X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Ffreedreno%2Ffreedreno_context.c;h=e783a8ac71599a29de7d4a072c395a94afd02800;hb=94f4ecba8dd8611133b94154a28b6e85cc9903de;hp=23f6a67734d1d5949d6cf61c86df87566e0b9078;hpb=c0766528baaef48902c87bbdaa4f5926c472269b;p=mesa.git

diff --git a/src/gallium/drivers/freedreno/freedreno_context.c b/src/gallium/drivers/freedreno/freedreno_context.c
index 23f6a67734d..e783a8ac715 100644
--- a/src/gallium/drivers/freedreno/freedreno_context.c
+++ b/src/gallium/drivers/freedreno/freedreno_context.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>
  *
@@ -27,111 +25,182 @@
  */
 
 #include "freedreno_context.h"
+#include "freedreno_blitter.h"
 #include "freedreno_draw.h"
+#include "freedreno_fence.h"
+#include "freedreno_log.h"
+#include "freedreno_program.h"
 #include "freedreno_resource.h"
 #include "freedreno_texture.h"
 #include "freedreno_state.h"
 #include "freedreno_gmem.h"
+#include "freedreno_query.h"
+#include "freedreno_query_hw.h"
 #include "freedreno_util.h"
+#include "util/u_upload_mgr.h"
+
+#if DETECT_OS_ANDROID
+#include "util/u_process.h"
+#include <sys/stat.h>
+#include <sys/types.h>
+#endif
 
-static struct fd_ringbuffer *next_rb(struct fd_context *ctx)
+static void
+fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fencep,
+		unsigned flags)
 {
-	struct fd_ringbuffer *ring;
-	uint32_t ts;
+	struct fd_context *ctx = fd_context(pctx);
+	struct pipe_fence_handle *fence = NULL;
+	// TODO we want to lookup batch if it exists, but not create one if not.
+	struct fd_batch *batch = fd_context_batch(ctx);
 
-	/* grab next ringbuffer: */
-	ring = ctx->rings[(ctx->rings_idx++) % ARRAY_SIZE(ctx->rings)];
+	DBG("%p: flush: flags=%x\n", ctx->batch, flags);
 
-	/* wait for new rb to be idle: */
-	ts = fd_ringbuffer_timestamp(ring);
-	if (ts) {
-		DBG("wait: %u", ts);
-		fd_pipe_wait(ctx->screen->pipe, ts);
-	}
+	/* In some sequence of events, we can end up with a last_fence that is
+	 * not an "fd" fence, which results in eglDupNativeFenceFDANDROID()
+	 * errors.
+	 */
+	if ((flags & PIPE_FLUSH_FENCE_FD) && ctx->last_fence &&
+			!fd_fence_is_fd(ctx->last_fence))
+		fd_fence_ref(&ctx->last_fence, NULL);
 
-	fd_ringbuffer_reset(ring);
+	/* if no rendering since last flush, ie. app just decided it needed
+	 * a fence, re-use the last one:
+	 */
+	if (ctx->last_fence) {
+		fd_fence_ref(&fence, ctx->last_fence);
+		fd_bc_dump(ctx->screen, "%p: reuse last_fence, remaining:\n", ctx);
+		goto out;
+	}
 
-	return ring;
-}
+	if (!batch) {
+		fd_bc_dump(ctx->screen, "%p: NULL batch, remaining:\n", ctx);
+		return;
+	}
 
-static void
-fd_context_next_rb(struct pipe_context *pctx)
-{
-	struct fd_context *ctx = fd_context(pctx);
-	struct fd_ringbuffer *ring;
+	/* Take a ref to the batch's fence (batch can be unref'd when flushed: */
+	fd_fence_ref(&fence, batch->fence);
 
-	fd_ringmarker_del(ctx->draw_start);
-	fd_ringmarker_del(ctx->draw_end);
+	if (flags & PIPE_FLUSH_FENCE_FD)
+		batch->needs_out_fence_fd = true;
 
-	ring = next_rb(ctx);
+	fd_bc_dump(ctx->screen, "%p: flushing %p<%u>, flags=0x%x, pending:\n",
+			ctx, batch, batch->seqno, flags);
 
-	ctx->draw_start = fd_ringmarker_new(ring);
-	ctx->draw_end = fd_ringmarker_new(ring);
+	if (!ctx->screen->reorder) {
+		fd_batch_flush(batch);
+	} else if (flags & PIPE_FLUSH_DEFERRED) {
+		fd_bc_flush_deferred(&ctx->screen->batch_cache, ctx);
+	} else {
+		fd_bc_flush(&ctx->screen->batch_cache, ctx);
+	}
 
