X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Ffreedreno%2Ffreedreno_resource.c;h=0f48e5878ffa315a3e2c763e6da74a87dacabdb0;hb=f9f5d3eb554e6096b762744125580a239d3809b0;hp=c2c0271e89a1fdb4199c39ec236cb1e5bc65533f;hpb=d1465b3aeee318a3a76a652325aa13d4a1946f37;p=mesa.git diff --git a/src/gallium/drivers/freedreno/freedreno_resource.c b/src/gallium/drivers/freedreno/freedreno_resource.c index c2c0271e89a..0f48e5878ff 100644 --- a/src/gallium/drivers/freedreno/freedreno_resource.c +++ b/src/gallium/drivers/freedreno/freedreno_resource.c @@ -1,5 +1,3 @@ -/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */ - /* * Copyright (C) 2012 Rob Clark * @@ -26,28 +24,40 @@ * Rob Clark */ -#include "util/u_format.h" -#include "util/u_format_rgtc.h" -#include "util/u_format_zs.h" +#include "util/format/u_format.h" +#include "util/format/u_format_rgtc.h" +#include "util/format/u_format_zs.h" #include "util/u_inlines.h" #include "util/u_transfer.h" #include "util/u_string.h" #include "util/u_surface.h" #include "util/set.h" +#include "util/u_drm.h" #include "freedreno_resource.h" #include "freedreno_batch_cache.h" +#include "freedreno_blitter.h" +#include "freedreno_fence.h" #include "freedreno_screen.h" #include "freedreno_surface.h" #include "freedreno_context.h" #include "freedreno_query_hw.h" #include "freedreno_util.h" +#include "drm-uapi/drm_fourcc.h" #include /* XXX this should go away, needed for 'struct winsys_handle' */ #include "state_tracker/drm_driver.h" +/* A private modifier for now, so we have a way to request tiled but not + * compressed. It would perhaps be good to get real modifiers for the + * tiled formats, but would probably need to do some work to figure out + * the layout(s) of the tiled modes, and whether they are the same + * across generations. + */ +#define FD_FORMAT_MOD_QCOM_TILED fourcc_mod_code(QCOM, 0xffffffff) + /** * Go through the entire state and see if the resource is bound * anywhere. If it is, mark the relevant state as dirty. This is @@ -84,6 +94,15 @@ rebind_resource(struct fd_context *ctx, struct pipe_resource *prsc) ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_TEX; } + /* Images */ + const unsigned num_images = util_last_bit(ctx->shaderimg[stage].enabled_mask); + for (unsigned i = 0; i < num_images; i++) { + if (ctx->dirty_shader[stage] & FD_DIRTY_SHADER_IMAGE) + break; + if (ctx->shaderimg[stage].si[i].resource == prsc) + ctx->dirty_shader[stage] |= FD_DIRTY_SHADER_IMAGE; + } + /* SSBOs */ const unsigned num_ssbos = util_last_bit(ctx->shaderbuf[stage].enabled_mask); for (unsigned i = 0; i < num_ssbos; i++) { @@ -98,9 +117,12 @@ rebind_resource(struct fd_context *ctx, struct pipe_resource *prsc) static void realloc_bo(struct fd_resource *rsc, uint32_t size) { + struct pipe_resource *prsc = &rsc->base; struct fd_screen *screen = fd_screen(rsc->base.screen); uint32_t flags = DRM_FREEDRENO_GEM_CACHE_WCOMBINE | - DRM_FREEDRENO_GEM_TYPE_KMEM; /* TODO */ + DRM_FREEDRENO_GEM_TYPE_KMEM | + COND(prsc->bind & PIPE_BIND_SCANOUT, DRM_FREEDRENO_GEM_SCANOUT); + /* TODO other flags? */ /* if we start using things other than write-combine, * be sure to check for PIPE_RESOURCE_FLAG_MAP_COHERENT @@ -109,7 +131,9 @@ realloc_bo(struct fd_resource *rsc, uint32_t size) if (rsc->bo) fd_bo_del(rsc->bo); - rsc->bo = fd_bo_new(screen->dev, size, flags); + rsc->bo = fd_bo_new(screen->dev, size, flags, "%ux%ux%u@%u:%x", + prsc->width0, prsc->height0, prsc->depth0, rsc->layout.cpp, prsc->bind); + rsc->seqno = p_atomic_inc_return(&screen->rsc_seqno); util_range_set_empty(&rsc->valid_buffer_range); fd_bc_invalidate_resource(rsc, true); } @@ -117,24 +141,27 @@ realloc_bo(struct fd_resource *rsc, uint32_t size) static void do_blit(struct fd_context *ctx, const struct pipe_blit_info *blit, bool fallback) { + struct pipe_context *pctx = &ctx->base; + /* TODO size threshold too?? */ - if (!fallback) { - /* do blit on gpu: */ - fd_blitter_pipe_begin(ctx, false, true, FD_STAGE_BLIT); - util_blitter_blit(ctx->blitter, blit); - fd_blitter_pipe_end(ctx); - } else { + if (fallback || !fd_blit(pctx, blit)) { /* do blit on cpu: */ - util_resource_copy_region(&ctx->base, + util_resource_copy_region(pctx, blit->dst.resource, blit->dst.level, blit->dst.box.x, blit->dst.box.y, blit->dst.box.z, blit->src.resource, blit->src.level, &blit->src.box); } } +/** + * @rsc: the resource to shadow + * @level: the level to discard (if box != NULL, otherwise ignored) + * @box: the box to discard (or NULL if none) + * @modifier: the modifier for the new buffer state + */ static bool fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc, - unsigned level, const struct pipe_box *box) + unsigned level, const struct pipe_box *box, uint64_t modifier) { struct pipe_context *pctx = &ctx->base; struct pipe_resource *prsc = &rsc->base; @@ -148,6 +175,7 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc, */ if (!pctx->screen->is_format_supported(pctx->screen, prsc->format, prsc->target, prsc->nr_samples, + prsc->nr_storage_samples, PIPE_BIND_RENDER_TARGET)) fallback = true; @@ -155,15 +183,16 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc, if (prsc->target == PIPE_BUFFER) fallback = true; - bool whole_level = util_texrange_covers_whole_level(prsc, level, + bool discard_whole_level = box && util_texrange_covers_whole_level(prsc, level, box->x, box->y, box->z, box->width, box->height, box->depth); /* TODO need to be more clever about current level */ - if ((prsc->target >= PIPE_TEXTURE_2D) && !whole_level) + if ((prsc->target >= PIPE_TEXTURE_2D) && box && !discard_whole_level) return false; struct pipe_resource *pshadow = - pctx->screen->resource_create(pctx->screen, prsc); + pctx->screen->resource_create_with_modifiers(pctx->screen, + prsc, &modifier, 1); if (!pshadow) return false; @@ -194,6 +223,12 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc, /* TODO valid_buffer_range?? */ swap(rsc->bo, shadow->bo); swap(rsc->write_batch, shadow->write_batch); + for (int level = 0; level <= prsc->last_level; level++) { + swap(rsc->layout.slices[level], shadow->layout.slices[level]); + swap(rsc->layout.ubwc_slices[level], shadow->layout.ubwc_slices[level]); + } + swap(rsc->layout.ubwc_layer_size, shadow->layout.ubwc_layer_size); + rsc->seqno = p_atomic_inc_return(&ctx->screen->rsc_seqno); /* at this point, the newly created shadow buffer is not referenced * by any batches, but the existing rsc (probably) is. We need to @@ -210,7 +245,7 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc, mtx_unlock(&ctx->screen->lock); - struct pipe_blit_info blit = {0}; + struct pipe_blit_info blit = {}; blit.dst.resource = prsc; blit.dst.format = prsc->format; blit.src.resource = pshadow; @@ -225,7 +260,7 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc, /* blit the other levels in their entirety: */ for (unsigned l = 0; l <= prsc->last_level; l++) { - if (l == level) + if (box && l == level) continue; /* just blit whole level: */ @@ -240,7 +275,7 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc, /* deal w/ current level specially, since we might need to split * it up into a couple blits: */ - if (!whole_level) { + if (box && !discard_whole_level) { set_box(level, level); switch (prsc->target) { @@ -278,15 +313,104 @@ fd_try_shadow_resource(struct fd_context *ctx, struct fd_resource *rsc, return true; } -static unsigned -fd_resource_layer_offset(struct fd_resource *rsc, - struct fd_resource_slice *slice, - unsigned layer) +/** + * Uncompress an UBWC compressed buffer "in place". This works basically + * like resource shadowing, creating a new resource, and doing an uncompress + * blit, and swapping the state between shadow and original resource so it + * appears to the state tracker as if nothing changed. + */ +void +fd_resource_uncompress(struct fd_context *ctx, struct fd_resource *rsc) { - if (rsc->layer_first) - return layer * rsc->layer_size; - else - return layer * slice->size0; + bool success = + fd_try_shadow_resource(ctx, rsc, 0, NULL, FD_FORMAT_MOD_QCOM_TILED); + + /* shadow should not fail in any cases where we need to uncompress: */ + debug_assert(success); + + /* + * TODO what if rsc is used in other contexts, we don't currently + * have a good way to rebind_resource() in other contexts. And an + * app that is reading one resource in multiple contexts, isn't + * going to expect that the resource is modified. + * + * Hopefully the edge cases where we need to uncompress are rare + * enough that they mostly only show up in deqp. + */ + + rebind_resource(ctx, &rsc->base); +} + +static struct fd_resource * +fd_alloc_staging(struct fd_context *ctx, struct fd_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 fd_resource(pstaging); +} + +static void +fd_blit_from_staging(struct fd_context *ctx, struct fd_transfer *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_prsc; + blit.src.format = trans->staging_prsc->format; + blit.src.level = 0; + blit.src.box = trans->staging_box; + blit.mask = util_format_get_mask(trans->staging_prsc->format); + blit.filter = PIPE_TEX_FILTER_NEAREST; + + do_blit(ctx, &blit, false); +} + +static void +fd_blit_to_staging(struct fd_context *ctx, struct fd_transfer *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_prsc; + blit.dst.format = trans->staging_prsc->format; + blit.dst.level = 0; + blit.dst.box = trans->staging_box; + blit.mask = util_format_get_mask(trans->staging_prsc->format); + blit.filter = PIPE_TEX_FILTER_NEAREST; + + do_blit(ctx, &blit, false); } static void fd_resource_transfer_flush_region(struct pipe_context *pctx, @@ -296,23 +420,76 @@ static void fd_resource_transfer_flush_region(struct pipe_context *pctx, struct fd_resource *rsc = fd_resource(ptrans->resource); if (ptrans->resource->target == PIPE_BUFFER) - util_range_add(&rsc->valid_buffer_range, + util_range_add(&rsc->base, &rsc->valid_buffer_range, ptrans->box.x + box->x, ptrans->box.x + box->x + box->width); } +static void +flush_resource(struct fd_context *ctx, struct fd_resource *rsc, unsigned usage) +{ + struct fd_batch *write_batch = NULL; + + mtx_lock(&ctx->screen->lock); + fd_batch_reference_locked(&write_batch, rsc->write_batch); + mtx_unlock(&ctx->screen->lock); + + if (usage & PIPE_TRANSFER_WRITE) { + struct fd_batch *batch, *batches[32] = {}; + uint32_t batch_mask; + + /* This is a bit awkward, probably a fd_batch_flush_locked() + * would make things simpler.. but we need to hold the lock + * to iterate the batches which reference this resource. So + * we must first grab references under a lock, then flush. + */ + mtx_lock(&ctx->screen->lock); + batch_mask = rsc->batch_mask; + foreach_batch(batch, &ctx->screen->batch_cache, batch_mask) + fd_batch_reference_locked(&batches[batch->idx], batch); + mtx_unlock(&ctx->screen->lock); + + foreach_batch(batch, &ctx->screen->batch_cache, batch_mask) + fd_batch_flush(batch); + + foreach_batch(batch, &ctx->screen->batch_cache, batch_mask) { + fd_batch_reference(&batches[batch->idx], NULL); + } + assert(rsc->batch_mask == 0); + } else if (write_batch) { + fd_batch_flush(write_batch); + } + + fd_batch_reference(&write_batch, NULL); + + assert(!rsc->write_batch); +} + +static void +fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc) +{ + flush_resource(fd_context(pctx), fd_resource(prsc), PIPE_TRANSFER_READ); +} + static void fd_resource_transfer_unmap(struct pipe_context *pctx, struct pipe_transfer *ptrans) { struct fd_context *ctx = fd_context(pctx); struct fd_resource *rsc = fd_resource(ptrans->resource); + struct fd_transfer *trans = fd_transfer(ptrans); + + if (trans->staging_prsc) { + if (ptrans->usage & PIPE_TRANSFER_WRITE) + fd_blit_from_staging(ctx, trans); + pipe_resource_reference(&trans->staging_prsc, NULL); + } if (!(ptrans->usage & PIPE_TRANSFER_UNSYNCHRONIZED)) { fd_bo_cpu_fini(rsc->bo); } - util_range_add(&rsc->valid_buffer_range, + util_range_add(&rsc->base, &rsc->valid_buffer_range, ptrans->box.x, ptrans->box.x + ptrans->box.width); @@ -329,7 +506,7 @@ fd_resource_transfer_map(struct pipe_context *pctx, { struct fd_context *ctx = fd_context(pctx); struct fd_resource *rsc = fd_resource(prsc); - struct fd_resource_slice *slice = fd_resource_slice(rsc, level); + struct fdl_slice *slice = fd_resource_slice(rsc, level); struct fd_transfer *trans; struct pipe_transfer *ptrans; enum pipe_format format = prsc->format; @@ -353,8 +530,49 @@ fd_resource_transfer_map(struct pipe_context *pctx, ptrans->level = level; ptrans->usage = usage; ptrans->box = *box; - ptrans->stride = util_format_get_nblocksx(format, slice->pitch) * rsc->cpp; - ptrans->layer_stride = rsc->layer_first ? rsc->layer_size : slice->size0; + ptrans->stride = util_format_get_nblocksx(format, slice->pitch) * rsc->layout.cpp; + ptrans->layer_stride = fd_resource_layer_stride(rsc, level); + + /* we always need a staging texture for tiled buffers: + * + * TODO we might sometimes want to *also* shadow the resource to avoid + * splitting a batch.. for ex, mid-frame texture uploads to a tiled + * texture. + */ + if (rsc->layout.tile_mode) { + struct fd_resource *staging_rsc; + + staging_rsc = fd_alloc_staging(ctx, rsc, level, box); + if (staging_rsc) { + struct fdl_slice *staging_slice = + fd_resource_slice(staging_rsc, 0); + // TODO for PIPE_TRANSFER_READ, need to do untiling blit.. + trans->staging_prsc = &staging_rsc->base; + trans->base.stride = util_format_get_nblocksx(format, + staging_slice->pitch) * staging_rsc->layout.cpp; + trans->base.layer_stride = fd_resource_layer_stride(staging_rsc, 0); + trans->staging_box = *box; + trans->staging_box.x = 0; + trans->staging_box.y = 0; + trans->staging_box.z = 0; + + if (usage & PIPE_TRANSFER_READ) { + fd_blit_to_staging(ctx, trans); + + fd_bo_cpu_prep(staging_rsc->bo, ctx->pipe, + DRM_FREEDRENO_PREP_READ); + } + + buf = fd_bo_map(staging_rsc->bo); + offset = 0; + + *pptrans = ptrans; + + ctx->stats.staging_uploads++; + + return buf; + } + } if (ctx->in_shadow && !(usage & PIPE_TRANSFER_READ)) usage |= PIPE_TRANSFER_UNSYNCHRONIZED; @@ -379,7 +597,9 @@ fd_resource_transfer_map(struct pipe_context *pctx, struct fd_batch *write_batch = NULL; /* hold a reference, so it doesn't disappear under us: */ - fd_batch_reference(&write_batch, rsc->write_batch); + fd_context_lock(ctx); + fd_batch_reference_locked(&write_batch, rsc->write_batch); + fd_context_unlock(ctx); if ((usage & PIPE_TRANSFER_WRITE) && write_batch && write_batch->back_blit) { @@ -403,40 +623,57 @@ fd_resource_transfer_map(struct pipe_context *pctx, */ if (ctx->screen->reorder && busy && !