This patch adds the ability to decompress DXTn textures, and is meant to be used for DXTn 3D textures because they are not supported properly in drivers.
Signed-off-by: Connor McAdams conmanx360@gmail.com --- dlls/wined3d/texture.c | 51 +++++++- dlls/wined3d/utils.c | 275 +++++++++++++++++++++++++++++++++++++---- dlls/wined3d/wined3d_private.h | 4 + 3 files changed, 303 insertions(+), 27 deletions(-)
diff --git a/dlls/wined3d/texture.c b/dlls/wined3d/texture.c index c316906..34e9e56 100644 --- a/dlls/wined3d/texture.c +++ b/dlls/wined3d/texture.c @@ -567,6 +567,8 @@ static void wined3d_texture_allocate_gl_mutable_storage(struct wined3d_texture * unsigned int level, level_count, layer, layer_count; GLsizei width, height, depth; GLenum target; + GLint gl_format; + GLint gl_type;
level_count = texture->level_count; if (texture->target == GL_TEXTURE_1D_ARRAY || texture->target == GL_TEXTURE_2D_ARRAY) @@ -578,6 +580,18 @@ static void wined3d_texture_allocate_gl_mutable_storage(struct wined3d_texture * { target = wined3d_texture_get_sub_resource_target(texture, layer * level_count);
+ if (texture->resource.format_flags & WINED3DFMT_FLAG_DECOMPRESS) + { + gl_internal_format = GL_RGBA8; + gl_format = GL_BGRA; + gl_type = GL_UNSIGNED_INT_8_8_8_8_REV; + } + else + { + gl_format = format->glFormat; + gl_type = format->glType; + } + for (level = 0; level < level_count; ++level) { width = wined3d_texture_get_level_pow2_width(texture, level); @@ -596,19 +610,19 @@ static void wined3d_texture_allocate_gl_mutable_storage(struct wined3d_texture * depth = wined3d_texture_get_level_depth(texture, level); GL_EXTCALL(glTexImage3D(target, level, gl_internal_format, width, height, target == GL_TEXTURE_2D_ARRAY ? texture->layer_count : depth, 0, - format->glFormat, format->glType, NULL)); + gl_format, gl_type, NULL)); checkGLcall("glTexImage3D"); } else if (target == GL_TEXTURE_1D) { gl_info->gl_ops.gl.p_glTexImage1D(target, level, gl_internal_format, - width, 0, format->glFormat, format->glType, NULL); + width, 0, gl_format, gl_type, NULL); } else { gl_info->gl_ops.gl.p_glTexImage2D(target, level, gl_internal_format, width, target == GL_TEXTURE_1D_ARRAY ? texture->layer_count : height, 0, - format->glFormat, format->glType, NULL); + gl_format, gl_type, NULL); checkGLcall("glTexImage2D"); } } @@ -624,6 +638,9 @@ static void wined3d_texture_allocate_gl_immutable_storage(struct wined3d_texture GLsizei height = wined3d_texture_get_level_pow2_height(texture, 0); GLsizei width = wined3d_texture_get_level_pow2_width(texture, 0);
+ if (texture->resource.format_flags & WINED3DFMT_FLAG_DECOMPRESS) + gl_internal_format = GL_RGBA8; + switch (texture->target) { case GL_TEXTURE_3D: @@ -1948,6 +1965,34 @@ void wined3d_texture_upload_data(struct wined3d_texture *texture, unsigned int s bo.addr += src_box->left * format->byte_count; }
