X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Ffreedreno%2Ffreedreno_gmem.c;h=7f6d330d3190cd84b8f7f9eabbf4b0f25a4b5070;hb=8d9ab0a33bb17a7f2eebe4369d8b130548ecbac4;hp=3d959c65dc759cb539e59a8a906831c4d9eb93e6;hpb=4c91930a257ac500029f56514c5504ad66b7cf8f;p=mesa.git

diff --git a/src/gallium/drivers/freedreno/freedreno_gmem.c b/src/gallium/drivers/freedreno/freedreno_gmem.c
index 3d959c65dc7..7f6d330d319 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 <robclark@freedesktop.org>
  *
@@ -26,15 +24,21 @@
  *    Rob Clark <robclark@freedesktop.org>
  */
 
+#include "util/debug.h"
 #include "pipe/p_state.h"
+#include "util/hash_table.h"
+#include "util/u_dump.h"
 #include "util/u_string.h"
 #include "util/u_memory.h"
 #include "util/u_inlines.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
 
 #include "freedreno_gmem.h"
 #include "freedreno_context.h"
+#include "freedreno_fence.h"
+#include "freedreno_log.h"
 #include "freedreno_resource.h"
+#include "freedreno_query_hw.h"
 #include "freedreno_util.h"
 
 /*
@@ -66,205 +70,728 @@
  * resolve.
  */
 
+#ifndef BIN_DEBUG
+#  define BIN_DEBUG 0
+#endif
+
+/*
+ * GMEM Cache:
+ *
+ * Caches GMEM state based on a given framebuffer state.  The key is
+ * meant to be the minimal set of data that results in a unique gmem
+ * configuration, avoiding multiple keys arriving at the same gmem
+ * state.  For example, the render target format is not part of the
+ * key, only the size per pixel.  And the max_scissor bounds is not
+ * part of they key, only the minx/miny (after clamping to tile
+ * alignment) and width/height.  This ensures that slightly different
+ * max_scissor which would result in the same gmem state, do not
+ * become different keys that map to the same state.
+ */
+
+struct gmem_key {
+	uint16_t minx, miny;
+	uint16_t width, height;
+	uint8_t gmem_page_align;      /* alignment in multiples of 0x1000 to reduce key size */
+	uint8_t nr_cbufs;
+	uint8_t cbuf_cpp[MAX_RENDER_TARGETS];
+	uint8_t zsbuf_cpp[2];
+};
+
+static uint32_t
+gmem_key_hash(const void *_key)
+{
+	const struct gmem_key *key = _key;
+	return _mesa_hash_data(key, sizeof(*key));
+}
+
+static bool
+gmem_key_equals(const void *_a, const void *_b)
+{
+	const struct gmem_key *a = _a;
+	const struct gmem_key *b = _b;
+	return memcmp(a, b, sizeof(*a)) == 0;
+}
+
 static void
-calculate_tiles(struct fd_context *ctx)
+dump_gmem_key(const struct gmem_key *key)
 {
-	struct fd_gmem_stateobj *gmem = &ctx->gmem;
-	struct pipe_scissor_state *scissor = &ctx->max_scissor;
-	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
-	uint32_t gmem_size = ctx->screen->gmemsize_bytes;
-	uint32_t minx, miny, width, height;
-	uint32_t nbins_x = 1, nbins_y = 1;
+	printf("{ .minx=%u, .miny=%u, .width=%u, .height=%u",
+			key->minx, key->miny, key->width, key->height);
+	printf(", .gmem_page_align=%u, .nr_cbufs=%u",
+			key->gmem_page_align, key->nr_cbufs);
+	printf(", .cbuf_cpp = {");
+	for (unsigned i = 0; i < ARRAY_SIZE(key->cbuf_cpp); i++)
+		printf("%u,", key->cbuf_cpp[i]);
+	printf("}, .zsbuf_cpp = {");
+	for (unsigned i = 0; i < ARRAY_SIZE(key->zsbuf_cpp); i++)
+		printf("%u,", key->zsbuf_cpp[i]);
+	printf("}},\n");
+}
+
+static void
+dump_gmem_state(const struct fd_gmem_stateobj *gmem)
+{
+	unsigned total = 0;
+	printf("GMEM LAYOUT: bin=%ux%u, nbins=%ux%u\n",
+			gmem->bin_w, gmem->bin_h, gmem->nbins_x, gmem->nbins_y);
+	for (int i = 0; i < ARRAY_SIZE(gmem->cbuf_base); i++) {
+		if (!gmem->cbuf_cpp[i])
+			continue;
+
+		unsigned size = gmem->cbuf_cpp[i] * gmem->bin_w * gmem->bin_h;
+		printf("  cbuf[%d]: base=0x%06x, size=0x%x, cpp=%u\n", i,
+				gmem->cbuf_base[i], size, gmem->cbuf_cpp[i]);
+
+		total = gmem->cbuf_base[i] + size;
+	}
+
+	for (int i = 0; i < ARRAY_SIZE(gmem->zsbuf_base); i++) {
+		if (!gmem->zsbuf_cpp[i])
+			continue;
+
+		unsigned size = gmem->zsbuf_cpp[i] * gmem->bin_w * gmem->bin_h;
+		printf("  zsbuf[%d]: base=0x%06x, size=0x%x, cpp=%u\n", i,
+				gmem->zsbuf_base[i], size, gmem->zsbuf_cpp[i]);
+
+		total = gmem->zsbuf_base[i] + size;
+	}
+
+	printf("total: 0x%06x (of 0x%06x)\n", total,
+			gmem->screen->gmemsize_bytes);
+}
+
+static uint32_t bin_width(struct fd_screen *screen)
+{
+	if (is_a4xx(screen) || is_a5xx(screen) || is_a6xx(screen))
+		return 1024;
+	if (is_a3xx(screen))
+		return 992;
+	return 512;
+}
+
+static unsigned
+div_align(unsigned num, unsigned denom, unsigned al)
+{
+	return util_align_npot(DIV_ROUND_UP(num, denom), al);
+}
+
+static bool
+layout_gmem(struct gmem_key *key, uint32_t nbins_x, uint32_t nbins_y,
+		struct fd_gmem_stateobj *gmem)
+{
+	struct fd_screen *screen = gmem->screen;
+	uint32_t gmem_align = key->gmem_page_align * 0x1000;
+	uint32_t total = 0, i;
+
+	if ((nbins_x == 0) || (nbins_y == 0))
+		return false;
+
 	uint32_t bin_w, bin_h;
-	uint32_t max_width = 992;
-	uint32_t cpp = 4;
+	bin_w = div_align(key->width, nbins_x, screen->tile_alignw);
+	bin_h = div_align(key->height, nbins_y, screen->tile_alignh);
 
