Optimize streaming texture updates

GL/private.h

@@ -379,6 +379,8 @@ GLenum _glGetCullFace();
 GLenum _glGetFrontFace();
 GLenum _glGetGpuBlendSrcFactor();
 GLenum _glGetGpuBlendDstFactor();
+GLint _glGetUnpackRowLength();
+GLint _glGetUnpackAlignment();
 
 extern PolyList OP_LIST;
 extern PolyList PT_LIST;

GL/state.c

@@ -66,6 +66,9 @@ static struct {
 
     GLfloat half_point_size;
     GLfloat half_line_width;
+
+    GLint unpack_row_length;
+    GLint unpack_alignment;
 } GPUState = {
     .is_dirty = GL_TRUE,
     .depth_func = GL_LESS,
@@ -102,6 +105,8 @@ static struct {
     .current_tex_coord1 = {0.0f, 0.0f},
     .half_point_size = 0.5f,
     .half_line_width = 0.5f,
+    .unpack_row_length = 0,
+    .unpack_alignment = 4,
 };
 
 float* _glCurrentColor() {
@@ -141,6 +146,14 @@ GLboolean _glGPUStateIsDirty() {
     return GPUState.is_dirty;
 }
 
+GLint _glGetUnpackRowLength() {
+    return GPUState.unpack_row_length;
+}
+
+GLint _glGetUnpackAlignment() {
+    return GPUState.unpack_alignment;
+}
+
 Material* _glActiveMaterial() {
     return &GPUState.material;
 }
@@ -873,8 +886,25 @@ void glColorMask(GLboolean red, GLboolean green, GLboolean blue, GLboolean alpha
 }
 
 void glPixelStorei(GLenum pname, GLint param) {
-    _GL_UNUSED(pname);
+    switch(pname) {
-    _GL_UNUSED(param);
+        case GL_UNPACK_ROW_LENGTH:
+            if(param < 0) {
+                _glKosThrowError(GL_INVALID_VALUE, __func__);
+                return;
+            }
+            GPUState.unpack_row_length = param;
+        break;
+        case GL_UNPACK_ALIGNMENT:
+            if(param != 1 && param != 2 && param != 4 && param != 8) {
+                _glKosThrowError(GL_INVALID_VALUE, __func__);
+                return;
+            }
+            GPUState.unpack_alignment = param;
+        break;
+        default:
+            _glKosThrowError(GL_INVALID_ENUM, __func__);
+        break;
+    }
 }
 
 
@@ -1118,6 +1148,12 @@ void APIENTRY glGetIntegerv(GLenum pname, GLint *params) {
         case GL_TEXTURE_INTERNAL_FORMAT_KOS:
             *params = _glGetTextureInternalFormat();
         break;
+        case GL_UNPACK_ROW_LENGTH:
+            *params = GPUState.unpack_row_length;
+        break;
+        case GL_UNPACK_ALIGNMENT:
+            *params = GPUState.unpack_alignment;
+        break;
 
     default:
         _glKosThrowError(GL_INVALID_ENUM, __func__);

GL/texture.c

@@ -1625,6 +1625,15 @@ static inline GLboolean is4BPPFormat(GLenum format) {
     return format == GL_COLOR_INDEX4_EXT || format == GL_COLOR_INDEX4_TWID_KOS;
 }
 
+static GLuint _glGetUnpackRowPitch(GLsizei width, GLint sourceStride, GLenum format) {
+    GLint rowLength = _glGetUnpackRowLength();
+    GLint unpackAlignment = _glGetUnpackAlignment();
+    GLuint rowPixels = (rowLength > 0) ? (GLuint) rowLength : (GLuint) width;
+    GLuint rowBytes = is4BPPFormat(format) ? ((rowPixels + 1) / 2) : (rowPixels * (GLuint) sourceStride);
+
+    return (rowBytes + (GLuint) unpackAlignment - 1) & ~((GLuint) unpackAlignment - 1);
+}
+
 void APIENTRY glTexImage2D(GLenum target, GLint level, GLint internalFormat,
                            GLsizei width, GLsizei height, GLint border,
                            GLenum format, GLenum type, const GLvoid *data) {
@@ -1745,6 +1754,12 @@ void APIENTRY glTexImage2D(GLenum target, GLint level, GLint internalFormat,
     GLubyte* targetData = (active->baseDataOffset == 0) ? active->data : _glGetMipmapLocation(active, level);
     gl_assert(targetData);
 
