X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fpanfrost%2Fpan_resource.c;h=a34656b5bdfa3a96516b4b043731ae068cd68b75;hb=39bf1fb3221c0ab79e4609ca3588e2eeb9e3e2da;hp=766edee3ca6f8f58dd8322c60d8b32e4c8dd5807;hpb=29d0e5c1776cf430f0fd7d262df55670a5ff8d05;p=mesa.git diff --git a/src/gallium/drivers/panfrost/pan_resource.c b/src/gallium/drivers/panfrost/pan_resource.c index 766edee3ca6..a34656b5bdf 100644 --- a/src/gallium/drivers/panfrost/pan_resource.c +++ b/src/gallium/drivers/panfrost/pan_resource.c @@ -1,6 +1,7 @@ /* * Copyright (C) 2008 VMware, Inc. - * Copyright (C) 2014 Broadcom + * Copyright (C) 2012 Rob Clark + * Copyright (C) 2014-2017 Broadcom * Copyright (C) 2018-2019 Alyssa Rosenzweig * Copyright (C) 2019 Collabora, Ltd. * @@ -33,30 +34,23 @@ #include #include "drm-uapi/drm_fourcc.h" -#include "state_tracker/winsys_handle.h" -#include "util/u_format.h" +#include "frontend/winsys_handle.h" +#include "util/format/u_format.h" #include "util/u_memory.h" #include "util/u_surface.h" #include "util/u_transfer.h" #include "util/u_transfer_helper.h" #include "util/u_gen_mipmap.h" +#include "util/u_drm.h" +#include "pan_bo.h" #include "pan_context.h" #include "pan_screen.h" #include "pan_resource.h" #include "pan_util.h" #include "pan_tiling.h" - -static void -panfrost_resource_reset_damage(struct panfrost_resource *pres) -{ - /* We set the damage extent to the full resource size but keep the - * damage box empty so that the FB content is reloaded by default. - */ - memset(&pres->damage, 0, sizeof(pres->damage)); - pres->damage.extent.maxx = pres->base.width0; - pres->damage.extent.maxy = pres->base.height0; -} +#include "decode.h" +#include "panfrost-quirks.h" static struct pipe_resource * panfrost_resource_from_handle(struct pipe_screen *pscreen, @@ -64,7 +58,7 @@ panfrost_resource_from_handle(struct pipe_screen *pscreen, struct winsys_handle *whandle, unsigned usage) { - struct panfrost_screen *screen = pan_screen(pscreen); + struct panfrost_device *dev = pan_device(pscreen); struct panfrost_resource *rsc; struct pipe_resource *prsc; @@ -81,14 +75,31 @@ panfrost_resource_from_handle(struct pipe_screen *pscreen, pipe_reference_init(&prsc->reference, 1); prsc->screen = pscreen; - rsc->bo = panfrost_bo_import(screen, whandle->handle); + rsc->bo = panfrost_bo_import(dev, whandle->handle); + rsc->internal_format = templat->format; + rsc->modifier = (whandle->modifier == DRM_FORMAT_MOD_INVALID) ? + DRM_FORMAT_MOD_LINEAR : whandle->modifier; rsc->slices[0].stride = whandle->stride; + rsc->slices[0].offset = whandle->offset; rsc->slices[0].initialized = true; - panfrost_resource_reset_damage(rsc); + panfrost_resource_set_damage_region(NULL, &rsc->base, 0, NULL); + + if (dev->quirks & IS_BIFROST && + templat->bind & PIPE_BIND_RENDER_TARGET) { + unsigned size = panfrost_compute_checksum_size( + &rsc->slices[0], templat->width0, templat->height0); + rsc->slices[0].checksum_bo = panfrost_bo_create(dev, size, 0); + rsc->checksummed = true; + } - if (screen->ro) { + if (drm_is_afbc(whandle->modifier)) { + rsc->slices[0].header_size = + panfrost_afbc_header_size(templat->width0, templat->height0); + } + + if (dev->ro) { rsc->scanout = - renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL); + renderonly_create_gpu_import_for_resource(prsc, dev->ro, NULL); /* failure is expected in some cases.. */ } @@ -102,11 +113,11 @@ panfrost_resource_get_handle(struct pipe_screen *pscreen, struct winsys_handle *handle, unsigned usage) { - struct panfrost_screen *screen = pan_screen(pscreen); + struct panfrost_device *dev = pan_device(pscreen); struct panfrost_resource *rsrc = (struct panfrost_resource *) pt; struct renderonly_scanout *scanout = rsrc->scanout; - handle->modifier = DRM_FORMAT_MOD_INVALID; + handle->modifier = rsrc->modifier; if (handle->type == WINSYS_HANDLE_TYPE_SHARED) { return false; @@ -116,6 +127,7 @@ panfrost_resource_get_handle(struct pipe_screen *pscreen, handle->handle = rsrc->bo->gem_handle; handle->stride = rsrc->slices[0].stride; + handle->offset = rsrc->slices[0].offset; return TRUE; } else if (handle->type == WINSYS_HANDLE_TYPE_FD) { if (scanout) { @@ -124,7 +136,7 @@ panfrost_resource_get_handle(struct pipe_screen *pscreen, .flags = DRM_CLOEXEC, }; - int ret = drmIoctl(screen->ro->kms_fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args); + int ret = drmIoctl(dev->ro->kms_fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args); if (ret == -1) return false; @@ -133,13 +145,14 @@ panfrost_resource_get_handle(struct pipe_screen *pscreen, return true; } else { - int fd = panfrost_bo_export(screen, rsrc->bo); + int fd = panfrost_bo_export(rsrc->bo); if (fd < 0) return false; handle->handle = fd; handle->stride = rsrc->slices[0].stride; + handle->offset = rsrc->slices[0].offset; return true; } } @@ -150,7 +163,7 @@ panfrost_resource_get_handle(struct pipe_screen *pscreen, static void panfrost_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc) { - //DBG("TODO %s\n", __func__); + /* TODO */ } static struct pipe_surface * @@ -172,6 +185,7 @@ panfrost_create_surface(struct pipe_context *pipe, assert(surf_tmpl->u.tex.level <= pt->last_level); ps->width = u_minify(pt->width0, surf_tmpl->u.tex.level); ps->height = u_minify(pt->height0, surf_tmpl->u.tex.level); + ps->nr_samples = surf_tmpl->nr_samples; ps->u.tex.level = surf_tmpl->u.tex.level; ps->u.tex.first_layer = surf_tmpl->u.tex.first_layer; ps->u.tex.last_layer = surf_tmpl->u.tex.last_layer; @@ -200,21 +214,54 @@ panfrost_surface_destroy(struct pipe_context *pipe, static struct pipe_resource * panfrost_create_scanout_res(struct pipe_screen *screen, - const struct pipe_resource *template) + const struct pipe_resource *template, + uint64_t modifier) { - struct panfrost_screen *pscreen = pan_screen(screen); - struct pipe_resource scanout_templat = *template; + struct panfrost_device *dev = pan_device(screen); struct renderonly_scanout *scanout; struct winsys_handle handle; struct pipe_resource *res; + struct pipe_resource scanout_templat = *template; + + /* Tiled formats need to be tile aligned */ + if (modifier == DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED) { + scanout_templat.width0 = ALIGN_POT(template->width0, 16); + scanout_templat.height0 = ALIGN_POT(template->height0, 16); + } + + /* AFBC formats need a header. Thankfully we don't care about the + * stride so we can just use wonky dimensions as long as the right + * number of bytes are allocated at the end of the day... this implies + * that stride/pitch is invalid for AFBC buffers */ + + if (drm_is_afbc(modifier)) { + /* Space for the header. We need to keep vaguely similar + * dimensions because... reasons... to allocate with renderonly + * as a dumb buffer. To do so, after the usual 16x16 alignment, + * we add on extra rows for the header. The order of operations + * matters here, the extra rows of padding can in fact be + * needed and missing them can lead to faults. */ + + unsigned header_size = panfrost_afbc_header_size( + template->width0, template->height0); + + unsigned pitch = ALIGN_POT(template->width0, 16) * + util_format_get_blocksize(template->format); + + unsigned header_rows = + DIV_ROUND_UP(header_size, pitch); + + scanout_templat.width0 = ALIGN_POT(template->width0, 16); + scanout_templat.height0 = ALIGN_POT(template->height0, 16) + header_rows; + } scanout = renderonly_scanout_for_resource(&scanout_templat, - pscreen->ro, &handle); + dev->ro, &handle); if (!scanout) return NULL; assert(handle.type == WINSYS_HANDLE_TYPE_FD); - /* TODO: handle modifiers? */ + handle.modifier = modifier; res = screen->resource_from_handle(screen, template, &handle, PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE); close(handle.handle); @@ -228,30 +275,7 @@ panfrost_create_scanout_res(struct pipe_screen *screen, return res; } -/* Computes sizes for checksumming, which is 8 bytes per 16x16 tile */ - -#define CHECKSUM_TILE_WIDTH 16 -#define CHECKSUM_TILE_HEIGHT 16 -#define CHECKSUM_BYTES_PER_TILE 8 - -static unsigned -panfrost_compute_checksum_sizes( - struct panfrost_slice *slice, - unsigned width, - unsigned height) -{ - unsigned aligned_width = ALIGN_POT(width, CHECKSUM_TILE_WIDTH); - unsigned aligned_height = ALIGN_POT(height, CHECKSUM_TILE_HEIGHT); - - unsigned tile_count_x = aligned_width / CHECKSUM_TILE_WIDTH; - unsigned tile_count_y = aligned_height / CHECKSUM_TILE_HEIGHT; - - slice->checksum_stride = tile_count_x * CHECKSUM_BYTES_PER_TILE; - - return slice->checksum_stride * tile_count_y; -} - -/* Setup the mip tree given a particular layout, possibly with checksumming */ +/* Setup the mip tree given a particular modifier, possibly with checksumming */ static void panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size) @@ -260,7 +284,17 @@ panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size) unsigned width = res->width0; unsigned height = res->height0; unsigned depth = res->depth0; - unsigned bytes_per_pixel = util_format_get_blocksize(res->format); + unsigned bytes_per_pixel = util_format_get_blocksize(pres->internal_format); + + /* MSAA is implemented as a 3D texture with z corresponding to the + * sample #, horrifyingly enough */ + + bool msaa = res->nr_samples > 1; + + if (msaa) { + assert(depth == 1); + depth = res->nr_samples; + } assert(depth > 0); @@ -270,10 +304,12 @@ panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size) * makes code a lot simpler */ bool renderable = res->bind & - (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL); - bool afbc = pres->layout == PAN_AFBC; - bool tiled = pres->layout == PAN_TILED; - bool should_align = renderable || tiled; + (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL) && + res->target != PIPE_BUFFER; + bool afbc = drm_is_afbc(pres->modifier); + bool tiled = pres->modifier == DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED; + bool linear = pres->modifier == DRM_FORMAT_MOD_LINEAR; + bool should_align = renderable || tiled || afbc; /* We don't know how to specify a 2D stride for 3D textures */ @@ -309,8 +345,11 @@ panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size) /* Compute the would-be stride */ unsigned stride = bytes_per_pixel * effective_width; + if (util_format_is_compressed(pres->internal_format)) + stride /= 4; + /* ..but cache-line align it for performance */ - if (can_align_stride && pres->layout == PAN_LINEAR) + if (can_align_stride && linear) stride = ALIGN_POT(stride, 64); slice->stride = stride; @@ -318,6 +357,8 @@ panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size) unsigned slice_one_size = slice->stride * effective_height; unsigned slice_full_size = slice_one_size * effective_depth; + slice->size0 = slice_one_size; + /* Report 2D size for 3D texturing */ if (l == 0) @@ -337,7 +378,7 @@ panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size) if (pres->checksummed) { slice->checksum_offset = offset; - unsigned size = panfrost_compute_checksum_sizes( + unsigned size = panfrost_compute_checksum_size( slice, width, height); offset += size; @@ -345,7 +386,10 @@ panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size) width = u_minify(width, 1); height = u_minify(height, 1); - depth = u_minify(depth, 1); + + /* Don't mipmap the sample count */ + if (!msaa) + depth = u_minify(depth, 1); } assert(res->array_size); @@ -364,47 +408,137 @@ panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size) } } -static void -panfrost_resource_create_bo(struct panfrost_screen *screen, struct panfrost_resource *pres) +/* Based on the usage, determine if it makes sense to use u-inteleaved tiling. + * We only have routines to tile 2D textures of sane bpps. On the hardware + * level, not all usages are valid for tiling. Finally, if the app is hinting + * that the contents frequently change, tiling will be a loss. + * + * Due to incomplete information on some platforms, we may need to force tiling + * in some cases. + * + * On platforms where it is supported, AFBC is even better. */ + +static bool +panfrost_can_linear(struct panfrost_device *dev, const struct panfrost_resource *pres) { - struct pipe_resource *res = &pres->base; + /* XXX: We should be able to do linear Z/S with the right bits.. */ + return !((pres->base.bind & PIPE_BIND_DEPTH_STENCIL) && + (dev->quirks & (MIDGARD_SFBD | IS_BIFROST))); +} - /* Based on the usage, figure out what storing will be used. There are - * various tradeoffs: - * - * Linear: the basic format, bad for memory bandwidth, bad for cache - * use. Zero-copy, though. Renderable. - * - * Tiled: Not compressed, but cache-optimized. Expensive to write into - * (due to software tiling), but cheap to sample from. Ideal for most - * textures. - * - * AFBC: Compressed and renderable (so always desirable for non-scanout - * rendertargets). Cheap to sample from. The format is black box, so we - * can't read/write from software. - */ +static bool +panfrost_should_afbc(struct panfrost_device *dev, const struct panfrost_resource *pres) +{ + /* AFBC resources may be rendered to, textured from, or shared across + * processes, but may not be used as e.g buffers */ + const unsigned valid_binding = + PIPE_BIND_DEPTH_STENCIL | + PIPE_BIND_RENDER_TARGET | + PIPE_BIND_BLENDABLE | + PIPE_BIND_SAMPLER_VIEW | + PIPE_BIND_DISPLAY_TARGET | + PIPE_BIND_SCANOUT | + PIPE_BIND_SHARED; + + if (pres->base.bind & ~valid_binding) + return false; - /* Tiling textures is almost always faster, unless we only use it once */ + /* AFBC introduced with Mali T760 */ + if (dev->quirks & MIDGARD_NO_AFBC) + return false; - bool is_texture = (res->bind & PIPE_BIND_SAMPLER_VIEW); - bool is_2d = res->depth0 == 1 && res->array_size == 1; - bool is_streaming = (res->usage != PIPE_USAGE_STREAM); + /* AFBC<-->staging is expensive */ + if (pres->base.usage == PIPE_USAGE_STREAM) + return false; - /* TODO: Reenable tiling on SFBD systems when we support rendering to - * tiled formats with SFBD */ - bool should_tile = is_streaming && is_texture && is_2d && !screen->require_sfbd; + /* Only a small selection of formats are AFBC'able */ + if (!panfrost_format_supports_afbc(pres->internal_format)) + return false; - /* Depth/stencil can't be tiled, only linear or AFBC */ - should_tile &= !(res->bind & PIPE_BIND_DEPTH_STENCIL); + /* AFBC does not support layered (GLES3 style) multisampling. Use + * EXT_multisampled_render_to_texture instead */ + if (pres->base.nr_samples > 1) + return false; + + /* TODO: Is AFBC of 3D textures possible? */ + if ((pres->base.target != PIPE_TEXTURE_2D) && (pres->base.target != PIPE_TEXTURE_RECT)) + return false; + + /* For one tile, AFBC is a loss compared to u-interleaved */ + if (pres->base.width0 <= 16 && pres->base.height0 <= 16) + return false; + + /* Otherwise, we'd prefer AFBC as it is dramatically more efficient + * than linear or usually even u-interleaved */ + return true; +} + +static bool +panfrost_should_tile(struct panfrost_device *dev, const struct panfrost_resource *pres) +{ + const unsigned valid_binding = + PIPE_BIND_DEPTH_STENCIL | + PIPE_BIND_RENDER_TARGET | + PIPE_BIND_BLENDABLE | + PIPE_BIND_SAMPLER_VIEW | + PIPE_BIND_DISPLAY_TARGET | + PIPE_BIND_SCANOUT | + PIPE_BIND_SHARED; + + unsigned bpp = util_format_get_blocksizebits(pres->internal_format); + + bool is_sane_bpp = + bpp == 8 || bpp == 16 || bpp == 24 || bpp == 32 || + bpp == 64 || bpp == 128; - /* FBOs we would like to checksum, if at all possible */ - bool can_checksum = !