-	if (pfb->cbufs[0])
-		cpp = util_format_get_blocksize(pfb->cbufs[0]->format);
+	gmem->bin_w = bin_w;
+	gmem->bin_h = bin_h;
 
-	if ((gmem->cpp == cpp) &&
-			!memcmp(&gmem->scissor, scissor, sizeof(gmem->scissor))) {
-		/* everything is up-to-date */
-		return;
+	/* due to aligning bin_w/h, we could end up with one too
+	 * many bins in either dimension, so recalculate:
+	 */
+	gmem->nbins_x = DIV_ROUND_UP(key->width, bin_w);
+	gmem->nbins_y = DIV_ROUND_UP(key->height, bin_h);
+
+	for (i = 0; i < MAX_RENDER_TARGETS; i++) {
+		if (key->cbuf_cpp[i]) {
+			gmem->cbuf_base[i] = util_align_npot(total, gmem_align);
+			total = gmem->cbuf_base[i] + key->cbuf_cpp[i] * bin_w * bin_h;
+		}
 	}
 
-	if (fd_mesa_debug & FD_DBG_DSCIS) {
-		minx = 0;
-		miny = 0;
-		width = pfb->width;
-		height = pfb->height;
-	} else {
-		minx = scissor->minx & ~31; /* round down to multiple of 32 */
-		miny = scissor->miny & ~31;
-		width = scissor->maxx - minx;
-		height = scissor->maxy - miny;
+	if (key->zsbuf_cpp[0]) {
+		gmem->zsbuf_base[0] = util_align_npot(total, gmem_align);
+		total = gmem->zsbuf_base[0] + key->zsbuf_cpp[0] * bin_w * bin_h;
+	}
+
+	if (key->zsbuf_cpp[1]) {
+		gmem->zsbuf_base[1] = util_align_npot(total, gmem_align);
+		total = gmem->zsbuf_base[1] + key->zsbuf_cpp[1] * bin_w * bin_h;
 	}
 