(usage & PIPE_TRANSFER_READ) && (usage & PIPE_TRANSFER_DISCARD_RANGE)) { - if (fd_try_shadow_resource(ctx, rsc, level, box)) { + /* try shadowing only if it avoids a flush, otherwise staging would + * be better: + */ + if (needs_flush && fd_try_shadow_resource(ctx, rsc, level, + box, DRM_FORMAT_MOD_LINEAR)) { needs_flush = busy = false; rebind_resource(ctx, prsc); - } - } + ctx->stats.shadow_uploads++; + } else { + struct fd_resource *staging_rsc; - if (needs_flush) { - if (usage & PIPE_TRANSFER_WRITE) { - struct fd_batch *batch, *batches[32] = {0}; - uint32_t batch_mask; - - /* This is a bit awkward, probably a fd_batch_flush_locked() - * would make things simpler.. but we need to hold the lock - * to iterate the batches which reference this resource. So - * we must first grab references under a lock, then flush. + if (needs_flush) { + flush_resource(ctx, rsc, usage); + needs_flush = false; + } + + /* in this case, we don't need to shadow the whole resource, + * since any draw that references the previous contents has + * already had rendering flushed for all tiles. So we can + * use a staging buffer to do the upload. */ - mtx_lock(&ctx->screen->lock); - batch_mask = rsc->batch_mask; - foreach_batch(batch, &ctx->screen->batch_cache, batch_mask) - fd_batch_reference(&batches[batch->idx], batch); - mtx_unlock(&ctx->screen->lock); - - foreach_batch(batch, &ctx->screen->batch_cache, batch_mask) - fd_batch_flush(batch, false, false); - - foreach_batch(batch, &ctx->screen->batch_cache, batch_mask) { - fd_batch_sync(batch); - fd_batch_reference(&batches[batch->idx], NULL); + staging_rsc = fd_alloc_staging(ctx, rsc, level, box); + if (staging_rsc) { + struct fdl_slice *staging_slice = + fd_resource_slice(staging_rsc, 0); + trans->staging_prsc = &staging_rsc->base; + trans->base.stride = util_format_get_nblocksx(format, + staging_slice->pitch) * staging_rsc->layout.cpp; + trans->base.layer_stride = + fd_resource_layer_stride(staging_rsc, 0); + trans->staging_box = *box; + trans->staging_box.x = 0; + trans->staging_box.y = 0; + trans->staging_box.z = 0; + buf = fd_bo_map(staging_rsc->bo); + offset = 0; + + *pptrans = ptrans; + + fd_batch_reference(&write_batch, NULL); + + ctx->stats.staging_uploads++; + + return buf; } - assert(rsc->batch_mask == 0); - } else { - fd_batch_flush(write_batch, true, false); } - assert(!rsc->write_batch); + } + + if (needs_flush) { + flush_resource(ctx, rsc, usage); + needs_flush = false; } fd_batch_reference(&write_batch, NULL); @@ -453,13 +690,13 @@ fd_resource_transfer_map(struct pipe_context *pctx, } buf = fd_bo_map(rsc->bo); - if (!buf) - goto fail; - - offset = slice->offset + + offset = box->y / util_format_get_blockheight(format) * ptrans->stride + - box->x / util_format_get_blockwidth(format) * rsc->cpp + - fd_resource_layer_offset(rsc, slice, box->z); + box->x / util_format_get_blockwidth(format) * rsc->layout.cpp + + fd_resource_offset(rsc, level, box->z); + + if (usage & PIPE_TRANSFER_WRITE) + rsc->valid = true; *pptrans = ptrans; @@ -478,11 +715,27 @@ fd_resource_destroy(struct pipe_screen *pscreen, fd_bc_invalidate_resource(rsc, true); if (rsc->bo) fd_bo_del(rsc->bo); + if (rsc->scanout) + renderonly_scanout_destroy(rsc->scanout, fd_screen(pscreen)->ro); + util_range_destroy(&rsc->valid_buffer_range); FREE(rsc); } -static boolean +static uint64_t +fd_resource_modifier(struct fd_resource *rsc) +{ + if (!rsc->layout.tile_mode) + return DRM_FORMAT_MOD_LINEAR; + + if (rsc->layout.ubwc_layer_size) + return DRM_FORMAT_MOD_QCOM_COMPRESSED; + + /* TODO invent a modifier for tiled but not UBWC buffers: */ + return DRM_FORMAT_MOD_INVALID; +} + +static bool fd_resource_get_handle(struct pipe_screen *pscreen, struct pipe_context *pctx, struct pipe_resource *prsc, @@ -491,8 +744,10 @@ fd_resource_get_handle(struct pipe_screen *pscreen, { struct fd_resource *rsc = fd_resource(prsc); - return fd_screen_bo_get_handle(pscreen, rsc->bo, - rsc->slices[0].pitch * rsc->cpp, handle); + handle->modifier = fd_resource_modifier(rsc); + + return fd_screen_bo_get_handle(pscreen, rsc->bo, rsc->scanout, + fd_resource_slice(rsc, 0)->pitch * rsc->layout.cpp, handle); } static uint32_t @@ -509,13 +764,10 @@ setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format forma /* in layer_first layout, the level (slice) contains just one * layer (since in fact the layer contains the slices) */ - uint32_t layers_in_level = rsc->layer_first ? 1 : prsc->array_size; - - if (is_a5xx(screen) && (rsc->base.target >= PIPE_TEXTURE_2D)) - height = align(height, screen->gmem_alignh); + uint32_t layers_in_level = rsc->layout.layer_first ? 