/**
* \brief A fast path for glReadPixels
*
* This fast path is taken when the source format is BGRA, RGBA,
* A or L and when the texture memory is X- or Y-tiled. It downloads
* the source data by directly mapping the memory without a GTT fence.
* This then needs to be de-tiled on the CPU before presenting the data to
* the user in the linear fasion.
*
* This is a performance win over the conventional texture download path.
* In the conventional texture download path, the texture is either mapped
* through the GTT or copied to a linear buffer with the blitter before
* handing off to a software path. This allows us to avoid round-tripping
* through the GPU (in the case where we would be blitting) and do only a
* single copy operation.
*/
static bool
intel_readpixels_tiled_memcpy(struct gl_context * ctx,
GLint xoffset, GLint yoffset,
GLsizei width, GLsizei height,
GLenum format, GLenum type,
GLvoid * pixels,
const struct gl_pixelstore_attrib *pack)
{
struct brw_context *brw = brw_context(ctx);
struct gl_renderbuffer *rb = ctx->ReadBuffer->_ColorReadBuffer;
const struct gen_device_info *devinfo = &brw->screen->devinfo;
/* This path supports reading from color buffers only */
if (rb == NULL)
return false;
struct intel_renderbuffer *irb = intel_renderbuffer(rb);
int dst_pitch;
/* The miptree's buffer. */
struct brw_bo *bo;
uint32_t cpp;
mem_copy_fn mem_copy = NULL;
/* This fastpath is restricted to specific renderbuffer types:
* a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
* more types.
*/
if (!devinfo->has_llc ||
!(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
pixels == NULL ||
_mesa_is_bufferobj(pack->BufferObj) ||
pack->Alignment > 4 ||
pack->SkipPixels > 0 ||
pack->SkipRows > 0 ||
(pack->RowLength != 0 && pack->RowLength != width) ||
pack->SwapBytes ||
pack->LsbFirst ||
pack->Invert)
return false;
/* Only a simple blit, no scale, bias or other mapping. */
if (ctx->_ImageTransferState)
return false;
/* It is possible that the renderbuffer (or underlying texture) is
* multisampled. Since ReadPixels from a multisampled buffer requires a
* multisample resolve, we can't handle this here
*/
if (rb->NumSamples > 1)
return false;
/* We can't handle copying from RGBX or BGRX because the tiled_memcpy
* function doesn't set the last channel to 1. Note this checks BaseFormat
* rather than TexFormat in case the RGBX format is being simulated with an
* RGBA format.
*/
if (rb->_BaseFormat == GL_RGB)
return false;
if (!intel_get_memcpy(rb->Format, format, type, &mem_copy, &cpp))
return false;
if (!irb->mt ||
(irb->mt->surf.tiling != ISL_TILING_X &&
irb->mt->surf.tiling != ISL_TILING_Y0)) {
/* The algorithm is written only for X- or Y-tiled memory. */
return false;
}
/* tiled_to_linear() assumes that if the object is swizzled, it is using
* I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y. This is only
* true on gen5 and above.
*
* The killer on top is that some gen4 have an L-shaped swizzle mode, where
* parts of the memory aren't swizzled at all. Userspace just can't handle
* that.
*/
if (devinfo->gen < 5 && brw->has_swizzling)
return false;
/* Since we are going to read raw data to the miptree, we need to resolve
* any pending fast color clears before we start.
*/
intel_miptree_access_raw(brw, irb->mt, irb->mt_level, irb->mt_layer, false);
bo = irb->mt->bo;
if (brw_batch_references(&brw->batch, bo)) {
perf_debug("Flushing before mapping a referenced bo.\n");
intel_batchbuffer_flush(brw);
}
void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
if (map == NULL) {
DBG("%s: failed to map bo\n", __func__);
return false;
}
unsigned slice_offset_x, slice_offset_y;
intel_miptree_get_image_offset(irb->mt, irb->mt_level, irb->mt_layer,
&slice_offset_x, &slice_offset_y);
xoffset += slice_offset_x;
yoffset += slice_offset_y;
dst_pitch = _mesa_image_row_stride(pack, width, format, type);
/* For a window-system renderbuffer, the buffer is actually flipped
* vertically, so we need to handle that. Since the detiling function
* can only really work in the forwards direction, we have to be a
* little creative. First, we compute the Y-offset of the first row of
* the renderbuffer (in renderbuffer coordinates). We then match that
* with the last row of the client's data. Finally, we give
* tiled_to_linear a negative pitch so that it walks through the
* client's data backwards as it walks through the renderbufer forwards.