-// TODO we probably could optimize this a bit if we know that
-// Z or stencil is not enabled for any of the draw calls..
-//	if (fd_stencil_enabled(ctx->zsa) || fd_depth_enabled(ctx->zsa)) {
-		gmem_size /= 2;
-		max_width = 256;
-//	}
+	return total <= screen->gmemsize_bytes;
+}
 
-	bin_w = align(width, 32);
-	bin_h = align(height, 32);
+static void
+calc_nbins(struct gmem_key *key, struct fd_gmem_stateobj *gmem)
+{
+	struct fd_screen *screen = gmem->screen;
+	uint32_t nbins_x = 1, nbins_y = 1;
+	uint32_t max_width = bin_width(screen);
+
+	if (fd_mesa_debug & FD_DBG_MSGS) {
+		debug_printf("binning input: cbuf cpp:");
+		for (unsigned i = 0; i < key->nr_cbufs; i++)
+			debug_printf(" %d", key->cbuf_cpp[i]);
+		debug_printf(", zsbuf cpp: %d; %dx%d\n",
+				key->zsbuf_cpp[0], key->width, key->height);
+	}
 
 	/* first, find a bin width that satisfies the maximum width
 	 * restrictions:
 	 */
-	while (bin_w > max_width) {
+	while (div_align(key->width, nbins_x, screen->tile_alignw) > max_width) {
 		nbins_x++;
-		bin_w = align(width / nbins_x, 32);
 	}
 
-	/* then find a bin height that satisfies the memory constraints:
+	/* then find a bin width/height that satisfies the memory
+	 * constraints:
+	 */
+	while (!layout_gmem(key, nbins_x, nbins_y, gmem)) {
+		if (nbins_y > nbins_x) {
+			nbins_x++;
+		} else {
+			nbins_y++;
+		}
+	}
+
+	/* Lets see if we can tweak the layout a bit and come up with
+	 * something better:
 	 */
-	while ((bin_w * bin_h * cpp) > gmem_size) {
+	if ((((nbins_x - 1) * (nbins_y + 1)) < (nbins_x * nbins_y)) &&
+			layout_gmem(key, nbins_x - 1, nbins_y + 1, gmem)) {
+		nbins_x--;
 		nbins_y++;
-		bin_h = align(height / nbins_y, 32);
+	} else if ((((nbins_x + 1) * (nbins_y - 1)) < (nbins_x * nbins_y)) &&
+			layout_gmem(key, nbins_x + 1, nbins_y - 1, gmem)) {
+		nbins_x++;
+		nbins_y--;
 	}
 
-	DBG("using %d bins of size %dx%d", nbins_x*nbins_y, bin_w, bin_h);
+	layout_gmem(key, nbins_x, nbins_y, gmem);
 
-	gmem->scissor = *scissor;
-	gmem->cpp = cpp;
-	gmem->minx = minx;
-	gmem->miny = miny;
-	gmem->bin_h = bin_h;
-	gmem->bin_w = bin_w;
-	gmem->nbins_x = nbins_x;
-	gmem->nbins_y = nbins_y;
-	gmem->width = width;
-	gmem->height = height;
 }
 