1 : prsc->array_size; for (level = 0; level <= prsc->last_level; level++) { - struct fd_resource_slice *slice = fd_resource_slice(rsc, level); + struct fdl_slice *slice = fd_resource_slice(rsc, level); uint32_t blocks; if (layout == UTIL_FORMAT_LAYOUT_ASTC) @@ -533,12 +785,12 @@ setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format forma */ if (prsc->target == PIPE_TEXTURE_3D && ( level == 1 || - (level > 1 && rsc->slices[level - 1].size0 > 0xf000))) - slice->size0 = align(blocks * rsc->cpp, alignment); - else if (level == 0 || rsc->layer_first || alignment == 1) - slice->size0 = align(blocks * rsc->cpp, alignment); + (level > 1 && fd_resource_slice(rsc, level - 1)->size0 > 0xf000))) + slice->size0 = align(blocks * rsc->layout.cpp, alignment); + else if (level == 0 || rsc->layout.layer_first || alignment == 1) + slice->size0 = align(blocks * rsc->layout.cpp, alignment); else - slice->size0 = rsc->slices[level - 1].size0; + slice->size0 = fd_resource_slice(rsc, level - 1)->size0; size += slice->size0 * depth * layers_in_level; @@ -551,12 +803,12 @@ setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format forma } static uint32_t -slice_alignment(struct pipe_screen *pscreen, const struct pipe_resource *tmpl) +slice_alignment(enum pipe_texture_target target) { /* on a3xx, 2d array and 3d textures seem to want their * layers aligned to page boundaries: */ - switch (tmpl->target) { + switch (target) { case PIPE_TEXTURE_3D: case PIPE_TEXTURE_1D_ARRAY: case PIPE_TEXTURE_2D_ARRAY: @@ -566,6 +818,36 @@ slice_alignment(struct pipe_screen *pscreen, const struct pipe_resource *tmpl) } } +/* cross generation texture layout to plug in to screen->setup_slices().. + * replace with generation specific one as-needed. + * + * TODO for a4xx probably can extract out the a4xx specific logic int + * a small fd4_setup_slices() wrapper that sets up layer_first, and then + * calls this. + */ +uint32_t +fd_setup_slices(struct fd_resource *rsc) +{ + uint32_t alignment; + + alignment = slice_alignment(rsc->base.target); + + struct fd_screen *screen = fd_screen(rsc->base.screen); + if (is_a4xx(screen)) { + switch (rsc->base.target) { + case PIPE_TEXTURE_3D: + rsc->layout.layer_first = false; + break; + default: + rsc->layout.layer_first = true; + alignment = 1; + break; + } + } + + return setup_slices(rsc, alignment, rsc->base.format); +} + /* special case to resize query buf after allocated.. */ void fd_resource_resize(struct pipe_resource *prsc, uint32_t sz) @@ -577,40 +859,71 @@ fd_resource_resize(struct pipe_resource *prsc, uint32_t sz) debug_assert(prsc->bind == PIPE_BIND_QUERY_BUFFER); prsc->width0 = sz; - realloc_bo(rsc, setup_slices(rsc, 1, prsc->format)); + realloc_bo(rsc, fd_screen(prsc->screen)->setup_slices(rsc)); } -// TODO common helper? -static bool -has_depth(enum pipe_format format) +static void +fd_resource_layout_init(struct pipe_resource *prsc) { - switch (format) { - case PIPE_FORMAT_Z16_UNORM: - case PIPE_FORMAT_Z32_UNORM: - case PIPE_FORMAT_Z32_FLOAT: - case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT: - case PIPE_FORMAT_Z24_UNORM_S8_UINT: - case PIPE_FORMAT_S8_UINT_Z24_UNORM: - case PIPE_FORMAT_Z24X8_UNORM: - case PIPE_FORMAT_X8Z24_UNORM: - return true; - default: - return false; - } + struct fd_resource *rsc = fd_resource(prsc); + struct fdl_layout *layout = &rsc->layout; + + layout->width0 = prsc->width0; + layout->height0 = prsc->height0; + layout->depth0 = prsc->depth0; + + layout->cpp = util_format_get_blocksize(prsc->format); + layout->cpp *= fd_resource_nr_samples(prsc); } /** * Create a new texture object, using the given template info. */ static struct pipe_resource * -fd_resource_create(struct pipe_screen *pscreen, - const struct pipe_resource *tmpl) +fd_resource_create_with_modifiers(struct pipe_screen *pscreen, + const struct pipe_resource *tmpl, + const uint64_t *modifiers, int count) { struct fd_screen *screen = fd_screen(pscreen); - struct fd_resource *rsc = CALLOC_STRUCT(fd_resource); - struct pipe_resource *prsc = &rsc->base; + struct fd_resource *rsc; + struct pipe_resource *prsc; enum pipe_format format = tmpl->format; - uint32_t size, alignment; + uint32_t size; + + /* when using kmsro, scanout buffers are allocated on the display device + * create_with_modifiers() doesn't give us usage flags, so we have to + * assume that all calls with modifiers are scanout-possible + */ + if (screen->ro && + ((tmpl->bind & PIPE_BIND_SCANOUT) || + !(count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID))) { + struct pipe_resource scanout_templat = *tmpl; + struct renderonly_scanout *scanout; + struct winsys_handle handle; + + /* apply freedreno alignment requirement */ + scanout_templat.width0 = align(tmpl->width0, screen->gmem_alignw); + + scanout = renderonly_scanout_for_resource(&scanout_templat, + screen->ro, &handle); + if (!scanout) + return NULL; + + renderonly_scanout_destroy(scanout, screen->ro); + + assert(handle.type == WINSYS_HANDLE_TYPE_FD); + rsc = fd_resource(pscreen->resource_from_handle(pscreen, tmpl, + &handle, + PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE)); + close(handle.handle); + if (!rsc) + return NULL; + + return &rsc->base; + } + + rsc = CALLOC_STRUCT(fd_resource); + prsc = &rsc->base; DBG("%p: target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, " "nr_samples=%u, usage=%u, bind=%x, flags=%x", prsc, @@ -623,48 +936,58 @@ fd_resource_create(struct pipe_screen *pscreen, return NULL; *prsc = *tmpl; + fd_resource_layout_init(prsc); - pipe_reference_init(&prsc->reference, 