+ if (texture->resource.format_flags & WINED3DFMT_FLAG_DECOMPRESS) + { + unsigned int dst_row_pitch, dst_slice_pitch; + void *src_mem; + + f = *format; + format = wined3d_get_format(gl_info, WINED3DFMT_B8G8R8A8_UNORM, WINED3DUSAGE_TEXTURE); + + wined3d_format_calculate_pitch(format, 1, update_w, update_h, &dst_row_pitch, &dst_slice_pitch); + + if (!(converted_mem = heap_calloc(update_d, dst_slice_pitch))) + { + ERR("Failed to allocate upload buffer.\n"); + return; + } + src_mem = context_map_bo_address(context, &bo, src_slice_pitch, + GL_PIXEL_UNPACK_BUFFER, WINED3D_MAP_READ); + f.decompress(src_mem, converted_mem, src_row_pitch, src_slice_pitch, + dst_row_pitch, dst_slice_pitch, update_w, update_h, update_d); + context_unmap_bo_address(context, &bo, GL_PIXEL_UNPACK_BUFFER); + + bo.buffer_object = 0; + bo.addr = converted_mem; + src_row_pitch = dst_row_pitch; + src_slice_pitch = dst_slice_pitch; + texture->resource.format_flags &= ~WINED3DFMT_FLAG_BLOCKS; + } + if (format->upload) { unsigned int dst_row_pitch, dst_slice_pitch; diff --git a/dlls/wined3d/utils.c b/dlls/wined3d/utils.c index 937c1bc..04128d4 100644 --- a/dlls/wined3d/utils.c +++ b/dlls/wined3d/utils.c @@ -339,6 +339,34 @@ static const struct wined3d_format_base_flags format_base_flags[] = {WINED3DFMT_RESZ, WINED3DFMT_FLAG_EXTENSION}, };
+static void convert_dxt1_b8g8r8a8_unorm(const BYTE *src, BYTE *dst, UINT src_row_pitch, UINT src_slice_pitch, + UINT dst_row_pitch, UINT dst_slice_pitch, UINT width, UINT height, UINT depth); +static void convert_dxt3_b8g8r8a8_unorm(const BYTE *src, BYTE *dst, UINT src_row_pitch, UINT src_slice_pitch, + UINT dst_row_pitch, UINT dst_slice_pitch, UINT width, UINT height, UINT depth); +static void convert_dxt5_b8g8r8a8_unorm(const BYTE *src, BYTE *dst, UINT src_row_pitch, UINT src_slice_pitch, + UINT dst_row_pitch, UINT dst_slice_pitch, UINT width, UINT height, UINT depth); + +struct wined3d_format_decompress_info +{ + enum wined3d_format_id id; + DWORD flags; + void (*decompress)(const BYTE *src, BYTE *dst, unsigned int src_row_pitch, unsigned int src_slice_pitch, + unsigned int dst_row_pitch, unsigned int dst_slice_pitch, + unsigned int width, unsigned int height, unsigned int depth); +}; + +static const struct wined3d_format_decompress_info format_decompress_info[] = +{ + {WINED3DFMT_DXT1, WINED3DFMT_FLAG_DECOMPRESS, convert_dxt1_b8g8r8a8_unorm}, + {WINED3DFMT_DXT2, WINED3DFMT_FLAG_DECOMPRESS, convert_dxt3_b8g8r8a8_unorm}, + {WINED3DFMT_DXT3, WINED3DFMT_FLAG_DECOMPRESS, convert_dxt3_b8g8r8a8_unorm}, + {WINED3DFMT_DXT4, WINED3DFMT_FLAG_DECOMPRESS, convert_dxt5_b8g8r8a8_unorm}, + {WINED3DFMT_DXT5, WINED3DFMT_FLAG_DECOMPRESS, convert_dxt5_b8g8r8a8_unorm}, + {WINED3DFMT_BC1_UNORM, WINED3DFMT_FLAG_DECOMPRESS, convert_dxt1_b8g8r8a8_unorm}, + {WINED3DFMT_BC2_UNORM, WINED3DFMT_FLAG_DECOMPRESS, convert_dxt3_b8g8r8a8_unorm}, + {WINED3DFMT_BC3_UNORM, WINED3DFMT_FLAG_DECOMPRESS, convert_dxt5_b8g8r8a8_unorm}, +}; + struct wined3d_format_block_info { enum wined3d_format_id id; @@ -437,6 +465,9 @@ struct wined3d_format_texture_info void (*download)(const BYTE *src, BYTE *dst, unsigned int src_row_pitch, unsigned int src_slice_pitch, unsigned int dst_row_pitch, unsigned int dst_slice_pitch, unsigned int width, unsigned int height, unsigned int depth); + void (*decompress)(const BYTE *src, BYTE *dst, unsigned int src_row_pitch, unsigned int src_slice_pitch, + unsigned int dst_row_pitch, unsigned int dst_slice_pitch, + unsigned int width, unsigned int height, unsigned int depth); };