-static void
-render_tiles(struct fd_context *ctx)
+static struct fd_gmem_stateobj *
+gmem_stateobj_init(struct fd_screen *screen, struct gmem_key *key)
 {
-	struct fd_gmem_stateobj *gmem = &ctx->gmem;
-	uint32_t i, yoff = gmem->miny;
+	struct fd_gmem_stateobj *gmem =
+			rzalloc(screen->gmem_cache.ht, struct fd_gmem_stateobj);
+	pipe_reference_init(&gmem->reference, 1);
+	gmem->screen = screen;
+	gmem->key = key;
+	list_inithead(&gmem->node);
+
+	const unsigned npipes = screen->num_vsc_pipes;
+	uint32_t i, j, t, xoff, yoff;
+	uint32_t tpp_x, tpp_y;
+	int tile_n[npipes];
+
+	calc_nbins(key, gmem);
+
+	DBG("using %d bins of size %dx%d", gmem->nbins_x * gmem->nbins_y,
+			gmem->bin_w, gmem->bin_h);
+
+	memcpy(gmem->cbuf_cpp, key->cbuf_cpp, sizeof(key->cbuf_cpp));
+	memcpy(gmem->zsbuf_cpp, key->zsbuf_cpp, sizeof(key->zsbuf_cpp));
+	gmem->minx = key->minx;
+	gmem->miny = key->miny;
+	gmem->width = key->width;
+	gmem->height = key->height;
+
+	if (BIN_DEBUG) {
+		dump_gmem_state(gmem);
+		dump_gmem_key(key);
+	}
+
+	/*
+	 * Assign tiles and pipes:
+	 *
+	 * At some point it might be worth playing with different
+	 * strategies and seeing if that makes much impact on
+	 * performance.
+	 */
+
+#define div_round_up(v, a)  (((v) + (a) - 1) / (a))
+	/* figure out number of tiles per pipe: */
+	if (is_a20x(screen)) {
+		/* for a20x we want to minimize the number of "pipes"
+		 * binning data has 3 bits for x/y (8x8) but the edges are used to
+		 * cull off-screen vertices with hw binning, so we have 6x6 pipes
+		 */
+		tpp_x = 6;
+		tpp_y = 6;
+	} else {
+		tpp_x = tpp_y = 1;
+		while (div_round_up(gmem->nbins_y, tpp_y) > npipes)
+			tpp_y += 2;
+		while ((div_round_up(gmem->nbins_y, tpp_y) *
+				div_round_up(gmem->nbins_x, tpp_x)) > npipes)
+			tpp_x += 1;
+	}
+
+#ifdef DEBUG
+	tpp_x = env_var_as_unsigned("TPP_X", tpp_x);
+	tpp_y = env_var_as_unsigned("TPP_Y", tpp_x);
+#endif
+
+	gmem->maxpw = tpp_x;
+	gmem->maxph = tpp_y;
+
+	/* configure pipes: */
+	xoff = yoff = 0;
+	for (i = 0; i < npipes; i++) {
+		struct fd_vsc_pipe *pipe = &gmem->vsc_pipe[i];
+
+		if (xoff >= gmem->nbins_x) {
+			xoff = 0;
+			yoff += tpp_y;
+		}
+
+		if (yoff >= gmem->nbins_y) {
+			break;
+		}
+
+		pipe->x = xoff;
+		pipe->y = yoff;
+		pipe->w = MIN2(tpp_x, gmem->nbins_x - xoff);
+		pipe->h = MIN2(tpp_y, gmem->nbins_y - yoff);
+
+		xoff += tpp_x;
+	}
 
-	ctx->emit_tile_init(ctx);
+	/* number of pipes to use for a20x */
+	gmem->num_vsc_pipes = MAX2(1, i);
 
+	for (; i < npipes; i++) {
+		struct fd_vsc_pipe *pipe = &gmem->vsc_pipe[i];
+		pipe->x = pipe->y = pipe->w = pipe->h = 0;
+	}
+
+	if (BIN_DEBUG) {
+		printf("%dx%d ... tpp=%dx%d\n", gmem->nbins_x, gmem->nbins_y, tpp_x, tpp_y);
+		for (i = 0; i < ARRAY_SIZE(gmem->vsc_pipe); i++) {
+			struct fd_vsc_pipe *pipe = &gmem->vsc_pipe[i];
+			printf("pipe[%d]: %ux%u @ %u,%u\n", i,
+					pipe->w, pipe->h, pipe->x, pipe->y);
+		}
+	}
+
+	/* configure tiles: */
+	t = 0;
+	yoff = key->miny;
+	memset(tile_n, 0, sizeof(tile_n));
 	for (i = 0; i < gmem->nbins_y; i++) {
-		uint32_t j, xoff = gmem->minx;
-		uint32_t bh = gmem->bin_h;
+		int bw, bh;
+
+		xoff = key->minx;
 
 		/* clip bin height: */
-		bh = MIN2(bh, gmem->miny + gmem->height - yoff);
+		bh = MIN2(gmem->bin_h, key->miny + key->height - yoff);
+		assert(bh > 0);
 
 		for (j = 0; j < gmem->nbins_x; j++) {
-			uint32_t bw = gmem->bin_w;
+			struct fd_tile *tile = &gmem->tile[t];
+			uint32_t p;
 
-			/* clip bin width: */
-			bw = MIN2(bw, gmem->minx + gmem->width - xoff);
-
-			DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
-					bh, yoff, bw, xoff);
+			assert(t < ARRAY_SIZE(gmem->tile));
 
-			ctx->emit_tile_prep(ctx, xoff, yoff, bw, bh);
+			/* pipe number: */
+			p = ((i / tpp_y) * div_round_up(gmem->nbins_x, tpp_x)) + (j / tpp_x);
+			assert(p < gmem->num_vsc_pipes);
 
