-/**************************************************************************
- *
- * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
+/*
+ * Copyright 2003 VMware, Inc.
* All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the
* "Software"), to deal in the Software without restriction, including
* without limitation the rights to use, copy, modify, merge, publish,
- * distribute, sub license, and/or sell copies of the Software, and to
+ * distribute, sublicense, and/or sell copies of the Software, and to
* permit persons to whom the Software is furnished to do so, subject to
* the following conditions:
*
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
* OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
- * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
- * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
* ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
- *
- **************************************************************************/
+ */
-#include "main/glheader.h"
#include "main/enums.h"
#include "main/mtypes.h"
#include "main/macros.h"
#include "main/bufferobj.h"
#include "main/readpix.h"
#include "main/state.h"
+#include "main/glformats.h"
+#include "drivers/common/meta.h"
#include "brw_context.h"
#include "intel_screen.h"
+#include "intel_batchbuffer.h"
#include "intel_blit.h"
#include "intel_buffers.h"
#include "intel_fbo.h"
#include "intel_mipmap_tree.h"
-#include "intel_regions.h"
#include "intel_pixel.h"
#include "intel_buffer_objects.h"
+#include "intel_tiled_memcpy.h"
#define FILE_DEBUG_FLAG DEBUG_PIXEL
-/* For many applications, the new ability to pull the source buffers
- * back out of the GTT and then do the packing/conversion operations
- * in software will be as much of an improvement as trying to get the
- * blitter and/or texture engine to do the work.
- *
- * This step is gated on private backbuffers.
- *
- * Obviously the frontbuffer can't be pulled back, so that is either
- * an argument for blit/texture readpixels, or for blitting to a
- * temporary and then pulling that back.
+/**
+ * \brief A fast path for glReadPixels
*
- * When the destination is a pbo, however, it's not clear if it is
- * ever going to be pulled to main memory (though the access param
- * will be a good hint). So it sounds like we do want to be able to
- * choose between blit/texture implementation on the gpu and pullback
- * and cpu-based copying.
+ * 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.
*
- * Unless you can magically turn client memory into a PBO for the
- * duration of this call, there will be a cpu-based copying step in
- * any case.
+ * 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
-do_blit_readpixels(struct gl_context * ctx,
- GLint x, GLint y, GLsizei width, GLsizei height,
- GLenum format, GLenum type,
- const struct gl_pixelstore_attrib *pack, GLvoid * pixels)
+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 intel_context *intel = intel_context(ctx);
- struct intel_buffer_object *dst = intel_buffer_object(pack->BufferObj);
- GLuint dst_offset;
- drm_intel_bo *dst_buffer;
- bool all;
- GLint dst_x, dst_y;
- GLuint dirty;
-
- DBG("%s\n", __FUNCTION__);
+ struct gl_renderbuffer *rb = ctx->ReadBuffer->_ColorReadBuffer;
- assert(_mesa_is_bufferobj(pack->BufferObj));
+ /* This path supports reading from color buffers only */
+ if (rb == NULL)
+ return false;
- struct gl_renderbuffer *rb = ctx->ReadBuffer->_ColorReadBuffer;
struct intel_renderbuffer *irb = intel_renderbuffer(rb);
+ int dst_pitch;
- if (ctx->_ImageTransferState ||
- !_mesa_format_matches_format_and_type(irb->mt->format, format, type,
- false)) {
- DBG("%s - bad format for blit\n", __FUNCTION__);
+ /* 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 (!brw->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;
- }
- if (pack->SwapBytes || pack->LsbFirst) {
- DBG("%s: bad packing params\n", __FUNCTION__);
+ /* Only a simple blit, no scale, bias or other mapping. */
+ if (ctx->_ImageTransferState)
return false;
- }
- int dst_stride = _mesa_image_row_stride(pack, width, format, type);
- bool dst_flip = false;
- /* Mesa flips the dst_stride for pack->Invert, but we want our mt to have a
- * normal dst_stride.