*/
if (rb->Name == 0) {
yoffset = rb->Height - yoffset - height;
pixels += (ptrdiff_t) (height - 1) * dst_pitch;
dst_pitch = -dst_pitch;
}
/* We postponed printing this message until having committed to executing
* the function.
*/
DBG("%s: x,y=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
"mesa_format=0x%x tiling=%d "
"pack=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
__func__, xoffset, yoffset, width, height,
format, type, rb->Format, irb->mt->surf.tiling,
pack->Alignment, pack->RowLength, pack->SkipPixels,
pack->SkipRows);
tiled_to_linear(
xoffset * cpp, (xoffset + width) * cpp,
yoffset, yoffset + height,
pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
map + irb->mt->offset,
dst_pitch, irb->mt->surf.row_pitch,
brw->has_swizzling,
irb->mt->surf.tiling,
mem_copy
);
brw_bo_unmap(bo);
return true;
}
bool
intel_miptree_copy(struct brw_context *brw,
struct intel_mipmap_tree *src_mt,
int src_level, int src_slice,
uint32_t src_x, uint32_t src_y,
struct intel_mipmap_tree *dst_mt,
int dst_level, int dst_slice,
uint32_t dst_x, uint32_t dst_y,
uint32_t src_width, uint32_t src_height)
{
/* The blitter doesn't understand multisampling at all. */
if (src_mt->surf.samples > 1 || dst_mt->surf.samples > 1)
return false;
if (src_mt->format == MESA_FORMAT_S_UINT8)
return false;
/* The blitter has no idea about HiZ or fast color clears, so we need to
* resolve the miptrees before we do anything.
*/
intel_miptree_access_raw(brw, src_mt, src_level, src_slice, false);
intel_miptree_access_raw(brw, dst_mt, dst_level, dst_slice, true);
uint32_t src_image_x, src_image_y;
intel_miptree_get_image_offset(src_mt, src_level, src_slice,
&src_image_x, &src_image_y);
if (_mesa_is_format_compressed(src_mt->format)) {
GLuint bw, bh;
_mesa_get_format_block_size(src_mt->format, &bw, &bh);
/* Compressed textures need not have dimensions that are a multiple of
* the block size. Rectangles in compressed textures do need to be a
* multiple of the block size. The one exception is that the right and
* bottom edges may be at the right or bottom edge of the miplevel even
* if it's not aligned.
*/
assert(src_x % bw == 0);
assert(src_y % bh == 0);
assert(src_width % bw == 0 ||
src_x + src_width ==
minify(src_mt->surf.logical_level0_px.width, src_level));
assert(src_height % bh == 0 ||
src_y + src_height ==
minify(src_mt->surf.logical_level0_px.height, src_level));
src_x /= (int)bw;
src_y /= (int)bh;
src_width = DIV_ROUND_UP(src_width, (int)bw);
src_height = DIV_ROUND_UP(src_height, (int)bh);
}
src_x += src_image_x;
src_y += src_image_y;
uint32_t dst_image_x, dst_image_y;
intel_miptree_get_image_offset(dst_mt, dst_level, dst_slice,
&dst_image_x, &dst_image_y);
if (_mesa_is_format_compressed(dst_mt->format)) {
GLuint bw, bh;
_mesa_get_format_block_size(dst_mt->format, &bw, &bh);
assert(dst_x % bw == 0);
assert(dst_y % bh == 0);
dst_x /= (int)bw;
dst_y /= (int)bh;
}
dst_x += dst_image_x;
dst_y += dst_image_y;
return emit_miptree_blit(brw, src_mt, src_x, src_y,
dst_mt, dst_x, dst_y,
src_width, src_height, false, COLOR_LOGICOP_COPY);
}
/*
* Render a bitmap.