-			if (ctx->restore)
-				ctx->emit_tile_mem2gmem(ctx, xoff, yoff, bw, bh);
+			/* clip bin width: */
+			bw = MIN2(gmem->bin_w, key->minx + key->width - xoff);
+			assert(bw > 0);
 
-			ctx->emit_tile_renderprep(ctx, xoff, yoff, bw, bh);
+			tile->n = !is_a20x(screen) ? tile_n[p]++ :
+				((i % tpp_y + 1) << 3 | (j % tpp_x + 1));
+			tile->p = p;
+			tile->bin_w = bw;
+			tile->bin_h = bh;
+			tile->xoff = xoff;
+			tile->yoff = yoff;
 
-			/* emit IB to drawcmds: */
-			OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
+			if (BIN_DEBUG) {
+				printf("tile[%d]: p=%u, bin=%ux%u+%u+%u\n", t,
+						p, bw, bh, xoff, yoff);
+			}
 
-			/* emit gmem2mem to transfer tile back to system memory: */
-			ctx->emit_tile_gmem2mem(ctx, xoff, yoff, bw, bh);
+			t++;
 
 			xoff += bw;
 		}
 
 		yoff += bh;
 	}
+
+	if (BIN_DEBUG) {
+		t = 0;
+		for (i = 0; i < gmem->nbins_y; i++) {
+			for (j = 0; j < gmem->nbins_x; j++) {
+				struct fd_tile *tile = &gmem->tile[t++];
+				printf("|p:%u n:%u|", tile->p, tile->n);
+			}
+			printf("\n");
+		}
+	}
+
+	return gmem;
+}
+
+void
+__fd_gmem_destroy(struct fd_gmem_stateobj *gmem)
+{
+	struct fd_gmem_cache *cache = &gmem->screen->gmem_cache;
+
+	fd_screen_assert_locked(gmem->screen);
+
+	_mesa_hash_table_remove_key(cache->ht, gmem->key);
+	list_del(&gmem->node);
+
+	ralloc_free(gmem->key);
+	ralloc_free(gmem);
+}
+
+static struct gmem_key *
+gmem_key_init(struct fd_batch *batch, bool assume_zs, bool no_scis_opt)
+{
+	struct fd_screen *screen = batch->ctx->screen;
+	struct pipe_framebuffer_state *pfb = &batch->framebuffer;
+	bool has_zs = pfb->zsbuf && !!(batch->gmem_reason & (FD_GMEM_DEPTH_ENABLED |
+		FD_GMEM_STENCIL_ENABLED | FD_GMEM_CLEARS_DEPTH_STENCIL));
+	struct gmem_key *key = rzalloc(screen->gmem_cache.ht, struct gmem_key);
+
+	if (has_zs || assume_zs) {
+		struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
+		key->zsbuf_cpp[0] = rsc->layout.cpp;
+		if (rsc->stencil)
+			key->zsbuf_cpp[1] = rsc->stencil->layout.cpp;
+	} else {
+		/* we might have a zsbuf, but it isn't used */
+		batch->restore &= ~(FD_BUFFER_DEPTH | FD_BUFFER_STENCIL);
+		batch->resolve &= ~(FD_BUFFER_DEPTH | FD_BUFFER_STENCIL);
+	}
+
+	key->nr_cbufs = pfb->nr_cbufs;
+	for (unsigned i = 0; i < pfb->nr_cbufs; i++) {
+		if (pfb->cbufs[i])
+			key->cbuf_cpp[i] = util_format_get_blocksize(pfb->cbufs[i]->format);
+		else
+			key->cbuf_cpp[i] = 4;
+		/* if MSAA, color buffers are super-sampled in GMEM: */
+		key->cbuf_cpp[i] *= pfb->samples;
+	}
+
+	/* NOTE: on a6xx, the max-scissor-rect is handled in fd6_gmem, and
+	 * we just rely on CP_COND_EXEC to skip bins with no geometry.
+	 */
+	if (no_scis_opt || is_a6xx(screen)) {
+		key->minx = 0;
+		key->miny = 0;
+		key->width = pfb->width;
+		key->height = pfb->height;
+	} else {
+		struct pipe_scissor_state *scissor = &batch->max_scissor;
+
+		if (fd_mesa_debug & FD_DBG_NOSCIS) {
+			scissor->minx = 0;
+			scissor->miny = 0;
+			scissor->maxx = pfb->width;
+			scissor->maxy = pfb->height;
+		}
+
+		/* round down to multiple of alignment: */
+		key->minx = scissor->minx & ~(screen->gmem_alignw - 1);
+		key->miny = scissor->miny & ~(screen->gmem_alignh - 1);