static void convert_l4a4_unorm(const BYTE *src, BYTE *dst, UINT src_row_pitch, UINT src_slice_pitch, @@ -840,6 +871,206 @@ static void convert_s8_uint_d24_float(const BYTE *src, BYTE *dst, UINT src_row_p } }
+static void dxtn_decompress_block(const BYTE *src, BYTE *dst, UINT width, UINT height, UINT depth, + UINT x_pos, UINT y_pos, UINT z_pos, UINT dst_slice_pitch, UINT64 cur_block, UINT fmt, BYTE block_offset) +{ + UINT64 alpha_block, alpha_index, color_block, color_index; + const UINT64 *source; + DWORD *dest; + DWORD alpha_lookup; + DWORD bgra; + DWORD temp; + DWORD i, x, y; + WORD color[2]; + BYTE alpha_val, color_val; + BYTE alpha[8]; + BYTE r[4]; + BYTE g[4]; + BYTE b[4]; + + source = (const UINT64 *)(src + cur_block * block_offset); + if (fmt == WINED3DFMT_DXT1) + { + color_block = source[0]; + alpha_block = 0; + } + else + { + alpha_block = source[0]; + color_block = source[1]; + } + + color[0] = color_block & 0xffff; + color[1] = (color_block >> 16) & 0xffff; + for (i = 0; i < 2; ++i) + { + temp = (color[i] >> 11) * 255 + 16; + r[i] = (temp / 32 + temp) / 32; + temp = ((color[i] >> 5) & 0x3f) * 255 + 32; + g[i] = (temp / 64 + temp) / 64; + temp = (color[i] & 0x1f) * 255 + 16; + b[i] = (temp / 32 + temp) / 32; + } + + if (fmt == WINED3DFMT_DXT3 || fmt == WINED3DFMT_DXT5 || color[0] > color[1]) + { + for (i = 0; i < 2; ++i) + { + r[2 + i] = (2 * r[0 + i] + r[1 - i]) / 3; + g[2 + i] = (2 * g[0 + i] + g[1 - i]) / 3; + b[2 + i] = (2 * b[0 + i] + b[1 - i]) / 3; + } + } + else if (fmt == WINED3DFMT_DXT1 && color[0] <= color[1]) + { + r[2] = (r[0] + r[1]) / 2; + g[2] = (g[0] + g[1]) / 2; + b[2] = (b[0] + b[1]) / 2; + + r[3] = 0; + g[3] = 0; + b[3] = 0; + } + + switch (fmt) + { + case WINED3DFMT_DXT1: + alpha_index = 0; + for (i = 0; i < ARRAY_SIZE(alpha); ++i) + alpha[i] = 255; + if (color[0] <= color[1]) + alpha[3] = 0; + break; + case WINED3DFMT_DXT3: + alpha_index = alpha_block; + for (i = 0; i < ARRAY_SIZE(alpha); ++i) + alpha[i] = 0; + break; + case WINED3DFMT_DXT5: + alpha_index = (alpha_block >> 16); + alpha[0] = alpha_block & 0xff; + alpha[1] = (alpha_block >> 8) & 0xff; + if (alpha[0] > alpha[1]) + { + for (i = 0; i < 6; ++i) + alpha[2 + i] = (((6 - i) * alpha[0]) + ((1 + i) * alpha[1])) / 7; + } + else if (alpha[0] <= alpha[1]) + { + for (i = 0; i < 4; ++i) + alpha[2 + i] = (((4 - i) * alpha[0]) + ((1 + i) * alpha[1])) / 5; + alpha[6] = 0; + alpha[7] = 255; + } + break; + default: + alpha_index = 0; + break; + } + + color_index = (color_block >> 32) & 0xffffffff; + dest = (DWORD *)(dst + z_pos * dst_slice_pitch); + for (y = 0; y < 4; ++y) + { + if (y_pos + y >= height) + break; + for (x = 0; x < 4; ++x) + { + if (x_pos + x >= width) + break; + + color_val = 0; + alpha_val = 0; + bgra = 0; + + color_val = (color_index >> (y * 8)); + color_val = (color_val >> (x * 2)) & 0x3; + switch (fmt) + { + case WINED3DFMT_DXT1: + alpha_val = color_val; + break; + case