<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Sat, Sep 3, 2016 at 12:14 AM, Pohjolainen, Topi <<a href="mailto:topi.pohjolainen@gmail.com" target="_blank">topi.pohjolainen@gmail.com</a>> wrote: <blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">On Wed, Aug 31, 2016 at 02:22:33PM -0700, Jason Ekstrand wrote: > --- > src/intel/blorp/blorp.h | 10 ++++ > src/intel/blorp/blorp_blit.c | 133 +++++++++++++++++++++++++++++++++++++++++++ So the idea is to copy a compression block as a pixel (by choosing large enough format), right? And this is only for compression schemes that do not use auxiliary buffers? </blockquote><div> </div><div>Yes. Particularly, it's for the sort of compressed formats that get provided directly by users. It is also useful if you want to, for instance, copy from RGBA8_UNORM to RG16_FLOAT bit-for-bit without doing any conversion. In short, it's a memcpy() for image data. GL and Vulkan both have such APIs: glCopyImageSubData and vkCopyImage. </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> A few small comments further down. <div><div class="h5"> > 2 files changed, 143 insertions(+) > > diff --git a/src/intel/blorp/blorp.h b/src/intel/blorp/blorp.h > index c1e93fd..6574124 100644 > --- a/src/intel/blorp/blorp.h > +++ b/src/intel/blorp/blorp.h > @@ -109,6 +109,16 @@ blorp_blit(struct blorp_batch *batch, > uint32_t filter, bool mirror_x, bool mirror_y); > > void > +blorp_copy(struct blorp_batch *batch, > + const struct blorp_surf *src_surf, > + unsigned src_level, unsigned src_layer, > + const struct blorp_surf *dst_surf, > + unsigned dst_level, unsigned dst_layer, > + uint32_t src_x, uint32_t src_y, > + uint32_t dst_x, uint32_t dst_y, > + uint32_t src_width, uint32_t src_height); > + > +void > blorp_fast_clear(struct blorp_batch *batch, > const struct blorp_surf *surf, > uint32_t level, uint32_t layer, enum isl_format format, > diff --git a/src/intel/blorp/blorp_blit.c b/src/intel/blorp/blorp_blit.c > index 3ab39a3..42a502c 100644 > --- a/src/intel/blorp/blorp_blit.c > +++ b/src/intel/blorp/blorp_blit.c > @@ -1685,3 +1685,136 @@ blorp_blit(struct blorp_batch *batch, > dst_x0, dst_y0, dst_x1, dst_y1, > mirror_x, mirror_y); > } > + > +static enum isl_format > +get_copy_format_for_bpb(unsigned bpb) > +{ > + /* The choice of UNORM and UINT formats is very intentional here. Most of > + * the time, we want to use a UINT format to avoid any rounding error in > + * the blit. For stencil blits, R8_UINT is required by the hardware. > + * (It's the only format allowed in conjunction with W-tiling.) Also we > + * intentionally use the 4-channel formats whenever we can. This is so > + * that, when we do a RGB <-> RGBX copy, the two formats will line up even > + * though one of them is 3/4 the size of the other. The choice of UNORM > + * vs. UINT is also very intentional because Haswell doesn't handle 8 or > + * 16-bit RGB UINT formats at all so we have to use UNORM there. > + * Fortunately, the only time we should ever use two different formats in > + * the table below is for RGB -> RGBA blits and so we will never have any > + * UNORM/UINT mismatch. > + */ > + switch (bpb) { > + case 8: return ISL_FORMAT_R8_UINT; > + case 16: return ISL_FORMAT_R8G8_UINT; > + case 24: return ISL_FORMAT_R8G8B8_UNORM; > + case 32: return ISL_FORMAT_R8G8B8A8_UNORM; > + case 48: return ISL_FORMAT_R16G16B16_UNORM; > + case 64: return ISL_FORMAT_R16G16B16A16_UNORM; > + case 96: return ISL_FORMAT_R32G32B32_UINT; > + case 128:return ISL_FORMAT_R32G32B32A32_UINT; > + default: > + unreachable("Unknown format bpb"); > + } > +} > + > +static void > +surf_convert_to_uncompressed(const struct isl_device *isl_dev, > + struct brw_blorp_surface_info *info, > + uint32_t *x, uint32_t *y, > + uint32_t *width, uint32_t *height) > +{ > + const struct isl_format_layout *fmtl = > + isl_format_get_layout(info->surf.format); > + > + assert(fmtl->bw > 1 || fmtl->bh > 1); > + > + /* This is a compressed surface. We need to convert it to a single > + * slice (becase compressed layouts don't perfectly match uncompressed </div></div> because <div><div class="h5"></div></div></blockquote><div> </div><div>Yup </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div><div class="h5"> > + * ones with the same bpb) and divide x, y, width, and height by the > + * block size. > + */ > + surf_convert_to_single_slice(isl_dev, info); > + > + if (width || height) { > + assert(*width % fmtl->bw == 0 || > + *x + *width == info->surf.logical_level0_px.width); > + assert(*height % fmtl->bh == 0 || > + *y + *height == info->surf.logical_level0_px.height); > + *width = DIV_ROUND_UP(*width, fmtl->bw); > + *height = DIV_ROUND_UP(*height, fmtl->bh); > + } > + > + assert(*x % fmtl->bw == 0); > + assert(*y % fmtl->bh == 0); > + *x /= fmtl->bw; > + *y /= fmtl->bh; > + > + info->surf.logical_level0_px.width = > + DIV_ROUND_UP(info->surf.logical_level0_px.width, fmtl->bw); > + info->surf.logical_level0_px.height = > + DIV_ROUND_UP(info->surf.logical_level0_px.height, fmtl->bh); > + > + assert(info->surf.phys_level0_sa.width % fmtl->bw == 0); > + assert(info->surf.phys_level0_sa.height % fmtl->bh == 0); > + info->surf.phys_level0_sa.width /= fmtl->bw; > + info->surf.phys_level0_sa.height /= fmtl->bh; > + > + assert(info->tile_x_sa % fmtl->bw == 0); > + assert(info->tile_y_sa % fmtl->bh == 0); > + info->tile_x_sa /= fmtl->bw; > + info->tile_y_sa /= fmtl->bh; > + > + /* It's now an uncompressed surface so we need an uncompressed format */ > + info->surf.format = get_copy_format_for_bpb(fmtl->bpb); > +} > + > +void > +blorp_copy(struct blorp_batch *batch, > + const struct blorp_surf *src_surf, > + unsigned src_level, unsigned src_layer, > + const struct blorp_surf *dst_surf, > + unsigned dst_level, unsigned dst_layer, > + uint32_t src_x, uint32_t src_y, > + uint32_t dst_x, uint32_t dst_y, > + uint32_t src_width, uint32_t src_height) > +{ > + struct blorp_params params; > + blorp_params_init(&params); > + > + brw_blorp_surface_info_init(batch->blorp, &params.src, src_surf, src_level, > + src_layer, ISL_FORMAT_UNSUPPORTED, false); > + brw_blorp_surface_info_init(batch->blorp, &params.dst, dst_surf, dst_level, > + dst_layer, ISL_FORMAT_UNSUPPORTED, true); > + > + struct brw_blorp_blit_prog_key wm_prog_key; > + memset(&wm_prog_key, 0, sizeof(wm_prog_key)); > + > + const struct isl_format_layout *src_fmtl = > + isl_format_get_layout(params.src.surf.format); > + const struct isl_format_layout *dst_fmtl = > + isl_format_get_layout(params.dst.surf.format); > + > + params.src.view.format = get_copy_format_for_bpb(src_fmtl->bpb); > + if (src_fmtl->bw > 1 || src_fmtl->bh > 1) { > + surf_convert_to_uncompressed(batch->blorp->isl_dev, &params.src, > + &src_x, &src_y, &src_width, &src_height); > + wm_prog_key.need_dst_offset = true; </div></div>I think you mean wm_prog_key.need_src_offset instead? </blockquote><div> </div><div>Drp... Thanks! </div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> > + } > + > + params.dst.view.format = get_copy_format_for_bpb(dst_fmtl->bpb); > + if (dst_fmtl->bw > 1 || dst_fmtl->bh > 1) { > + surf_convert_to_uncompressed(batch->blorp->isl_dev, &params.dst, > + &dst_x, &dst_y, NULL, NULL); > + wm_prog_key.need_dst_offset = true; > + } > + > + /* Once both surfaces are stompped to uncompressed as needed, the > + * destination size is the same as the source size. > + */ > + uint32_t dst_width = src_width; > + uint32_t dst_height = src_height; > + > + do_blorp_blit(batch, &params, &wm_prog_key, > + src_x, src_y, src_x + src_width, src_y + src_height, > + dst_x, dst_y, dst_x + dst_width, dst_y + dst_height, > + false, false); > +} > -- > 2.5.0.400.gff86faf > > _______________________________________________ > mesa-dev mailing list > <a href="mailto:mesa-dev@lists.freedesktop.org">mesa-dev@lists.freedesktop.org</a> > <a href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev" rel="noreferrer" target="_blank">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a> </blockquote></div> </div></div>