+ /* 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 (pack->Invert) {
- dst_stride = -dst_stride;
- dst_flip = true;
+ 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->tiling != I915_TILING_X &&
+ irb->mt->tiling != I915_TILING_Y)) {
+ /* The algorithm is written only for X- or Y-tiled memory. */
+ return false;
}
- dst_offset = (GLintptr)pixels;
- dst_offset += _mesa_image_offset(2, pack, width, height,
- format, type, 0, 0, 0);
+ /* 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 (brw->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 (!_mesa_clip_copytexsubimage(ctx,
- &dst_x, &dst_y,
- &x, &y,
- &width, &height)) {
- return true;
+ if (brw_batch_references(&brw->batch, bo)) {
+ perf_debug("Flushing before mapping a referenced bo.\n");
+ intel_batchbuffer_flush(brw);
}
- dirty = intel->front_buffer_dirty;
- intel_prepare_render(brw);
- intel->front_buffer_dirty = dirty;
-
- all = (width * height * irb->mt->cpp == dst->Base.Size &&
- x == 0 && dst_offset == 0);
-
- dst_buffer = intel_bufferobj_buffer(brw, dst,
- all ? INTEL_WRITE_FULL :
- INTEL_WRITE_PART);
-
- struct intel_mipmap_tree *pbo_mt =
- intel_miptree_create_for_bo(brw,
- dst_buffer,
- irb->mt->format,
- dst_offset,
- width, height,
- dst_stride, I915_TILING_NONE);
-
- if (!intel_miptree_blit(brw,
- irb->mt, irb->mt_level, irb->mt_layer,
- x, y, _mesa_is_winsys_fbo(ctx->ReadBuffer),
- pbo_mt, 0, 0,
- 0, 0, dst_flip,
- width, height, GL_COPY)) {
+ void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
+ if (map == NULL) {
+ DBG("%s: failed to map bo\n", __func__);
return false;
}
- intel_miptree_release(&pbo_mt);
+ xoffset += irb->mt->level[irb->mt_level].slice[irb->mt_layer].x_offset;
+ yoffset += irb->mt->level[irb->mt_level].slice[irb->mt_layer].y_offset;
- DBG("%s - DONE\n", __FUNCTION__);
+ 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->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->pitch,
+ brw->has_swizzling,
+ irb->mt->tiling,
+ mem_copy
+ );
+
+ brw_bo_unmap(bo);
return true;
}
GLenum format, GLenum type,
const struct gl_pixelstore_attrib *pack, GLvoid * pixels)
{
+ bool ok;
+
struct brw_context *brw = brw_context(ctx);
- struct intel_context *intel = intel_context(ctx);
bool dirty;
- DBG("%s\n", __FUNCTION__);
+ DBG("%s\n", __func__);
if (_mesa_is_bufferobj(pack->BufferObj)) {
- /* Using PBOs, so try the BLT based path. */
- if (do_blit_readpixels(ctx, x, y, width, height, format, type, pack,
- pixels)) {
+ if (_mesa_meta_pbo_GetTexSubImage(ctx, 2, NULL, x, y, 0, width, height, 1,
+ format, type, pixels, pack)) {
+ /* _mesa_meta_pbo_GetTexSubImage() implements PBO transfers by
+ * binding the user-provided BO as a fake framebuffer and rendering
+ * to it. This breaks the invariant of the GL that nothing is able
+ * to render to a BO, causing nondeterministic corruption issues
+ * because the render cache is not coherent with a number of other
+ * caches that the BO could potentially be bound to afterwards.
+ *
+ * This could be solved in the same way that we guarantee texture
+ * coherency after a texture is attached to a framebuffer and
+ * rendered to, but that would involve checking *all* BOs bound to
+ * the pipeline for the case we need to emit a cache flush due to
+ * previous rendering to any of them -- Including vertex, index,
+ * uniform, atomic counter, shader image, transform feedback,
+ * indirect draw buffers, etc.
+ *
+ * That would increase the per-draw call overhead even though it's
+ * very unlikely that any of the BOs bound to the pipeline has been
+ * rendered to via a PBO at any point, so it seems better to just
+ * flush here unconditionally.
+ */
+ brw_emit_mi_flush(brw);
return;
}
- perf_debug("%s: fallback to CPU mapping in PBO case\n", __FUNCTION__);
+ perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
}
- /* glReadPixels() wont dirty the front buffer, so reset the dirty
+ /* Reading pixels wont dirty the front buffer, so reset the dirty
* flag after calling intel_prepare_render(). */
- dirty = intel->front_buffer_dirty;
+ dirty = brw->front_buffer_dirty;
intel_prepare_render(brw);
- intel->front_buffer_dirty = dirty;
+ brw->front_buffer_dirty = dirty;
+
+ ok = intel_readpixels_tiled_memcpy(ctx, x, y, width, height,
+ format, type, pixels, pack);
+ if(ok)
+ return;
/* Update Mesa state before calling _mesa_readpixels().
* XXX this may not be needed since ReadPixels no longer uses the
_mesa_readpixels(ctx, x, y, width, height, format, type, pack, pixels);
/* There's an intel_prepare_render() call in intelSpanRenderStart(). */
- intel->front_buffer_dirty = dirty;
+ brw->front_buffer_dirty = dirty;
}