+		key->width = scissor->maxx - key->minx;
+		key->height = scissor->maxy - key->miny;
+	}
+
+	if (is_a20x(screen) && batch->cleared) {
+		/* under normal circumstances the requirement would be 4K
+		 * but the fast clear path requires an alignment of 32K
+		 */
+		key->gmem_page_align = 8;
+	} else if (is_a6xx(screen)) {
+		key->gmem_page_align = is_a650(screen) ? 3 : 1;
+	} else {
+		// TODO re-check this across gens.. maybe it should only
+		// be a single page in some cases:
+		key->gmem_page_align = 4;
+	}
+
+	return key;
+}
+
+static struct fd_gmem_stateobj *
+lookup_gmem_state(struct fd_batch *batch, bool assume_zs, bool no_scis_opt)
+{
+	struct fd_screen *screen = batch->ctx->screen;
+	struct fd_gmem_cache *cache = &screen->gmem_cache;
+	struct fd_gmem_stateobj *gmem = NULL;
+	struct gmem_key *key = gmem_key_init(batch, assume_zs, no_scis_opt);
+	uint32_t hash = gmem_key_hash(key);
+
+	fd_screen_lock(screen);
+
+	struct hash_entry *entry =
+		_mesa_hash_table_search_pre_hashed(cache->ht, hash, key);
+	if (entry) {
+		ralloc_free(key);
+		goto found;
+	}
+
+	/* limit the # of cached gmem states, discarding the least
+	 * recently used state if needed:
+	 */
+	if (cache->ht->entries >= 20) {
+		struct fd_gmem_stateobj *last =
+			list_last_entry(&cache->lru, struct fd_gmem_stateobj, node);
+		fd_gmem_reference(&last, NULL);
+	}
+
+	entry = _mesa_hash_table_insert_pre_hashed(cache->ht,
+			hash, key, gmem_stateobj_init(screen, key));
+
+found:
+	fd_gmem_reference(&gmem, entry->data);
+	/* Move to the head of the LRU: */
+	list_delinit(&gmem->node);
+	list_add(&gmem->node, &cache->lru);
+
+	fd_screen_unlock(screen);
+
+	return gmem;
+}
+
+/*
+ * GMEM render pass
+ */
+
+static void
+render_tiles(struct fd_batch *batch, struct fd_gmem_stateobj *gmem)
+{
+	struct fd_context *ctx = batch->ctx;
+	int i;
+
+	mtx_lock(&ctx->gmem_lock);
+
+	ctx->emit_tile_init(batch);
+
+	if (batch->restore)
+		ctx->stats.batch_restore++;
+
+	for (i = 0; i < (gmem->nbins_x * gmem->nbins_y); i++) {
+		struct fd_tile *tile = &gmem->tile[i];
+
+		fd_log(batch, "bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
+			tile->bin_h, tile->yoff, tile->bin_w, tile->xoff);
+
+		ctx->emit_tile_prep(batch, tile);
+
+		if (batch->restore) {
+			ctx->emit_tile_mem2gmem(batch, tile);
+		}
+
+		ctx->emit_tile_renderprep(batch, tile);
+
+		if (ctx->query_prepare_tile)
+			ctx->query_prepare_tile(batch, i, batch->gmem);
+
+		/* emit IB to drawcmds: */
+		fd_log(batch, "TILE[%d]: START DRAW IB", i);
+		if (ctx->emit_tile) {
+			ctx->emit_tile(batch, tile);
+		} else {
+			ctx->screen->emit_ib(batch->gmem, batch->draw);
+		}
+
+		fd_log(batch, "TILE[%d]: END DRAW IB", i);
+		fd_reset_wfi(batch);
+
+		/* emit gmem2mem to transfer tile back to system memory: */
+		ctx->emit_tile_gmem2mem(batch, tile);
+	}
+
+	if (ctx->emit_tile_fini)
+		ctx->emit_tile_fini(batch);
+
+	mtx_unlock(&ctx->gmem_lock);
 }
 
 static void
-render_sysmem(struct fd_context *ctx)
+render_sysmem(struct fd_batch *batch)
 {
-	ctx->emit_sysmem_prep(ctx);
+	struct fd_context *ctx = batch->ctx;
+
+	ctx->emit_sysmem_prep(batch);
+
+	if (ctx->query_prepare_tile)
+		ctx->query_prepare_tile(batch, 0, batch->gmem);
 