WINED3DFMT_DXT3: + alpha_lookup = (alpha_index >> (y * 16)) & 0xffff; + alpha_val = (alpha_lookup >> (x * 4)) & 0xf; + temp = alpha_val * 255 + 8; + alpha[0] = (temp / 16 + temp) / 16; + alpha_val = 0; + break; + case WINED3DFMT_DXT5: + alpha_lookup = (alpha_index >> (y * 12)) & 0xfff; + alpha_val = (alpha_lookup >> (x * 3)) & 0x7; + break; + } + bgra = ((alpha[alpha_val] << 24) | (r[color_val] << 16) | (g[color_val] << 8) | b[color_val]); + dest[(y_pos + y) * width + (x_pos + x)] = bgra; + } + } +} + +static void convert_dxt5_b8g8r8a8_unorm(const BYTE *src, BYTE *dst, UINT src_row_pitch, UINT src_slice_pitch, + UINT dst_row_pitch, UINT dst_slice_pitch, UINT width, UINT height, UINT depth) +{ + UINT64 current_block; + DWORD x, y, z; + + current_block = 0; + + for (z = 0; z < depth; ++z) + { + for (y = 0; y < height; y += 4) + { + for (x = 0; x < width; x += 4) + { + dxtn_decompress_block(src, dst, width, height, depth, x, y, z, dst_slice_pitch, current_block, WINED3DFMT_DXT5, 16); + current_block++; + } + } + } +} + +static void convert_dxt3_b8g8r8a8_unorm(const BYTE *src, BYTE *dst, UINT src_row_pitch, UINT src_slice_pitch, + UINT dst_row_pitch, UINT dst_slice_pitch, UINT width, UINT height, UINT depth) +{ + UINT64 current_block; + DWORD x, y, z; + + current_block = 0; + + for (z = 0; z < depth; ++z) + { + for (y = 0; y < height; y += 4) + { + for (x = 0; x < width; x += 4) + { + dxtn_decompress_block(src, dst, width, height, depth, x, y, z, dst_slice_pitch, current_block, WINED3DFMT_DXT3, 16); + current_block++; + } + } + } +} + +static void convert_dxt1_b8g8r8a8_unorm(const BYTE *src, BYTE *dst, UINT src_row_pitch, UINT src_slice_pitch, + UINT dst_row_pitch, UINT dst_slice_pitch, UINT width, UINT height, UINT depth) +{ + UINT64 current_block; + DWORD x, y, z; + + current_block = 0; + + for (z = 0; z < depth; ++z) + { + for (y = 0; y < height; y += 4) + { + for (x = 0; x < width; x += 4) + { + dxtn_decompress_block(src, dst, width, height, depth, x, y, z, dst_slice_pitch, current_block, WINED3DFMT_DXT1, 8); + current_block++; + } + } + } +} + static void x8_d24_unorm_upload(const BYTE *src, BYTE *dst, unsigned int src_row_pitch, unsigned int src_slice_pitch, unsigned int dst_row_pitch, unsigned int dst_slice_pitch, @@ -1066,7 +1297,7 @@ static const struct wined3d_format_texture_info format_texture_info[] = /* format id gl_internal gl_srgb_internal gl_rt_internal gl_format gl_type conv_byte_count flags - extension upload download */ + extension upload download decompress */ /* FourCC formats */ /* GL_APPLE_ycbcr_422 claims that its '2YUV' format, which is supported via the UNSIGNED_SHORT_8_8_REV_APPLE type * is equivalent to 'UYVY' format on Windows, and the 'YUVS' via UNSIGNED_SHORT_8_8_APPLE equates to 'YUY2'. The @@ -1897,6 +2128,23 @@ static BOOL init_format_block_info(struct wined3d_gl_info *gl_info) return TRUE; }