*/
static bool
do_blit_bitmap( struct gl_context *ctx,
GLint dstx, GLint dsty,
GLsizei width, GLsizei height,
const struct gl_pixelstore_attrib *unpack,
const GLubyte *bitmap )
{
struct brw_context *brw = brw_context(ctx);
struct gl_framebuffer *fb = ctx->DrawBuffer;
struct intel_renderbuffer *irb;
GLfloat tmpColor[4];
GLubyte ubcolor[4];
GLuint color;
GLsizei bitmap_width = width;
GLsizei bitmap_height = height;
GLint px, py;
GLuint stipple[32];
GLint orig_dstx = dstx;
GLint orig_dsty = dsty;
/* Update draw buffer bounds */
_mesa_update_state(ctx);
if (ctx->Depth.Test) {
/* The blit path produces incorrect results when depth testing is on.
* It seems the blit Z coord is always 1.0 (the far plane) so fragments
* will likely be obscured by other, closer geometry.
*/
return false;
}
intel_prepare_render(brw);
if (fb->_NumColorDrawBuffers != 1) {
perf_debug("accelerated glBitmap() only supports rendering to a "
"single color buffer\n");
return false;
}
irb = intel_renderbuffer(fb->_ColorDrawBuffers[0]);
if (_mesa_is_bufferobj(unpack->BufferObj)) {
bitmap = map_pbo(ctx, width, height, unpack, bitmap);
if (bitmap == NULL)
return true; /* even though this is an error, we're done */
}
COPY_4V(tmpColor, ctx->Current.RasterColor);
if (_mesa_need_secondary_color(ctx)) {
ADD_3V(tmpColor, tmpColor, ctx->Current.RasterSecondaryColor);
}
UNCLAMPED_FLOAT_TO_UBYTE(ubcolor[0], tmpColor[0]);
UNCLAMPED_FLOAT_TO_UBYTE(ubcolor[1], tmpColor[1]);
UNCLAMPED_FLOAT_TO_UBYTE(ubcolor[2], tmpColor[2]);
UNCLAMPED_FLOAT_TO_UBYTE(ubcolor[3], tmpColor[3]);
switch (_mesa_get_render_format(ctx, intel_rb_format(irb))) {
case MESA_FORMAT_B8G8R8A8_UNORM:
case MESA_FORMAT_B8G8R8X8_UNORM:
color = PACK_COLOR_8888(ubcolor[3], ubcolor[0], ubcolor[1], ubcolor[2]);
break;
case MESA_FORMAT_B5G6R5_UNORM:
color = PACK_COLOR_565(ubcolor[0], ubcolor[1], ubcolor[2]);
break;
default:
perf_debug("Unsupported format %s in accelerated glBitmap()\n",
_mesa_get_format_name(irb->mt->format));
return false;
}
if (!intel_check_blit_fragment_ops(ctx, tmpColor[3] == 1.0F))
return false;
/* Clip to buffer bounds and scissor. */
if (!_mesa_clip_to_region(fb->_Xmin, fb->_Ymin,
fb->_Xmax, fb->_Ymax,
&dstx, &dsty, &width, &height))
goto out;
dsty = y_flip(fb, dsty, height);
#define DY 32
#define DX 32
/* The blitter has no idea about fast color clears, so we need to resolve
* the miptree before we do anything.
*/
intel_miptree_access_raw(brw, irb->mt, irb->mt_level, irb->mt_layer, true);
/* Chop it all into chunks that can be digested by hardware: */
for (py = 0; py < height; py += DY) {
for (px = 0; px < width; px += DX) {
int h = MIN2(DY, height - py);
int w = MIN2(DX, width - px);
GLuint sz = ALIGN(ALIGN(w,8) * h, 64)/8;
GLenum logic_op = ctx->Color.ColorLogicOpEnabled ?
ctx->Color.LogicOp : GL_COPY;
assert(sz <= sizeof(stipple));
memset(stipple, 0, sz);
/* May need to adjust this when padding has been introduced in
* sz above:
*
* Have to translate destination coordinates back into source
* coordinates.