 	/* emit IB to drawcmds: */
-	OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
+	fd_log(batch, "SYSMEM: START DRAW IB");
+	ctx->screen->emit_ib(batch->gmem, batch->draw);
+	fd_log(batch, "SYSMEM: END DRAW IB");
+	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;
+
+	if (unlikely(fd_mesa_debug & FD_DBG_NOHW))
+		return;
+
+	fd_submit_flush(batch->submit, batch->in_fence_fd,
+			batch->needs_out_fence_fd ? &out_fence_fd : NULL,
+			&timestamp);
+
+	fd_fence_populate(batch->fence, timestamp, out_fence_fd);
+	fd_log_flush(batch);
 }
 
 void
-fd_gmem_render_tiles(struct pipe_context *pctx)
+fd_gmem_render_tiles(struct fd_batch *batch)
 {
-	struct fd_context *ctx = fd_context(pctx);
-	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
-	uint32_t timestamp = 0;
+	struct fd_context *ctx = batch->ctx;
+	struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 	bool sysmem = false;
 
-	if (ctx->emit_sysmem_prep) {
-		if (ctx->cleared || ctx->gmem_reason || (ctx->num_draws > 5)) {
-			DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u",
-				ctx->cleared, ctx->gmem_reason, ctx->num_draws);
-		} else {
+	if (ctx->emit_sysmem_prep && !batch->nondraw) {
+		if (batch->cleared || batch->gmem_reason ||
+				((batch->num_draws > 5) && !batch->blit) ||
+				(pfb->samples > 1)) {
+			fd_log(batch, "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;
 		}
+
+		/* For ARB_framebuffer_no_attachments: */
+		if ((pfb->nr_cbufs == 0) && !pfb->zsbuf) {
+			sysmem = true;
+		}
+	}
+
+	if (fd_mesa_debug & FD_DBG_NOGMEM)
+		sysmem = true;
+
+	/* Layered rendering always needs bypass. */
+	for (unsigned i = 0; i < pfb->nr_cbufs; i++) {
+		struct pipe_surface *psurf = pfb->cbufs[i];
+		if (!psurf)
+			continue;
+		if (psurf->u.tex.first_layer < psurf->u.tex.last_layer)
+			sysmem = true;
+	}
+
+	/* Tessellation doesn't seem to support tiled rendering so fall back to
+	 * bypass.
+	 */
+	if (batch->tessellation) {
+		debug_assert(ctx->emit_sysmem_prep);
+		sysmem = true;
 	}
 
-	/* mark the end of the clear/draw cmds before emitting per-tile cmds: */
-	fd_ringmarker_mark(ctx->draw_end);
+	fd_reset_wfi(batch);
+
+	ctx->stats.batch_total++;
 
-	if (sysmem) {
-		DBG("rendering sysmem (%s/%s)",
+	if (unlikely(fd_mesa_debug & FD_DBG_LOG) && !batch->nondraw) {
+		fd_log_stream(batch, stream, util_dump_framebuffer_state(stream, pfb));
+		for (unsigned i = 0; i < pfb->nr_cbufs; i++) {
+			fd_log_stream(batch, stream, util_dump_surface(stream, pfb->cbufs[i]));
+		}
+		fd_log_stream(batch, stream, util_dump_surface(stream, pfb->zsbuf));
+	}
+
+	if (batch->nondraw) {
+		DBG("%p: rendering non-draw", batch);
+		render_sysmem(batch);
+		ctx->stats.batch_nondraw++;
+	} else if (sysmem) {
+		fd_log(batch, "%p: rendering sysmem %ux%u (%s/%s), num_draws=%u",
+			batch, pfb->width, pfb->height,
 			util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
-			util_format_short_name(pipe_surface_format(pfb->zsbuf)));
-		render_sysmem(ctx);
+			util_format_short_name(pipe_surface_format(pfb->zsbuf)),
+			batch->num_draws);
+		if (ctx->query_prepare)
+			ctx->query_prepare(batch, 1);
+		render_sysmem(batch);
+		ctx->stats.batch_sysmem++;
 	} else {
-		struct fd_gmem_stateobj *gmem = &ctx->gmem;
-		calculate_tiles(ctx);
-		DBG("rendering %dx%d tiles (%s/%s)", gmem->nbins_x, gmem->nbins_y,
+		struct fd_gmem_stateobj *gmem = lookup_gmem_state(batch, false, false);
+		batch->gmem_state = gmem;
+		fd_log(batch, "%p: rendering %dx%d tiles %ux%u (%s/%s)",
+			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)));
-		render_tiles(ctx);
+		if (ctx->query_prepare)
+			ctx->query_prepare(batch, gmem->nbins_x * gmem->nbins_y);
+		render_tiles(batch, gmem);
+		batch->gmem_state = NULL;
+
+		fd_screen_lock(ctx->screen);
+		fd_gmem_reference(&gmem, NULL);
+		fd_screen_unlock(ctx->screen);
+
+		ctx->stats.batch_gmem++;
 	}
 
