast, trgHeight); + if (yFirst >= yLast || srcHeight <= 0 || srcWidth <= 0) return; + + const double scaleX = static_cast(trgWidth ) / srcWidth; + const double scaleY = static_cast(trgHeight) / srcHeight; + + //perf notes: + // -> double-based calculation is (slightly) faster than float + // -> pre-calculation gives significant boost; std::vector<> memory allocation is negligible! + struct CoeffsX + { + int x1 = 0; + int x2 = 0; + double xx1 = 0; + double x2x = 0; + }; + std::vector buf(trgWidth); + for (int x = 0; x < trgWidth; ++x) + { + const int x1 = srcWidth * x / trgWidth; + int x2 = x1 + 1; + if (x2 == srcWidth) --x2; + + const double xx1 = x / scaleX - x1; + const double x2x = 1 - xx1; + + buf[x] = { x1, x2, xx1, x2x }; + } + + for (int y = yFirst; y < yLast; ++y) + { + const int y1 = srcHeight * y / trgHeight; + int y2 = y1 + 1; + if (y2 == srcHeight) --y2; + + const double yy1 = y / scaleY - y1; + const double y2y = 1 - yy1; + + const uint32_t* const srcLine = byteAdvance(src, y1 * srcPitch); + const uint32_t* const srcLineNext = byteAdvance(src, y2 * srcPitch); + PixTrg* const trgLine = byteAdvance(trg, y * trgPitch); + + for (int x = 0; x < trgWidth; ++x) + { + //perf: do NOT "simplify" the variable layout without measurement! + const int x1 = buf[x].x1; + const int x2 = buf[x].x2; + const double xx1 = buf[x].xx1; + const double x2x = buf[x].x2x; + + const double x2xy2y = x2x * y2y; + const double xx1y2y = xx1 * y2y; + const double x2xyy1 = x2x * yy1; + const double xx1yy1 = xx1 * yy1; + + auto interpolate = [=](int offset) + { + /* https://en.wikipedia.org/wiki/Bilinear_interpolation + (c11(x2 - x) + c21(x - x1)) * (y2 - y ) + + (c12(x2 - x) + c22(x - x1)) * (y - y1) */ + const auto c11 = (srcLine [x1] >> (8 * offset)) & 0xff; + const auto c21 = (srcLine [x2] >> (8 * offset)) & 0xff; + const auto c12 = (srcLineNext[x1] >> (8 * offset)) & 0xff; + const auto c22 = (srcLineNext[x2] >> (8 * offset)) & 0xff; + + return c11 * x2xy2y + c21 * xx1y2y + + c12 * x2xyy1 + c22 * xx1yy1; + }; + + const double bi = interpolate(0); + const double gi = interpolate(1); + const double ri = interpolate(2); + const double ai = interpolate(3); + + const auto b = static_cast(bi + 0.5); + const auto g = static_cast(gi + 0.5); + const auto r = static_cast(ri + 0.5); + const auto a = static_cast(ai + 0.5); + + const uint32_t trgPix = (a << 24) | (r << 16) | (g << 8) | b; + + trgLine[x] = pixCvrt(trgPix); + } + } +} +} + +#endif //XBRZ_TOOLS_H_825480175091875