*/
int count = get_bitmap_rect(bitmap_width, bitmap_height, unpack,
bitmap,
-orig_dstx + (dstx + px),
-orig_dsty + y_flip(fb, dsty + py, h),
w, h,
(GLubyte *)stipple,
8,
_mesa_is_winsys_fbo(fb));
if (count == 0)
continue;
if (!intelEmitImmediateColorExpandBlit(brw,
irb->mt->cpp,
(GLubyte *)stipple,
sz,
color,
irb->mt->surf.row_pitch,
irb->mt->bo,
0,
irb->mt->surf.tiling,
dstx + px,
dsty + py,
w, h,
logic_op)) {
return false;
}
if (ctx->Query.CurrentOcclusionObject)
ctx->Query.CurrentOcclusionObject->Result += count;
}
}
out:
if (unlikely(INTEL_DEBUG & DEBUG_SYNC))
intel_batchbuffer_flush(brw);
if (_mesa_is_bufferobj(unpack->BufferObj)) {
/* done with PBO so unmap it now */
ctx->Driver.UnmapBuffer(ctx, unpack->BufferObj, MAP_INTERNAL);
}
return true;
}
/**
* Implements a rectangular block transfer (blit) of pixels between two
* miptrees.
*
* Our blitter can operate on 1, 2, or 4-byte-per-pixel data, with generous,
* but limited, pitches and sizes allowed.
*
* The src/dst coordinates are relative to the given level/slice of the
* miptree.
*
* If @src_flip or @dst_flip is set, then the rectangle within that miptree
* will be inverted (including scanline order) when copying. This is common
* in GL when copying between window system and user-created
* renderbuffers/textures.
*/
bool
intel_miptree_blit(struct brw_context *brw,
struct intel_mipmap_tree *src_mt,
int src_level, int src_slice,
uint32_t src_x, uint32_t src_y, bool src_flip,
struct intel_mipmap_tree *dst_mt,
int dst_level, int dst_slice,
uint32_t dst_x, uint32_t dst_y, bool dst_flip,
uint32_t width, uint32_t height,
enum gl_logicop_mode logicop)
{
/* The blitter doesn't understand multisampling at all. */
if (src_mt->surf.samples > 1 || dst_mt->surf.samples > 1)
return false;
/* No sRGB decode or encode is done by the hardware blitter, which is
* consistent with what we want in many callers (glCopyTexSubImage(),
* texture validation, etc.).
*/
mesa_format src_format = _mesa_get_srgb_format_linear(src_mt->format);
mesa_format dst_format = _mesa_get_srgb_format_linear(dst_mt->format);
/* The blitter doesn't support doing any format conversions. We do also
* support blitting ARGB8888 to XRGB8888 (trivial, the values dropped into
* the X channel don't matter), and XRGB8888 to ARGB8888 by setting the A
* channel to 1.0 at the end. Also trivially ARGB2101010 to XRGB2101010,
* but not XRGB2101010 to ARGB2101010 yet.
*/
if (!intel_miptree_blit_compatible_formats(src_format, dst_format)) {
perf_debug("%s: Can't use hardware blitter from %s to %s, "
"falling back.\n", __func__,
_mesa_get_format_name(src_format),
_mesa_get_format_name(dst_format));
return false;
}
/* The blitter has no idea about HiZ or fast color clears, so we need to
* resolve the miptrees before we do anything.
*/
intel_miptree_access_raw(brw, src_mt, src_level, src_slice, false);
intel_miptree_access_raw(brw, dst_mt, dst_level, dst_slice, true);
if (src_flip) {
const unsigned h0 = src_mt->surf.phys_level0_sa.height;
src_y = minify(h0, src_level - src_mt->first_level) - src_y - height;
}
if (dst_flip) {
const unsigned h0 = dst_mt->surf.phys_level0_sa.height;
dst_y = minify(h0, dst_level - dst_mt->first_level) - dst_y - height;
}
uint32_t src_image_x, src_image_y, dst_image_x, dst_image_y;
intel_miptree_get_image_offset(src_mt, src_level, src_slice,
&src_image_x, &src_image_y);
intel_miptree_get_image_offset(dst_mt, dst_level, dst_slice,
&dst_image_x, &dst_image_y);
src_x += src_image_x;
src_y += src_image_y;
dst_x += dst_image_x;
dst_y += dst_image_y;
if (!emit_miptree_blit(brw, src_mt, src_x, src_y,
dst_mt, dst_x, dst_y, width, height,
src_flip != dst_flip, logicop)) {
return false;
}
/* XXX This could be done in a single pass using XY_FULL_MONO_PATTERN_BLT */
if (_mesa_get_format_bits(src_format, GL_ALPHA_BITS) == 0 &&
_mesa_get_format_bits(dst_format, GL_ALPHA_BITS) > 0) {
intel_miptree_set_alpha_to_one(brw, dst_mt,
dst_x, dst_y,
width, height);
}
return true;
}