1); +#define LINEAR \ + (PIPE_BIND_SCANOUT | \ + PIPE_BIND_LINEAR | \ + PIPE_BIND_DISPLAY_TARGET) - prsc->screen = pscreen; + bool linear = drm_find_modifier(DRM_FORMAT_MOD_LINEAR, modifiers, count); + if (tmpl->bind & LINEAR) + linear = true; - util_range_init(&rsc->valid_buffer_range); + if (fd_mesa_debug & FD_DBG_NOTILE) + linear = true; - rsc->internal_format = format; - rsc->cpp = util_format_get_blocksize(format); + /* Normally, for non-shared buffers, allow buffer compression if + * not shared, otherwise only allow if QCOM_COMPRESSED modifier + * is requested: + * + * TODO we should probably also limit tiled in a similar way, + * except we don't have a format modifier for tiled. (We probably + * should.) + */ + bool allow_ubwc = drm_find_modifier(DRM_FORMAT_MOD_INVALID, modifiers, count); + if (tmpl->bind & PIPE_BIND_SHARED) + allow_ubwc = drm_find_modifier(DRM_FORMAT_MOD_QCOM_COMPRESSED, modifiers, count); - assert(rsc->cpp); + allow_ubwc &= !(fd_mesa_debug & FD_DBG_NOUBWC); - // XXX probably need some extra work if we hit rsc shadowing path w/ lrz.. - if (is_a5xx(screen) && (fd_mesa_debug & FD_DBG_LRZ) && has_depth(format)) { - const uint32_t flags = DRM_FREEDRENO_GEM_CACHE_WCOMBINE | - DRM_FREEDRENO_GEM_TYPE_KMEM; /* TODO */ - unsigned lrz_pitch = align(DIV_ROUND_UP(tmpl->width0, 8), 32); - unsigned lrz_height = DIV_ROUND_UP(tmpl->height0, 8); - unsigned size = lrz_pitch * lrz_height * 2; + pipe_reference_init(&prsc->reference, 1); - size += 0x1000; /* for GRAS_LRZ_FAST_CLEAR_BUFFER */ + prsc->screen = pscreen; - rsc->lrz_height = lrz_height; - rsc->lrz_width = lrz_pitch; - rsc->lrz_pitch = lrz_pitch; - rsc->lrz = fd_bo_new(screen->dev, size, flags); + if (screen->tile_mode && + (tmpl->target != PIPE_BUFFER) && + !linear) { + rsc->layout.tile_mode = screen->tile_mode(prsc); } - alignment = slice_alignment(pscreen, tmpl); - if (is_a4xx(screen) || is_a5xx(screen)) { - switch (tmpl->target) { - case PIPE_TEXTURE_3D: - rsc->layer_first = false; - break; - default: - rsc->layer_first = true; - alignment = 1; - break; - } + util_range_init(&rsc->valid_buffer_range); + + rsc->internal_format = format; + + if (prsc->target == PIPE_BUFFER) { + assert(prsc->format == PIPE_FORMAT_R8_UNORM); + size = prsc->width0; + fdl_layout_buffer(&rsc->layout, size); + } else { + size = screen->setup_slices(rsc); } - size = setup_slices(rsc, alignment, format); + if (allow_ubwc && screen->fill_ubwc_buffer_sizes && rsc->layout.tile_mode) + size += screen->fill_ubwc_buffer_sizes(rsc); /* special case for hw-query buffer, which we need to allocate before we * know the size: @@ -675,11 +998,15 @@ fd_resource_create(struct pipe_screen *pscreen, return prsc; } - if (rsc->layer_first) { - rsc->layer_size = align(size, 4096); - size = rsc->layer_size * prsc->array_size; + /* Set the layer size if the (non-a6xx) backend hasn't done so. */ + if (rsc->layout.layer_first && !rsc->layout.layer_size) { + rsc->layout.layer_size = align(size, 4096); + size = rsc->layout.layer_size * prsc->array_size; } + if (fd_mesa_debug & FD_DBG_LAYOUT) + fdl_dump_layout(&rsc->layout); + realloc_bo(rsc, size); if (!rsc->bo) goto fail; @@ -690,6 +1017,34 @@ fail: return NULL; } +static struct pipe_resource * +fd_resource_create(struct pipe_screen *pscreen, + const struct pipe_resource *tmpl) +{ + const uint64_t mod = DRM_FORMAT_MOD_INVALID; + return fd_resource_create_with_modifiers(pscreen, tmpl, &mod, 1); +} + +static bool +is_supported_modifier(struct pipe_screen *pscreen, enum pipe_format pfmt, + uint64_t mod) +{ + int count; + + /* Get the count of supported modifiers: */ + pscreen->query_dmabuf_modifiers(pscreen, pfmt, 0, NULL, NULL, &count); + + /* Get the supported modifiers: */ + uint64_t modifiers[count]; + pscreen->query_dmabuf_modifiers(pscreen, pfmt, count, modifiers, NULL, &count); + + for (int i = 0; i < count; i++) + if (modifiers[i] == mod) + return true; + + return false; +} + /** * Create a texture from a winsys_handle. The handle is often created in * another process by first creating a pipe texture and then calling @@ -700,8 +1055,9 @@ fd_resource_from_handle(struct pipe_screen *pscreen, const struct pipe_resource *tmpl, struct winsys_handle *handle, unsigned usage) { + struct fd_screen *screen = fd_screen(pscreen); struct fd_resource *rsc = CALLOC_STRUCT(fd_resource); - struct fd_resource_slice *slice = &rsc->slices[0]; + struct fdl_slice *slice = fd_resource_slice(rsc, 0); struct pipe_resource *prsc = &rsc->base; uint32_t pitchalign = fd_screen(pscreen)->gmem_alignw; @@ -716,6 +1072,7 @@ fd_resource_from_handle(struct pipe_screen *pscreen, return NULL; *prsc = *tmpl; + fd_resource_layout_init(prsc); pipe_reference_init(&prsc->reference, 1); @@ -727,8 +1084,8 @@ fd_resource_from_handle(struct pipe_screen *pscreen, if (!rsc->bo) goto fail; - rsc->cpp = util_format_get_blocksize(tmpl->format); - slice->pitch = handle->stride / rsc->cpp; + rsc->internal_format = tmpl->format; + slice->pitch = handle->stride / rsc->layout.cpp; slice->offset = handle->offset; slice->size0 = handle->stride * prsc->height0; @@ -736,7 +1093,28 @@ fd_resource_from_handle(struct pipe_screen *pscreen, (slice->pitch & (pitchalign - 1))) goto fail; - assert(rsc->cpp); + if (handle->modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED) { + if (!is_supported_modifier(pscreen, tmpl->format, + DRM_FORMAT_MOD_QCOM_COMPRESSED)) { + DBG("bad modifier: %"PRIx64, handle->modifier); + goto fail; + } + debug_assert(screen->fill_ubwc_buffer_sizes); + screen->fill_ubwc_buffer_sizes(rsc); + } else if (handle->modifier && + (handle->modifier != DRM_FORMAT_MOD_INVALID)) { + goto fail; + } + + assert(rsc->layout.cpp); + + if (screen->ro) { + rsc->scanout = + renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL); + /* failure is expected in some cases.. */ + } + + rsc->valid = true; return prsc; @@ -745,68 +1123,6 @@ fail: return NULL; } -/** - * _copy_region using pipe (3d engine) - */ -static bool -fd_blitter_pipe_copy_region(struct fd_context *ctx, - struct pipe_resource *dst, - unsigned dst_level, - unsigned dstx, unsigned dsty, unsigned dstz, - struct pipe_resource *src, - unsigned src_level, - const struct pipe_box *src_box) -{ - /* not until we allow rendertargets to be buffers */ - if (dst->target == PIPE_BUFFER || src->target == PIPE_BUFFER) - return false; - - if (!util_blitter_is_copy_supported(ctx->blitter, dst, src)) - return false; - - /* TODO we could discard if dst box covers dst level fully.. */ - fd_blitter_pipe_begin(ctx, false, false, FD_STAGE_BLIT); - util_blitter_copy_texture(ctx->blitter, - dst, dst_level, dstx, dsty, dstz, - src, src_level, src_box); - fd_blitter_pipe_end(ctx); - - return true; -} - -/** - * Copy a block of pixels from one resource to another. - * The resource must be of the same format. - * Resources with nr_samples > 1 are not allowed. - */ -static void -fd_resource_copy_region(struct pipe_context *pctx, - struct pipe_resource *dst, - unsigned dst_level, - unsigned dstx, unsigned dsty, unsigned dstz, - struct pipe_resource *src, - unsigned src_level, - const struct pipe_box *src_box) -{ - struct fd_context *ctx = fd_context(pctx); - - /* TODO if we have 2d core, or other DMA engine that could be used - * for simple copies and reasonably easily synchronized with the 3d - * core, this is where we'd plug it in.. - */ - - /* try blit on 3d pipe: */ - if (fd_blitter_pipe_copy_region(ctx, - dst, dst_level, dstx, dsty, dstz, - src, src_level, src_box)) - return; - - /* else fallback to pure sw: */ - util_resource_copy_region(pctx, - dst, dst_level, dstx, dsty, dstz, - src, src_level, src_box); -} - bool fd_render_condition_check(struct pipe_context *pctx) { @@ -826,133 +1142,33 @@ fd_render_condition_check(struct pipe_context *pctx) return true; } -/** - * Optimal hardware path for blitting pixels. - * Scaling, format conversion, up- and downsampling (resolve) are allowed. - */ -static void -fd_blit(struct pipe_context *pctx, const struct pipe_blit_info *blit_info) -{ - struct fd_context *ctx = fd_context(pctx); - struct pipe_blit_info info = *blit_info; - bool discard = false; - - if (info.src.resource->nr_samples > 1 && - info.dst.resource->nr_samples <= 1 && - !util_format_is_depth_or_stencil(info.src.resource->format) && - !util_format_is_pure_integer(info.src.resource->format)) { - DBG("color resolve unimplemented"); - return; - } - - if (info.render_condition_enable && !fd_render_condition_check(pctx)) - return; - - if (!info.scissor_enable && !info.alpha_blend) { - discard = util_texrange_covers_whole_level(info.dst.resource, - info.dst.level, info.dst.box.x, info.dst.box.y, - info.dst.box.z, info.dst.box.width, - info.dst.box.height, info.dst.box.depth); - } - - if (util_try_blit_via_copy_region(pctx, &info)) { - return; /* done */ - } - - if (info.mask & PIPE_MASK_S) { - DBG("cannot blit stencil, skipping"); - info.mask &= ~PIPE_MASK_S; - } - - if (!util_blitter_is_blit_supported(ctx->blitter, &info)) { - DBG("blit unsupported %s -> %s", - util_format_short_name(info.src.resource->format), - util_format_short_name(info.dst.resource->format)); - return; - } - - fd_blitter_pipe_begin(ctx, info.render_condition_enable, discard, FD_STAGE_BLIT); - util_blitter_blit(ctx->blitter, &info); - fd_blitter_pipe_end(ctx); -} - -void -fd_blitter_pipe_begin(struct fd_context *ctx, bool render_cond, bool discard, - enum fd_render_stage stage) -{ - util_blitter_save_fragment_constant_buffer_slot(ctx->blitter, - ctx->constbuf[PIPE_SHADER_FRAGMENT].cb); - util_blitter_save_vertex_buffer_slot(ctx->blitter, ctx->vtx.vertexbuf.vb); - util_blitter_save_vertex_elements(ctx->blitter, ctx->vtx.vtx); - util_blitter_save_vertex_shader(ctx->blitter, ctx->prog.vp); - util_blitter_save_so_targets(ctx->blitter, ctx->streamout.num_targets, - ctx->streamout.targets); - util_blitter_save_rasterizer(ctx->blitter, ctx->rasterizer); - util_blitter_save_viewport(ctx->blitter, &ctx->viewport); - util_blitter_save_scissor(ctx->blitter, &ctx->scissor); - util_blitter_save_fragment_shader(ctx->blitter, ctx->prog.fp); - util_blitter_save_blend(ctx->blitter, ctx->blend); - util_blitter_save_depth_stencil_alpha(ctx->blitter, ctx->zsa); - util_blitter_save_stencil_ref(ctx->blitter, &ctx->stencil_ref); - util_blitter_save_sample_mask(ctx->blitter, ctx->sample_mask); - util_blitter_save_framebuffer(ctx->blitter, - ctx->batch ? &ctx->batch->framebuffer : NULL); - util_blitter_save_fragment_sampler_states(ctx->blitter, - ctx->tex[PIPE_SHADER_FRAGMENT].num_samplers, - (void **)ctx->tex[PIPE_SHADER_FRAGMENT].samplers); - util_blitter_save_fragment_sampler_views(ctx->blitter, - ctx->tex[PIPE_SHADER_FRAGMENT].num_textures, - ctx->tex[PIPE_SHADER_FRAGMENT].textures); - if (!render_cond) - util_blitter_save_render_condition(ctx->blitter, - ctx->cond_query, ctx->cond_cond, ctx->cond_mode); - - if (ctx->batch) - fd_batch_set_stage(ctx->batch, stage); - - ctx->in_blit = discard; -} - -void -fd_blitter_pipe_end(struct fd_context *ctx) -{ - if (ctx->batch) - fd_batch_set_stage(ctx->batch, FD_STAGE_NULL); - ctx->in_blit = false; -} - -static void -fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc) -{ - struct fd_resource *rsc = fd_resource(prsc); - - if (rsc->write_batch) - fd_batch_flush(rsc->write_batch, true, false); - - assert(!rsc->write_batch); -} - static void fd_invalidate_resource(struct pipe_context *pctx, struct pipe_resource *prsc) { + struct fd_context *ctx = fd_context(pctx); struct fd_resource *rsc = fd_resource(prsc); /* * TODO I guess we could track that the resource is invalidated and * use that as a hint to realloc rather than stall in _transfer_map(), * even in the non-DISCARD_WHOLE_RESOURCE case? + * + * Note: we set dirty bits to trigger invalidate logic fd_draw_vbo */ if (rsc->write_batch) { struct fd_batch *batch = rsc->write_batch; struct pipe_framebuffer_state *pfb = &batch->framebuffer; - if (pfb->zsbuf && pfb->zsbuf->texture == prsc) + if (pfb->zsbuf && pfb->zsbuf->texture == prsc) { batch->resolve &= ~(FD_BUFFER_DEPTH | FD_BUFFER_STENCIL); + ctx->dirty |= FD_DIRTY_ZSA; + } for (unsigned i = 0; i < pfb->nr_cbufs; i++) { if (pfb->cbufs[i] && pfb->cbufs[i]->texture == prsc) { batch->resolve &= ~(PIPE_CLEAR_COLOR0 << i); + ctx->dirty |= FD_DIRTY_FRAMEBUFFER; } } } @@ -996,15 +1212,74 @@ static const struct u_transfer_vtbl transfer_vtbl = { void fd_resource_screen_init(struct pipe_screen *pscreen) { - bool fake_rgtc = fd_screen(pscreen)->gpu_id < 400; + struct fd_screen *screen = fd_screen(pscreen); + bool fake_rgtc = screen->gpu_id < 400; pscreen->resource_create = u_transfer_helper_resource_create; + /* NOTE: u_transfer_helper does not yet support the _with_modifiers() + * variant: + */ + pscreen->resource_create_with_modifiers = fd_resource_create_with_modifiers; pscreen->resource_from_handle = fd_resource_from_handle; pscreen->resource_get_handle = fd_resource_get_handle; pscreen->resource_destroy = u_transfer_helper_resource_destroy; pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl, - true, fake_rgtc, true); + true, false, fake_rgtc, true); + + if (!screen->setup_slices) + screen->setup_slices = fd_setup_slices; +} + +static void +fd_get_sample_position(struct pipe_context *context, + unsigned sample_count, unsigned sample_index, + float *pos_out) +{ + /* The following is copied from nouveau/nv50 except for position + * values, which are taken from blob driver */ + static const uint8_t pos1[1][2] = { { 0x8, 0x8 } }; + static const uint8_t pos2[2][2] = { + { 0xc, 0xc }, { 0x4, 0x4 } }; + static const uint8_t pos4[4][2] = { + { 0x6, 0x2 }, { 0xe, 0x6 }, + { 0x2, 0xa }, { 0xa, 0xe } }; + /* TODO needs to be verified on supported hw */ + static const uint8_t pos8[8][2] = { + { 0x9, 0x5 }, { 0x7, 0xb }, + { 0xd, 0x9 }, { 0x5, 0x3 }, + { 0x3, 0xd }, { 0x1, 0x7 }, + { 0xb, 0xf }, { 0xf, 0x1 } }; + + const uint8_t (*ptr)[2]; + + switch (sample_count) { + case 1: + ptr = pos1; + break; + case 2: + ptr = pos2; + break; + case 4: + ptr = pos4; + break; + case 8: + ptr = pos8; + break; + default: + assert(0); + return; + } + + pos_out[0] = ptr[sample_index][0] / 16.0f; + pos_out[1] = ptr[sample_index][1] / 16.0f; +} + +static void +fd_blit_pipe(struct pipe_context *pctx, const struct pipe_blit_info *blit_info) +{ + /* wrap fd_blit to return void */ + fd_blit(pctx, blit_info); } void @@ -1018,7 +1293,8 @@ fd_resource_context_init(struct pipe_context *pctx) pctx->create_surface = fd_create_surface; pctx->surface_destroy = fd_surface_destroy; pctx->resource_copy_region = fd_resource_copy_region; - pctx->blit = fd_blit; + pctx->blit = fd_blit_pipe; pctx->flush_resource = fd_flush_resource; pctx->invalidate_resource = fd_invalidate_resource; + pctx->get_sample_position = fd_get_sample_position; }