+static BOOL init_format_decompress_info(struct wined3d_gl_info *gl_info) +{ + struct wined3d_format *format; + unsigned int i; + + for (i = 0; i < ARRAY_SIZE(format_decompress_info); ++i) + { + if (!(format = get_format_internal(gl_info, format_decompress_info[i].id))) + return FALSE; + + format->flags[WINED3D_GL_RES_TYPE_TEX_3D] |= format_decompress_info[i].flags; + format->decompress = format_decompress_info[i].decompress; + } + + return TRUE; +} + static GLenum wined3d_gl_type_to_enum(enum wined3d_gl_resource_type type) { switch (type) @@ -3394,29 +3642,7 @@ static void apply_format_fixups(struct wined3d_adapter *adapter, struct wined3d_ * for dx9 GPUs support it, some do not, so not supporting DXTn volumes is OK for d3d9. * * Note that GL_NV_texture_compression_vtc adds this functionality to OpenGL, but the - * block layout is not compatible with the one used by d3d. See volume_dxt5_test. */ - idx = get_format_idx(WINED3DFMT_DXT1); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_DXT2); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_DXT3); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_DXT4); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_DXT5); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_BC1_UNORM); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_BC1_UNORM_SRGB); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_BC2_UNORM); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_BC2_UNORM_SRGB); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_BC3_UNORM); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; - idx = get_format_idx(WINED3DFMT_BC3_UNORM_SRGB); - gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; + * block layout is not compatible with the one used by d3d. See volume_dxtn_test. */ /* Similarly with ATI1N / ATI2N and GL_ARB_texture_compression_rgtc. */ idx = get_format_idx(WINED3DFMT_ATI1N); gl_info->formats[idx].flags[WINED3D_GL_RES_TYPE_TEX_3D] &= ~WINED3DFMT_FLAG_TEXTURE; @@ -3742,6 +3968,7 @@ BOOL wined3d_adapter_init_format_info(struct wined3d_adapter *adapter, struct wi
if (!init_format_base_info(gl_info)) return FALSE; if (!init_format_block_info(gl_info)) goto fail; + if (!init_format_decompress_info(gl_info)) goto fail;
if (!ctx) /* WINED3D_NO3D */ return TRUE; diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h index 9aa7081..b1d8b24 100644 --- a/dlls/wined3d/wined3d_private.h +++ b/dlls/wined3d/wined3d_private.h @@ -4247,6 +4247,7 @@ extern enum wined3d_format_id pixelformat_for_depth(DWORD depth) DECLSPEC_HIDDEN #define WINED3DFMT_FLAG_BLOCKS_NO_VERIFY 0x00100000 #define WINED3DFMT_FLAG_INTEGER 0x00200000 #define WINED3DFMT_FLAG_GEN_MIPMAP 0x00400000 +#define WINED3DFMT_FLAG_DECOMPRESS 0x00800000
struct wined3d_rational { @@ -4306,6 +4307,9 @@ struct wined3d_format void (*download)(const BYTE *src, BYTE *dst, unsigned int src_row_pitch, unsigned int src_slice_pitch, unsigned int dst_row_pitch, unsigned dst_slice_pitch, unsigned int width, unsigned int height, unsigned int depth); + void (*decompress)(const BYTE *src, BYTE *dst, unsigned int src_row_pitch, unsigned int src_slice_pitch, + unsigned int dst_row_pitch, unsigned dst_slice_pitch, + unsigned int width, unsigned int height, unsigned int depth);
enum wined3d_format_id typeless_id; GLenum gl_view_class;