X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Ffreedreno%2Ffreedreno_gmem.c;h=bb15f0a3e164d529c4fddb02ad49fd24cf2f2ab6;hb=cc6a5e937b45bc54a990f8b25b8bc57de35feae9;hp=ed013d9d037d63a7ddd81c73070a8a376649b9cd;hpb=00bed8a794de3d80a46b65b9ab23c6df83e416a8;p=mesa.git diff --git a/src/gallium/drivers/freedreno/freedreno_gmem.c b/src/gallium/drivers/freedreno/freedreno_gmem.c index ed013d9d037..bb15f0a3e16 100644 --- a/src/gallium/drivers/freedreno/freedreno_gmem.c +++ b/src/gallium/drivers/freedreno/freedreno_gmem.c @@ -1,5 +1,3 @@ -/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */ - /* * Copyright (C) 2012 Rob Clark * @@ -34,6 +32,7 @@ #include "freedreno_gmem.h" #include "freedreno_context.h" +#include "freedreno_fence.h" #include "freedreno_resource.h" #include "freedreno_query_hw.h" #include "freedreno_util.h" @@ -69,7 +68,7 @@ static uint32_t bin_width(struct fd_screen *screen) { - if (is_a4xx(screen)) + if (is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen)) return 1024; if (is_a3xx(screen)) return 992; @@ -106,19 +105,23 @@ static void calculate_tiles(struct fd_batch *batch) { struct fd_context *ctx = batch->ctx; + struct fd_screen *screen = ctx->screen; struct fd_gmem_stateobj *gmem = &ctx->gmem; struct pipe_scissor_state *scissor = &batch->max_scissor; struct pipe_framebuffer_state *pfb = &batch->framebuffer; - uint32_t gmem_size = ctx->screen->gmemsize_bytes; + const uint32_t gmem_alignw = screen->gmem_alignw; + const uint32_t gmem_alignh = screen->gmem_alignh; + const unsigned npipes = screen->num_vsc_pipes; + const uint32_t gmem_size = screen->gmemsize_bytes; uint32_t minx, miny, width, height; uint32_t nbins_x = 1, nbins_y = 1; uint32_t bin_w, bin_h; - uint32_t max_width = bin_width(ctx->screen); + uint32_t max_width = bin_width(screen); uint8_t cbuf_cpp[MAX_RENDER_TARGETS] = {0}, zsbuf_cpp[2] = {0}; uint32_t i, j, t, xoff, yoff; uint32_t tpp_x, tpp_y; bool has_zs = !!(batch->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL)); - int tile_n[ARRAY_SIZE(ctx->pipe)]; + int tile_n[npipes]; if (has_zs) { struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture); @@ -131,6 +134,8 @@ calculate_tiles(struct fd_batch *batch) cbuf_cpp[i] = util_format_get_blocksize(pfb->cbufs[i]->format); else cbuf_cpp[i] = 4; + /* if MSAA, color buffers are super-sampled in GMEM: */ + cbuf_cpp[i] *= pfb->samples; } if (!memcmp(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp)) && @@ -146,21 +151,22 @@ calculate_tiles(struct fd_batch *batch) width = pfb->width; height = pfb->height; } else { - minx = scissor->minx & ~31; /* round down to multiple of 32 */ - miny = scissor->miny & ~31; + /* round down to multiple of alignment: */ + minx = scissor->minx & ~(gmem_alignw - 1); + miny = scissor->miny & ~(gmem_alignh - 1); width = scissor->maxx - minx; height = scissor->maxy - miny; } - bin_w = align(width, 32); - bin_h = align(height, 32); + bin_w = align(width, gmem_alignw); + bin_h = align(height, gmem_alignh); /* first, find a bin width that satisfies the maximum width * restrictions: */ while (bin_w > max_width) { nbins_x++; - bin_w = align(width / nbins_x, 32); + bin_w = align(width / nbins_x, gmem_alignw); } if (fd_mesa_debug & FD_DBG_MSGS) { @@ -177,10 +183,10 @@ calculate_tiles(struct fd_batch *batch) while (total_size(cbuf_cpp, zsbuf_cpp, bin_w, bin_h, gmem) > gmem_size) { if (bin_w > bin_h) { nbins_x++; - bin_w = align(width / nbins_x, 32); + bin_w = align(width / nbins_x, gmem_alignw); } else { nbins_y++; - bin_h = align(height / nbins_y, 32); + bin_h = align(height / nbins_y, gmem_alignh); } } @@ -209,16 +215,19 @@ calculate_tiles(struct fd_batch *batch) #define div_round_up(v, a) (((v) + (a) - 1) / (a)) /* figure out number of tiles per pipe: */ tpp_x = tpp_y = 1; - while (div_round_up(nbins_y, tpp_y) > 8) + while (div_round_up(nbins_y, tpp_y) > screen->num_vsc_pipes) tpp_y += 2; while ((div_round_up(nbins_y, tpp_y) * - div_round_up(nbins_x, tpp_x)) > 8) + div_round_up(nbins_x, tpp_x)) > screen->num_vsc_pipes) tpp_x += 1; + gmem->maxpw = tpp_x; + gmem->maxph = tpp_y; + /* configure pipes: */ xoff = yoff = 0; - for (i = 0; i < ARRAY_SIZE(ctx->pipe); i++) { - struct fd_vsc_pipe *pipe = &ctx->pipe[i]; + for (i = 0; i < npipes; i++) { + struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i]; if (xoff >= nbins_x) { xoff = 0; @@ -237,8 +246,8 @@ calculate_tiles(struct fd_batch *batch) xoff += tpp_x; } - for (; i < ARRAY_SIZE(ctx->pipe); i++) { - struct fd_vsc_pipe *pipe = &ctx->pipe[i]; + for (; i < npipes; i++) { + struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i]; pipe->x = pipe->y = pipe->w = pipe->h = 0; } @@ -323,24 +332,24 @@ render_tiles(struct fd_batch *batch) ctx->emit_tile_prep(batch, tile); if (batch->restore) { - fd_hw_query_set_stage(batch, batch->gmem, FD_STAGE_MEM2GMEM); ctx->emit_tile_mem2gmem(batch, tile); - fd_hw_query_set_stage(batch, batch->gmem, FD_STAGE_NULL); } ctx->emit_tile_renderprep(batch, tile); - fd_hw_query_prepare_tile(batch, i, batch->gmem); + if (ctx->query_prepare_tile) + ctx->query_prepare_tile(batch, i, batch->gmem); /* emit IB to drawcmds: */ ctx->emit_ib(batch->gmem, batch->draw); - fd_reset_wfi(ctx); + fd_reset_wfi(batch); /* emit gmem2mem to transfer tile back to system memory: */ - fd_hw_query_set_stage(batch, batch->gmem, FD_STAGE_GMEM2MEM); ctx->emit_tile_gmem2mem(batch, tile); - fd_hw_query_set_stage(batch, batch->gmem, FD_STAGE_NULL); } + + if (ctx->emit_tile_fini) + ctx->emit_tile_fini(batch); } static void @@ -350,11 +359,28 @@ render_sysmem(struct fd_batch *batch) ctx->emit_sysmem_prep(batch); - fd_hw_query_prepare_tile(batch, 0, batch->gmem); + if (ctx->query_prepare_tile) + ctx->query_prepare_tile(batch, 0, batch->gmem); /* emit IB to drawcmds: */ ctx->emit_ib(batch->gmem, batch->draw); - fd_reset_wfi(ctx); + fd_reset_wfi(batch); + + if (ctx->emit_sysmem_fini) + ctx->emit_sysmem_fini(batch); +} + +static void +flush_ring(struct fd_batch *batch) +{ + uint32_t timestamp; + int out_fence_fd = -1; + + fd_submit_flush(batch->submit, batch->in_fence_fd, + batch->needs_out_fence_fd ? &out_fence_fd : NULL, + ×tamp); + + fd_fence_populate(batch->fence, timestamp, out_fence_fd); } void @@ -364,30 +390,37 @@ fd_gmem_render_tiles(struct fd_batch *batch) struct pipe_framebuffer_state *pfb = &batch->framebuffer; bool sysmem = false; - if (ctx->emit_sysmem_prep) { - if (batch->cleared || batch->gmem_reason || (batch->num_draws > 5)) { - DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u", - batch->cleared, batch->gmem_reason, batch->num_draws); + if (ctx->emit_sysmem_prep && !batch->nondraw) { + if (batch->cleared || batch->gmem_reason || + ((batch->num_draws > 5) && !batch->blit) || + (pfb->samples > 1)) { + DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u, samples=%u", + batch->cleared, batch->gmem_reason, batch->num_draws, + pfb->samples); } else if (!(fd_mesa_debug & FD_DBG_NOBYPASS)) { sysmem = true; } - } - /* close out the draw cmds by making sure any active queries are - * paused: - */ - fd_hw_query_set_stage(batch, batch->draw, FD_STAGE_NULL); + /* For ARB_framebuffer_no_attachments: */ + if ((pfb->nr_cbufs == 0) && !pfb->zsbuf) { + sysmem = true; + } + } - fd_reset_wfi(ctx); + fd_reset_wfi(batch); ctx->stats.batch_total++; - if (sysmem) { + if (batch->nondraw) { + DBG("%p: rendering non-draw", batch); + ctx->stats.batch_nondraw++; + } else if (sysmem) { DBG("%p: rendering sysmem %ux%u (%s/%s)", batch, pfb->width, pfb->height, util_format_short_name(pipe_surface_format(pfb->cbufs[0])), util_format_short_name(pipe_surface_format(pfb->zsbuf))); - fd_hw_query_prepare(batch, 1); + if (ctx->query_prepare) + ctx->query_prepare(batch, 1); render_sysmem(batch); ctx->stats.batch_sysmem++; } else { @@ -397,28 +430,13 @@ fd_gmem_render_tiles(struct fd_batch *batch) batch, pfb->width, pfb->height, gmem->nbins_x, gmem->nbins_y, util_format_short_name(pipe_surface_format(pfb->cbufs[0])), util_format_short_name(pipe_surface_format(pfb->zsbuf))); - fd_hw_query_prepare(batch, gmem->nbins_x * gmem->nbins_y); + if (ctx->query_prepare) + ctx->query_prepare(batch, gmem->nbins_x * gmem->nbins_y); render_tiles(batch); ctx->stats.batch_gmem++; } - fd_ringbuffer_flush(batch->gmem); - - fd_reset_wfi(ctx); -} - -/* tile needs restore if it isn't completely contained within the - * cleared scissor: - */ -static bool -skip_restore(struct pipe_scissor_state *scissor, struct fd_tile *tile) -{ - unsigned minx = tile->xoff; - unsigned maxx = tile->xoff + tile->bin_w; - unsigned miny = tile->yoff; - unsigned maxy = tile->yoff + tile->bin_h; - return (minx >= scissor->minx) && (maxx <= scissor->maxx) && - (miny >= scissor->miny) && (maxy <= scissor->maxy); + flush_ring(batch); } /* When deciding whether a tile needs mem2gmem, we need to take into @@ -433,21 +451,5 @@ fd_gmem_needs_restore(struct fd_batch *batch, struct fd_tile *tile, if (!(batch->restore & buffers)) return false; - /* if buffers partially cleared, then slow-path to figure out - * if this particular tile needs restoring: - */ - if ((buffers & FD_BUFFER_COLOR) && - (batch->partial_cleared & FD_BUFFER_COLOR) && - skip_restore(&batch->cleared_scissor.color, tile)) - return false; - if ((buffers & FD_BUFFER_DEPTH) && - (batch->partial_cleared & FD_BUFFER_DEPTH) && - skip_restore(&batch->cleared_scissor.depth, tile)) - return false; - if ((buffers & FD_BUFFER_STENCIL) && - (batch->partial_cleared & FD_BUFFER_STENCIL) && - skip_restore(&batch->cleared_scissor.stencil, tile)) - return false; - return true; }