-	/* GPU executes starting from tile cmds, which IB back to draw cmds: */
-	fd_ringmarker_flush(ctx->draw_end);
+	flush_ring(batch);
+}
 
-	/* mark start for next draw cmds: */
-	fd_ringmarker_mark(ctx->draw_start);
+/* Determine a worst-case estimate (ie. assuming we don't eliminate an
+ * unused depth/stencil) number of bins per vsc pipe.
+ */
+unsigned
+fd_gmem_estimate_bins_per_pipe(struct fd_batch *batch)
+{
+	struct pipe_framebuffer_state *pfb = &batch->framebuffer;
+	struct fd_screen *screen = batch->ctx->screen;
+	struct fd_gmem_stateobj *gmem = lookup_gmem_state(batch, !!pfb->zsbuf, true);
+	unsigned nbins = gmem->maxpw * gmem->maxph;
 
-	/* update timestamps on render targets: */
-	timestamp = fd_ringbuffer_timestamp(ctx->ring);
-	if (pfb->cbufs[0])
-		fd_resource(pfb->cbufs[0]->texture)->timestamp = timestamp;
-	if (pfb->zsbuf)
-		fd_resource(pfb->zsbuf->texture)->timestamp = timestamp;
+	fd_screen_lock(screen);
+	fd_gmem_reference(&gmem, NULL);
+	fd_screen_unlock(screen);
 
-	/* reset maximal bounds: */
-	ctx->max_scissor.minx = ctx->max_scissor.miny = ~0;
-	ctx->max_scissor.maxx = ctx->max_scissor.maxy = 0;
+	return nbins;
+}
 
-	/* Note that because the per-tile setup and mem2gmem/gmem2mem are emitted
-	 * after the draw/clear calls, but executed before, we need to preemptively
-	 * flag some state as dirty before the first draw/clear call.
-	 *
-	 * TODO maybe we need to mark all state as dirty to not worry about state
-	 * being clobbered by other contexts?
-	 */
-	ctx->dirty |= FD_DIRTY_ZSA |
-			FD_DIRTY_RASTERIZER |
-			FD_DIRTY_FRAMEBUFFER |
-			FD_DIRTY_SAMPLE_MASK |
-			FD_DIRTY_VIEWPORT |
-			FD_DIRTY_CONSTBUF |
-			FD_DIRTY_PROG |
-			FD_DIRTY_SCISSOR |
-			/* probably only needed if we need to mem2gmem on the next
-			 * draw..  but not sure if there is a good way to know?
-			 */
-			FD_DIRTY_VERTTEX |
-			FD_DIRTY_FRAGTEX |
-			FD_DIRTY_BLEND;
-
-	if (fd_mesa_debug & FD_DBG_DGMEM)
-		ctx->dirty = 0xffffffff;
+/* When deciding whether a tile needs mem2gmem, we need to take into
+ * account the scissor rect(s) that were cleared.  To simplify we only
+ * consider the last scissor rect for each buffer, since the common
+ * case would be a single clear.
+ */
+bool
+fd_gmem_needs_restore(struct fd_batch *batch, const struct fd_tile *tile,
+		uint32_t buffers)
+{
+	if (!(batch->restore & buffers))
+		return false;
+
+	return true;
+}
+
+void
+fd_gmem_screen_init(struct pipe_screen *pscreen)
+{
+	struct fd_gmem_cache *cache = &fd_screen(pscreen)->gmem_cache;
+
+	cache->ht = _mesa_hash_table_create(NULL, gmem_key_hash, gmem_key_equals);
+	list_inithead(&cache->lru);
+}
+
+void
+fd_gmem_screen_fini(struct pipe_screen *pscreen)
+{
+	struct fd_gmem_cache *cache = &fd_screen(pscreen)->gmem_cache;
+
+	_mesa_hash_table_destroy(cache->ht, NULL);
 }