(res->bind & (PIPE_BIND_SCANOUT | PIPE_BIND_SHARED)); - bool should_checksum = res->bind & PIPE_BIND_RENDER_TARGET; + bool is_2d = (pres->base.target == PIPE_TEXTURE_2D) + || (pres->base.target == PIPE_TEXTURE_RECT); + + bool can_tile = is_2d && is_sane_bpp && ((pres->base.bind & ~valid_binding) == 0); + + if (!panfrost_can_linear(dev, pres)) { + assert(can_tile); + return true; + } + + return can_tile && (pres->base.usage != PIPE_USAGE_STREAM); +} + +static uint64_t +panfrost_best_modifier(struct panfrost_device *dev, + const struct panfrost_resource *pres) +{ + if (panfrost_should_afbc(dev, pres)) { + uint64_t afbc = + AFBC_FORMAT_MOD_BLOCK_SIZE_16x16 | + AFBC_FORMAT_MOD_SPARSE; + + if (panfrost_afbc_can_ytr(pres->base.format)) + afbc |= AFBC_FORMAT_MOD_YTR; + + return DRM_FORMAT_MOD_ARM_AFBC(afbc); + } else if (panfrost_should_tile(dev, pres)) + return DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED; + else + return DRM_FORMAT_MOD_LINEAR; +} + +static void +panfrost_resource_create_bo(struct panfrost_device *dev, struct panfrost_resource *pres, + uint64_t modifier) +{ + struct pipe_resource *res = &pres->base; - pres->checksummed = can_checksum && should_checksum; + pres->modifier = (modifier != DRM_FORMAT_MOD_INVALID) ? modifier : + panfrost_best_modifier(dev, pres); + pres->checksummed = (res->bind & PIPE_BIND_RENDER_TARGET); - /* Set the layout appropriately */ - pres->layout = should_tile ? PAN_TILED : PAN_LINEAR; + /* We can only switch tiled->linear if the resource isn't already + * linear, and if we control the modifier, and if the resource can be + * linear. */ + pres->modifier_constant = !((pres->modifier != DRM_FORMAT_MOD_LINEAR) + && (modifier == DRM_FORMAT_INVALID) + && panfrost_can_linear(dev, pres)); size_t bo_size; @@ -412,7 +546,7 @@ panfrost_resource_create_bo(struct panfrost_screen *screen, struct panfrost_reso /* We create a BO immediately but don't bother mapping, since we don't * care to map e.g. FBOs which the CPU probably won't touch */ - pres->bo = panfrost_bo_create(screen, bo_size, PAN_ALLOCATE_DELAY_MMAP); + pres->bo = panfrost_bo_create(dev, bo_size, PAN_BO_DELAY_MMAP); } void @@ -422,57 +556,29 @@ panfrost_resource_set_damage_region(struct pipe_screen *screen, const struct pipe_box *rects) { struct panfrost_resource *pres = pan_resource(res); - struct pipe_box *damage_rect = &pres->damage.biggest_rect; struct pipe_scissor_state *damage_extent = &pres->damage.extent; unsigned int i; - if (!nrects) { - panfrost_resource_reset_damage(pres); - return; - } - - /* We keep track of 2 different things here: - * 1 the damage extent: the quad including all damage regions. Will be - * used restrict the rendering area - * 2 the biggest damage rectangle: when there are more than one damage - * rect we keep the biggest one and will generate 4 wallpaper quads - * out of it (see panfrost_draw_wallpaper() for more details). We - * might want to do something smarter at some point. - * - * _________________________________ - * | | - * | _________________________ | - * | | rect1| _________| | - * | |______|_____ | rect 3: | | - * | | | rect2 | | biggest | | - * | | |_______| | rect | | - * | |_______________|_________| | - * | damage extent | - * |_______________________________| - * resource - */ + if (pres->damage.inverted_rects) + ralloc_free(pres->damage.inverted_rects); + memset(&pres->damage, 0, sizeof(pres->damage)); + + pres->damage.inverted_rects = + pan_subtract_damage(pres, + res->width0, res->height0, + nrects, rects, &pres->damage.inverted_len); + + /* Track the damage extent: the quad including all damage regions. Will + * be used restrict the rendering area */ + damage_extent->minx = 0xffff; damage_extent->miny = 0xffff; + for (i = 0; i < nrects; i++) { int x = rects[i].x, w = rects[i].width, h = rects[i].height; int y = res->height0 - (rects[i].y + h); - /* Clamp x,y,w,h to prevent negative values. */ - if (x < 0) { - h += x; - x = 0; - } - if (y < 0) { - w += y; - y = 0; - } - w = MAX2(w, 0); - h = MAX2(h, 0); - - if (damage_rect->width * damage_rect->height < w * h) - u_box_2d(x, y, w, h, damage_rect); - damage_extent->minx = MIN2(damage_extent->minx, x); damage_extent->miny = MIN2(damage_extent->miny, y); damage_extent->maxx = MAX2(damage_extent->maxx, @@ -480,12 +586,23 @@ panfrost_resource_set_damage_region(struct pipe_screen *screen, damage_extent->maxy = MAX2(damage_extent->maxy, MIN2(y + h, res->height0)); } + + if (nrects == 0) { + damage_extent->minx = 0; + damage_extent->miny = 0; + damage_extent->maxx = res->width0; + damage_extent->maxy = res->height0; + } + } static struct pipe_resource * -panfrost_resource_create(struct pipe_screen *screen, - const struct pipe_resource *template) +panfrost_resource_create_with_modifier(struct pipe_screen *screen, + const struct pipe_resource *template, + uint64_t modifier) { + struct panfrost_device *dev = pan_device(screen); + /* Make sure we're familiar */ switch (template->target) { case PIPE_BUFFER: @@ -494,50 +611,161 @@ panfrost_resource_create(struct pipe_screen *screen, case PIPE_TEXTURE_3D: case PIPE_TEXTURE_CUBE: case PIPE_TEXTURE_RECT: + case PIPE_TEXTURE_1D_ARRAY: case PIPE_TEXTURE_2D_ARRAY: break; default: - DBG("Unknown texture target %d\n", template->target); - assert(0); + unreachable("Unknown texture target\n"); } - if (template->bind & - (PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT | PIPE_BIND_SHARED)) - return panfrost_create_scanout_res(screen, template); + if (dev->ro && (template->bind & + (PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT | PIPE_BIND_SHARED))) + return panfrost_create_scanout_res(screen, template, modifier); struct panfrost_resource *so = rzalloc(screen, struct panfrost_resource); - struct panfrost_screen *pscreen = (struct panfrost_screen *) screen; - so->base = *template; so->base.screen = screen; + so->internal_format = template->format; pipe_reference_init(&so->base.reference, 1); util_range_init(&so->valid_buffer_range); - panfrost_resource_create_bo(pscreen, so); - panfrost_resource_reset_damage(so); + panfrost_resource_create_bo(dev, so, modifier); + panfrost_resource_set_damage_region(NULL, &so->base, 0, NULL); + + if (template->bind & PIPE_BIND_INDEX_BUFFER) + so->index_cache = rzalloc(so, struct panfrost_minmax_cache); return (struct pipe_resource *)so; } +/* Default is to create a resource as don't care */ + +static struct pipe_resource * +panfrost_resource_create(struct pipe_screen *screen, + const struct pipe_resource *template) +{ + return panfrost_resource_create_with_modifier(screen, template, + DRM_FORMAT_MOD_INVALID); +} + +/* If no modifier is specified, we'll choose. Otherwise, the order of + * preference is compressed, tiled, linear. */ + +static struct pipe_resource * +panfrost_resource_create_with_modifiers(struct pipe_screen *screen, + const struct pipe_resource *template, + const uint64_t *modifiers, int count) +{ + for (unsigned i = 0; i < PAN_MODIFIER_COUNT; ++i) { + if (drm_find_modifier(pan_best_modifiers[i], modifiers, count)) { + return panfrost_resource_create_with_modifier(screen, template, + pan_best_modifiers[i]); + } + } + + /* If we didn't find one, app specified invalid */ + assert(count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID); + return panfrost_resource_create(screen, template); +} + static void panfrost_resource_destroy(struct pipe_screen *screen, struct pipe_resource *pt) { - struct panfrost_screen *pscreen = pan_screen(screen); + struct panfrost_device *dev = pan_device(screen); struct panfrost_resource *rsrc = (struct panfrost_resource *) pt; if (rsrc->scanout) - renderonly_scanout_destroy(rsrc->scanout, pscreen->ro); + renderonly_scanout_destroy(rsrc->scanout, dev->ro); if (rsrc->bo) - panfrost_bo_unreference(screen, rsrc->bo); + panfrost_bo_unreference(rsrc->bo); + + if (rsrc->slices[0].checksum_bo) + panfrost_bo_unreference(rsrc->slices[0].checksum_bo); util_range_destroy(&rsrc->valid_buffer_range); ralloc_free(rsrc); } +/* Most of the time we can do CPU-side transfers, but sometimes we need to use + * the 3D pipe for this. Let's wrap u_blitter to blit to/from staging textures. + * Code adapted from freedreno */ + +static struct panfrost_resource * +pan_alloc_staging(struct panfrost_context *ctx, struct panfrost_resource *rsc, + unsigned level, const struct pipe_box *box) +{ + struct pipe_context *pctx = &ctx->base; + struct pipe_resource tmpl = rsc->base; + + tmpl.width0 = box->width; + tmpl.height0 = box->height; + /* for array textures, box->depth is the array_size, otherwise + * for 3d textures, it is the depth: + */ + if (tmpl.array_size > 1) { + if (tmpl.target == PIPE_TEXTURE_CUBE) + tmpl.target = PIPE_TEXTURE_2D_ARRAY; + tmpl.array_size = box->depth; + tmpl.depth0 = 1; + } else { + tmpl.array_size = 1; + tmpl.depth0 = box->depth; + } + tmpl.last_level = 0; + tmpl.bind |= PIPE_BIND_LINEAR; + + struct pipe_resource *pstaging = + pctx->screen->resource_create(pctx->screen, &tmpl); + if (!pstaging) + return NULL; + + return pan_resource(pstaging); +} + +static void +pan_blit_from_staging(struct pipe_context *pctx, struct panfrost_gtransfer *trans) +{ + struct pipe_resource *dst = trans->base.resource; + struct pipe_blit_info blit = {}; + + blit.dst.resource = dst; + blit.dst.format = dst->format; + blit.dst.level = trans->base.level; + blit.dst.box = trans->base.box; + blit.src.resource = trans->staging.rsrc; + blit.src.format = trans->staging.rsrc->format; + blit.src.level = 0; + blit.src.box = trans->staging.box; + blit.mask = util_format_get_mask(trans->staging.rsrc->format); + blit.filter = PIPE_TEX_FILTER_NEAREST; + + panfrost_blit(pctx, &blit); +} + +static void +pan_blit_to_staging(struct pipe_context *pctx, struct panfrost_gtransfer *trans) +{ + struct pipe_resource *src = trans->base.resource; + struct pipe_blit_info blit = {}; + + blit.src.resource = src; + blit.src.format = src->format; + blit.src.level = trans->base.level; + blit.src.box = trans->base.box; + blit.dst.resource = trans->staging.rsrc; + blit.dst.format = trans->staging.rsrc->format; + blit.dst.level = 0; + blit.dst.box = trans->staging.box; + blit.mask = util_format_get_mask(trans->staging.rsrc->format); + blit.filter = PIPE_TEX_FILTER_NEAREST; + + panfrost_blit(pctx, &blit); +} + static void * panfrost_transfer_map(struct pipe_context *pctx, struct pipe_resource *resource, @@ -546,106 +774,177 @@ panfrost_transfer_map(struct pipe_context *pctx, const struct pipe_box *box, struct pipe_transfer **out_transfer) { - int bytes_per_pixel = util_format_get_blocksize(resource->format); + struct panfrost_context *ctx = pan_context(pctx); + struct panfrost_device *dev = pan_device(pctx->screen); struct panfrost_resource *rsrc = pan_resource(resource); + int bytes_per_pixel = util_format_get_blocksize(rsrc->internal_format); struct panfrost_bo *bo = rsrc->bo; + /* Can't map tiled/compressed directly */ + if ((usage & PIPE_TRANSFER_MAP_DIRECTLY) && rsrc->modifier != DRM_FORMAT_MOD_LINEAR) + return NULL; + struct panfrost_gtransfer *transfer = rzalloc(pctx, struct panfrost_gtransfer); transfer->base.level = level; transfer->base.usage = usage; transfer->base.box = *box; pipe_resource_reference(&transfer->base.resource, resource); - *out_transfer = &transfer->base; - /* If we haven't already mmaped, now's the time */ + /* We don't have s/w routines for AFBC, so use a staging texture */ + if (drm_is_afbc(rsrc->modifier)) { + struct panfrost_resource *staging = pan_alloc_staging(ctx, rsrc, level, box); + transfer->base.stride = staging->slices[0].stride; + transfer->base.layer_stride = transfer->base.stride * box->height; - if (!bo->cpu) { - struct panfrost_screen *screen = pan_screen(pctx->screen); - panfrost_bo_mmap(screen, bo); - } + transfer->staging.rsrc = &staging->base; - /* Check if we're bound for rendering and this is a read pixels. If so, - * we need to flush */ + transfer->staging.box = *box; + transfer->staging.box.x = 0; + transfer->staging.box.y = 0; + transfer->staging.box.z = 0; - struct panfrost_context *ctx = pan_context(pctx); - struct pipe_framebuffer_state *fb = &ctx->pipe_framebuffer; + assert(transfer->staging.rsrc != NULL); - bool is_bound = false; + /* TODO: Eliminate this flush. It's only there to determine if + * we're initialized or not, when the initialization could come + * from a pending batch XXX */ + panfrost_flush_batches_accessing_bo(ctx, rsrc->bo, true); - for (unsigned c = 0; c < fb->nr_cbufs; ++c) { - /* If cbufs is NULL, we're definitely not bound here */ + if ((usage & PIPE_TRANSFER_READ) && rsrc->slices[level].initialized) { + pan_blit_to_staging(pctx, transfer); + panfrost_flush_batches_accessing_bo(ctx, staging->bo, true); + panfrost_bo_wait(staging->bo, INT64_MAX, false); + } - if (fb->cbufs[c]) - is_bound |= fb->cbufs[c]->texture == resource; + panfrost_bo_mmap(staging->bo); + return staging->bo->cpu; } - if (is_bound && (usage & PIPE_TRANSFER_READ)) { - assert(level == 0); - panfrost_flush(pctx, NULL, PIPE_FLUSH_END_OF_FRAME); - } + /* If we haven't already mmaped, now's the time */ + panfrost_bo_mmap(bo); + + if (dev->debug & (PAN_DBG_TRACE | PAN_DBG_SYNC)) + pandecode_inject_mmap(bo->gpu, bo->cpu, bo->size, NULL); + + bool create_new_bo = usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE; + bool copy_resource = false; - /* TODO: Respect usage flags */ + if (!create_new_bo && + !(usage & PIPE_TRANSFER_UNSYNCHRONIZED) && + (usage & PIPE_TRANSFER_WRITE) && + !(resource->target == PIPE_BUFFER + && !util_ranges_intersect(&rsrc->valid_buffer_range, box->x, box->x + box->width)) && + panfrost_pending_batches_access_bo(ctx, bo)) { - if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) { - /* TODO: reallocate */ - //printf("debug: Missed reallocate\n"); + /* When a resource to be modified is already being used by a + * pending batch, it is often faster to copy the whole BO than + * to flush and split the frame in two. This also mostly + * mitigates broken depth reload. + */ + + panfrost_flush_batches_accessing_bo(ctx, bo, false); + panfrost_bo_wait(bo, INT64_MAX, false); + + create_new_bo = true; + copy_resource = true; + } + + if (create_new_bo) { + /* If the BO is used by one of the pending batches or if it's + * not ready yet (still accessed by one of the already flushed + * batches), we try to allocate a new one to avoid waiting. + */ + if (panfrost_pending_batches_access_bo(ctx, bo) || + !panfrost_bo_wait(bo, 0, true)) { + /* We want the BO to be MMAPed. */ + uint32_t flags = bo->flags & ~PAN_BO_DELAY_MMAP; + struct panfrost_bo *newbo = NULL; + + /* When the BO has been imported/exported, we can't + * replace it by another one, otherwise the + * importer/exporter wouldn't see the change we're + * doing to it. + */ + if (!(bo->flags & PAN_BO_SHARED)) + newbo = panfrost_bo_create(dev, bo->size, + flags); + + if (newbo) { + if (copy_resource) + memcpy(newbo->cpu, rsrc->bo->cpu, bo->size); + + panfrost_bo_unreference(bo); + rsrc->bo = newbo; + bo = newbo; + } else { + /* Allocation failed or was impossible, let's + * fall back on a flush+wait. + */ + panfrost_flush_batches_accessing_bo(ctx, bo, true); + panfrost_bo_wait(bo, INT64_MAX, true); + } + } } else if ((usage & PIPE_TRANSFER_WRITE) && resource->target == PIPE_BUFFER && !util_ranges_intersect(&rsrc->valid_buffer_range, box->x, box->x + box->width)) { /* No flush for writes to uninitialized */ } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) { if (usage & PIPE_TRANSFER_WRITE) { - /* STUB: flush reading */ - //printf("debug: missed reading flush %d\n", resource->target); + panfrost_flush_batches_accessing_bo(ctx, bo, true); + panfrost_bo_wait(bo, INT64_MAX, true); } else if (usage & PIPE_TRANSFER_READ) { - /* STUB: flush writing */ - //printf("debug: missed writing flush %d (%d-%d)\n", resource->target, box->x, box->x + box->width); - } else { - /* Why are you even mapping?! */ + panfrost_flush_batches_accessing_bo(ctx, bo, false); + panfrost_bo_wait(bo, INT64_MAX, false); } } - if (rsrc->layout != PAN_LINEAR) { - /* Non-linear resources need to be indirectly mapped */ - - if (usage & PIPE_TRANSFER_MAP_DIRECTLY) - return NULL; - + if (rsrc->modifier == DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED) { transfer->base.stride = box->width * bytes_per_pixel; transfer->base.layer_stride = transfer->base.stride * box->height; - transfer->map = rzalloc_size(transfer, transfer->base.layer_stride * box->depth); + transfer->map = ralloc_size(transfer, transfer->base.layer_stride * box->depth); assert(box->depth == 1); if ((usage & PIPE_TRANSFER_READ) && rsrc->slices[level].initialized) { - if (rsrc->layout == PAN_AFBC) { - DBG("Unimplemented: reads from AFBC"); - } else if (rsrc->layout == PAN_TILED) { - panfrost_load_tiled_image( + panfrost_load_tiled_image( transfer->map, bo->cpu + rsrc->slices[level].offset, - box, + box->x, box->y, box->width, box->height, transfer->base.stride, rsrc->slices[level].stride, - util_format_get_blocksize(resource->format)); - } + rsrc->internal_format); } return transfer->map; } else { + assert (rsrc->modifier == DRM_FORMAT_MOD_LINEAR); + + /* Direct, persistent writes create holes in time for + * caching... I don't know if this is actually possible but we + * should still get it right */ + + unsigned dpw = PIPE_TRANSFER_MAP_DIRECTLY | PIPE_TRANSFER_WRITE | PIPE_TRANSFER_PERSISTENT; + + if ((usage & dpw) == dpw && rsrc->index_cache) + return NULL; + transfer->base.stride = rsrc->slices[level].stride; - transfer->base.layer_stride = rsrc->cubemap_stride; + transfer->base.layer_stride = panfrost_get_layer_stride( + rsrc->slices, rsrc->base.target == PIPE_TEXTURE_3D, + rsrc->cubemap_stride, level); /* By mapping direct-write, we're implicitly already * initialized (maybe), so be conservative */ - if ((usage & PIPE_TRANSFER_WRITE) && (usage & PIPE_TRANSFER_MAP_DIRECTLY)) + if (usage & PIPE_TRANSFER_WRITE) { rsrc->slices[level].initialized = true; + panfrost_minmax_cache_invalidate(rsrc->index_cache, &transfer->base); + } return bo->cpu + rsrc->slices[level].offset - + transfer->base.box.z * rsrc->cubemap_stride + + transfer->base.box.z * transfer->base.layer_stride + transfer->base.box.y * rsrc->slices[level].stride + transfer->base.box.x * bytes_per_pixel; } @@ -660,35 +959,83 @@ panfrost_transfer_unmap(struct pipe_context *pctx, struct panfrost_gtransfer *trans = pan_transfer(transfer); struct panfrost_resource *prsrc = (struct panfrost_resource *) transfer->resource; - /* Mark whatever we wrote as written */ - if (transfer->usage & PIPE_TRANSFER_WRITE) - prsrc->slices[transfer->level].initialized = true; + /* AFBC will use a staging resource. `initialized` will be set when the + * fragment job is created; this is deferred to prevent useless surface + * reloads that can cascade into DATA_INVALID_FAULTs due to reading + * malformed AFBC data if uninitialized */ + if (trans->staging.rsrc) { + if (transfer->usage & PIPE_TRANSFER_WRITE) { + pan_blit_from_staging(pctx, trans); + panfrost_flush_batches_accessing_bo(pan_context(pctx), pan_resource(trans->staging.rsrc)->bo, true); + } + + pipe_resource_reference(&trans->staging.rsrc, NULL); + } + + /* Tiling will occur in software from a staging cpu buffer */ if (trans->map) { struct panfrost_bo *bo = prsrc->bo; if (transfer->usage & PIPE_TRANSFER_WRITE) { - if (prsrc->layout == PAN_AFBC) { - DBG("Unimplemented: writes to AFBC\n"); - } else if (prsrc->layout == PAN_TILED) { + prsrc->slices[transfer->level].initialized = true; + + if (prsrc->modifier == DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED) { assert(transfer->box.depth == 1); - panfrost_store_tiled_image( - bo->cpu + prsrc->slices[transfer->level].offset, - trans->map, - &transfer->box, - prsrc->slices[transfer->level].stride, - transfer->stride, - util_format_get_blocksize(prsrc->base.format)); + /* Do we overwrite the entire resource? If so, + * we don't need an intermediate blit so it's a + * good time to switch the modifier. */ + + bool discards_content = prsrc->base.last_level == 0 + && transfer->box.width == prsrc->base.width0 + && transfer->box.height == prsrc->base.height0 + && transfer->box.x == 0 + && transfer->box.y == 0 + && !prsrc->modifier_constant; + + /* It also serves as a good heuristic for + * streaming textures (e.g. in video players), + * but we could do better */ + + if (discards_content) + ++prsrc->modifier_updates; + + if (prsrc->modifier_updates >= LAYOUT_CONVERT_THRESHOLD) + { + prsrc->modifier = DRM_FORMAT_MOD_LINEAR; + + util_copy_rect( + bo->cpu + prsrc->slices[0].offset, + prsrc->base.format, + prsrc->slices[0].stride, + 0, 0, + transfer->box.width, + transfer->box.height, + trans->map, + transfer->stride, + 0, 0); + } else { + panfrost_store_tiled_image( + bo->cpu + prsrc->slices[transfer->level].offset, + trans->map, + transfer->box.x, transfer->box.y, + transfer->box.width, transfer->box.height, + prsrc->slices[transfer->level].stride, + transfer->stride, + prsrc->internal_format); + } } } } - util_range_add(&prsrc->valid_buffer_range, + util_range_add(&prsrc->base, &prsrc->valid_buffer_range, transfer->box.x, transfer->box.x + transfer->box.width); + panfrost_minmax_cache_invalidate(prsrc->index_cache, transfer); + /* Derefence the resource */ pipe_resource_reference(&transfer->resource, NULL); @@ -704,7 +1051,7 @@ panfrost_transfer_flush_region(struct pipe_context *pctx, struct panfrost_resource *rsc = pan_resource(transfer->resource); if (transfer->resource->target == PIPE_BUFFER) { - util_range_add(&rsc->valid_buffer_range, + util_range_add(&rsc->base, &rsc->valid_buffer_range, transfer->box.x + box->x, transfer->box.x + box->x + box->width); } else { @@ -716,12 +1063,14 @@ panfrost_transfer_flush_region(struct pipe_context *pctx, static void panfrost_invalidate_resource(struct pipe_context *pctx, struct pipe_resource *prsc) { - //DBG("TODO %s\n", __func__); + /* TODO */ } static enum pipe_format -panfrost_resource_get_internal_format(struct pipe_resource *prsrc) { - return prsrc->format; +panfrost_resource_get_internal_format(struct pipe_resource *rsrc) +{ + struct panfrost_resource *prsrc = (struct panfrost_resource *) rsrc; + return prsrc->internal_format; } static bool @@ -734,7 +1083,6 @@ panfrost_generate_mipmap( unsigned first_layer, unsigned last_layer) { - struct panfrost_context *ctx = pan_context(pctx); struct panfrost_resource *rsrc = pan_resource(prsrc); /* Generating a mipmap invalidates the written levels, so make that @@ -745,19 +1093,7 @@ panfrost_generate_mipmap( for (unsigned l = base_level + 1; l <= last_level; ++l) rsrc->slices[l].initialized = false; - /* Beyond that, we just delegate the hard stuff. We're careful to - * include flushes on both ends to make sure the data is really valid. - * We could be doing a lot better perf-wise, especially once we have - * reorder-type optimizations in place. But for now prioritize - * correctness. */ - - struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx); - bool has_draws = batch->last_job.gpu; - - if (has_draws) - panfrost_flush(pctx, NULL, PIPE_FLUSH_END_OF_FRAME); - - /* We've flushed the original buffer if needed, now trigger a blit */ + /* Beyond that, we just delegate the hard stuff. */ bool blit_res = util_gen_mipmap( pctx, prsrc, format, @@ -765,12 +1101,6 @@ panfrost_generate_mipmap( first_layer, last_layer, PIPE_TEX_FILTER_LINEAR); - /* If the blit was successful, flush once more. If it wasn't, well, let - * the state tracker deal with it. */ - - if (blit_res) - panfrost_flush(pctx, NULL, PIPE_FLUSH_END_OF_FRAME); - return blit_res; } @@ -779,72 +1109,10 @@ panfrost_generate_mipmap( mali_ptr panfrost_get_texture_address( struct panfrost_resource *rsrc, - unsigned level, unsigned face) -{ - unsigned level_offset = rsrc->slices[level].offset; - unsigned face_offset = face * rsrc->cubemap_stride; - - return rsrc->bo->gpu + level_offset + face_offset; -} - -/* Given a resource that has already been allocated, hint that it should use a - * given layout. These are suggestions, not commands; it is perfectly legal to - * stub out this function, but there will be performance implications. */ - -void -panfrost_resource_hint_layout( - struct panfrost_screen *screen, - struct panfrost_resource *rsrc, - enum panfrost_memory_layout layout, - signed weight) + unsigned level, unsigned face, unsigned sample) { - /* Nothing to do, although a sophisticated implementation might store - * the hint */ - - if (rsrc->layout == layout) - return; - - /* We don't use the weight yet, but we should check that it's positive - * (semantically meaning that we should choose the given `layout`) */ - - if (weight <= 0) - return; - - /* Check if the preferred layout is legal for this buffer */ - - if (layout == PAN_AFBC) { - bool can_afbc = panfrost_format_supports_afbc(rsrc->base.format); - bool is_scanout = rsrc->base.bind & - (PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT | PIPE_BIND_SHARED); - - if (!can_afbc || is_scanout) - return; - } - - /* Simple heuristic so far: if the resource is uninitialized, switch to - * the hinted layout. If it is initialized, keep the original layout. - * This misses some cases where it would be beneficial to switch and - * blit. */ - - bool is_initialized = false; - - for (unsigned i = 0; i < MAX_MIP_LEVELS; ++i) - is_initialized |= rsrc->slices[i].initialized; - - if (is_initialized) - return; - - /* We're uninitialized, so do a layout switch. Reinitialize slices. */ - - size_t new_size; - rsrc->layout = layout; - panfrost_setup_slices(rsrc, &new_size); - - /* If we grew in size, reallocate the BO */ - if (new_size > rsrc->bo->size) { - panfrost_bo_release(screen, rsrc->bo, true); - rsrc->bo = panfrost_bo_create(screen, new_size, PAN_ALLOCATE_DELAY_MMAP); - } + bool is_3d = rsrc->base.target == PIPE_TEXTURE_3D; + return rsrc->bo->gpu + panfrost_texture_offset(rsrc->slices, is_3d, rsrc->cubemap_stride, level, face, sample); } static void @@ -872,17 +1140,21 @@ static const struct u_transfer_vtbl transfer_vtbl = { }; void -panfrost_resource_screen_init(struct panfrost_screen *pscreen) +panfrost_resource_screen_init(struct pipe_screen *pscreen) { - //pscreen->base.resource_create_with_modifiers = - // panfrost_resource_create_with_modifiers; - pscreen->base.resource_create = u_transfer_helper_resource_create; - pscreen->base.resource_destroy = u_transfer_helper_resource_destroy; - pscreen->base.resource_from_handle = panfrost_resource_from_handle; - pscreen->base.resource_get_handle = panfrost_resource_get_handle; - pscreen->base.transfer_helper = u_transfer_helper_create(&transfer_vtbl, + struct panfrost_device *dev = pan_device(pscreen); + + bool fake_rgtc = !panfrost_supports_compressed_format(dev, MALI_BC4_UNORM); + + pscreen->resource_create_with_modifiers = + panfrost_resource_create_with_modifiers; + pscreen->resource_create = u_transfer_helper_resource_create; + pscreen->resource_destroy = u_transfer_helper_resource_destroy; + pscreen->resource_from_handle = panfrost_resource_from_handle; + pscreen->resource_get_handle = panfrost_resource_get_handle; + pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl, true, false, - true, true); + fake_rgtc, true); } void