+    if(!data) {
+        gl_assert(active->index == originalId);
+        _glGPUStateMarkDirty();
+        return;
+    }
+
     if(needs_conversion < 0) {
         _glKosThrowError(GL_INVALID_VALUE, __func__);
         INFO_MSG("Couldn't find necessary texture conversion\n");
@@ -2164,18 +2179,19 @@ void APIENTRY glTexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint y
         return;
     }
 
-    GLboolean isPaletted = (
-        active->internalFormat == GL_COLOR_INDEX8_EXT ||
-        active->internalFormat == GL_COLOR_INDEX4_EXT ||
-        active->internalFormat == GL_COLOR_INDEX4_TWID_KOS ||
-        active->internalFormat == GL_COLOR_INDEX8_TWID_KOS
-    ) ? GL_TRUE : GL_FALSE;
-
     GLenum cleanInternalFormat = _cleanInternalFormat(active->internalFormat);
 
     // Determine source stride
     GLint sourceStride = _determineStride(format, type);
-    GLuint srcBytes = (width * height * sourceStride);
+    if (sourceStride < 0) {
+        INFO_MSG("Stride was not detected\n");
+        _glKosThrowError(GL_INVALID_OPERATION, __func__);
+        return;
+    }
+
+    GLuint sourceRowWidth = is4BPPFormat(format) ? (((GLuint) width + 1) / 2) : ((GLuint) width * (GLuint) sourceStride);
+    GLuint sourcePitch = _glGetUnpackRowPitch(width, sourceStride, format);
+    GLuint srcBytes = height ? (sourcePitch * ((GLuint) height - 1) + sourceRowWidth) : 0;
 
     // Calculate destination stride (this accounts for both POT and NPOT)
     GLint destStride = _determineStrideInternal(cleanInternalFormat);
@@ -2220,6 +2236,7 @@ void APIENTRY glTexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint y
         }
         if (needs_conversion == CONVERSION_TYPE_CONVERT) {
             for (uint32_t y = 0; y < height; ++y) {
+                src = (const GLubyte*) data + (y * sourcePitch);
                 dst = conversionBuffer + ((y + yoffset) * textureWidth + xoffset) * destStride;
                 for (uint32_t x = 0; x < width; ++x) {
                     conversion(src, dst);
@@ -2233,21 +2250,24 @@ void APIENTRY glTexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint y
 
             for (uint32_t y = yoffset; y < yoffset + height; ++y) {
                 for (uint32_t x = xoffset; x < xoffset + width; ++x) {
-                    uint32_t srcIndex = (y - yoffset) * width + (x - xoffset);
                     uint32_t newLocation = twid_compute_index(x, y, maskX, maskY);
                     dst = conversionBuffer + (destStride * newLocation);
+                    src = (const GLubyte*) data + ((y - yoffset) * sourcePitch) + ((x - xoffset) * sourceStride);
 
                     if (needs_conversion == 3) {
-                        conversion(src + srcIndex * sourceStride, dst);
+                        conversion(src, dst);
                     } else {
-                        memcpy(dst, src + srcIndex * sourceStride, destStride);
+                        memcpy(dst, src, destStride);
                     }
                 }
             }
         }
 
         if (pack) {
-            assert(isPaletted);
+            assert(active->internalFormat == GL_COLOR_INDEX8_EXT ||
+                   active->internalFormat == GL_COLOR_INDEX4_EXT ||
+                   active->internalFormat == GL_COLOR_INDEX4_TWID_KOS ||
+                   active->internalFormat == GL_COLOR_INDEX8_TWID_KOS);
             size_t dst_byte = 0;
             for (size_t src_byte = 0; src_byte < destBytes; ++src_byte) {
                 uint8_t v = conversionBuffer[src_byte];
@@ -2267,10 +2287,19 @@ void APIENTRY glTexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint y
         free(conversionBuffer);
     } else {
         // No conversion necessary, we can update data directly
+        if (xoffset == 0 &&
+            yoffset == 0 &&
+            width == textureWidth &&
+            sourceStride == destStride &&
+            sourcePitch == textureWidth * destStride) {
+            FASTCPY(targetData, data, height * sourcePitch);
+            _glGPUStateMarkDirty();
+            return;
+        }
+
         for (GLsizei y = 0; y < height; ++y) {
-            GLsizei srcRowWidth = width * sourceStride;
             GLubyte* destRow = targetData + ((y + yoffset) * textureWidth + xoffset) * destStride;
-            FASTCPY(destRow, (GLubyte*)data + y * srcRowWidth, srcRowWidth);
+            FASTCPY(destRow, (GLubyte*)data + y * sourcePitch, sourceRowWidth);
         }
     }