-	fd_ringbuffer_set_parent(ring, NULL);
-	ctx->ring = ring;
+	fd_bc_dump(ctx->screen, "%p: remaining:\n", ctx);
 
-	fd_ringmarker_del(ctx->binning_start);
-	fd_ringmarker_del(ctx->binning_end);
+out:
+	if (fencep)
+		fd_fence_ref(fencep, fence);
 
-	ring = next_rb(ctx);
+	fd_fence_ref(&ctx->last_fence, fence);
 
-	ctx->binning_start = fd_ringmarker_new(ring);
-	ctx->binning_end = fd_ringmarker_new(ring);
+	fd_fence_ref(&fence, NULL);
 
-	fd_ringbuffer_set_parent(ring, ctx->ring);
-	ctx->binning_ring = ring;
+	if (flags & PIPE_FLUSH_END_OF_FRAME)
+		fd_log_eof(ctx);
 }
 
-/* emit accumulated render cmds, needed for example if render target has
- * changed, or for flush()
- */
-void
-fd_context_render(struct pipe_context *pctx)
+static void
+fd_texture_barrier(struct pipe_context *pctx, unsigned flags)
 {
-	struct fd_context *ctx = fd_context(pctx);
-	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
+	if (flags == PIPE_TEXTURE_BARRIER_FRAMEBUFFER) {
+		struct fd_context *ctx = fd_context(pctx);
 
-	DBG("needs_flush: %d", ctx->needs_flush);
+		if (ctx->framebuffer_barrier) {
+			ctx->framebuffer_barrier(ctx);
+			return;
+		}
+	}
+
+	/* On devices that could sample from GMEM we could possibly do better.
+	 * Or if we knew that we were doing GMEM bypass we could just emit a
+	 * cache flush, perhaps?  But we don't know if future draws would cause
+	 * us to use GMEM, and a flush in bypass isn't the end of the world.
+	 */
+	fd_context_flush(pctx, NULL, 0);
+}
 
-	if (!ctx->needs_flush)
+static void
+fd_memory_barrier(struct pipe_context *pctx, unsigned flags)
+{
+	if (!(flags & ~PIPE_BARRIER_UPDATE))
 		return;
 
-	fd_gmem_render_tiles(pctx);
+	fd_context_flush(pctx, NULL, 0);
+	/* TODO do we need to check for persistently mapped buffers and fd_bo_cpu_prep()?? */
+}
+
+static void
+emit_string_tail(struct fd_ringbuffer *ring, const char *string, int len)
+{
+	const uint32_t *buf = (const void *)string;
 
-	DBG("%p/%p/%p", ctx->ring->start, ctx->ring->cur, ctx->ring->end);
+	while (len >= 4) {
+		OUT_RING(ring, *buf);
+		buf++;
+		len -= 4;
+	}
 
-	/* if size in dwords is more than half the buffer size, then wait and
-	 * wrap around:
-	 */
-	if ((ctx->ring->cur - ctx->ring->start) > ctx->ring->size/8)
-		fd_context_next_rb(pctx);
+	/* copy remainder bytes without reading past end of input string: */
+	if (len > 0) {
+		uint32_t w = 0;
+		memcpy(&w, buf, len);
+		OUT_RING(ring, w);
+	}
+}
 
-	ctx->needs_flush = false;
-	ctx->cleared = ctx->restore = ctx->resolve = 0;
-	ctx->gmem_reason = 0;
-	ctx->num_draws = 0;
+/* for prior to a5xx: */
+void
+fd_emit_string(struct fd_ringbuffer *ring,
+		const char *string, int len)
+{
+	/* max packet size is 0x3fff+1 dwords: */
+	len = MIN2(len, 0x4000 * 4);
+
+	OUT_PKT3(ring, CP_NOP, align(len, 4) / 4);
+	emit_string_tail(ring, string, len);
+}
+
+/* for a5xx+ */
+void
+fd_emit_string5(struct fd_ringbuffer *ring,
+		const char *string, int len)
+{
+	/* max packet size is 0x3fff dwords: */
+	len = MIN2(len, 0x3fff * 4);
 
-	if (pfb->cbufs[0])
-		fd_resource(pfb->cbufs[0]->texture)->dirty = false;
-	if (pfb->zsbuf)
-		fd_resource(pfb->zsbuf->texture)->dirty = false;
+	OUT_PKT7(ring, CP_NOP, align(len, 4) / 4);
+	emit_string_tail(ring, string, len);
 }
 
+/**
+ * emit marker string as payload of a no-op packet, which can be
+ * decoded by cffdump.
+ */
 static void
-fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fence,
-		unsigned flags)
+fd_emit_string_marker(struct pipe_context *pctx, const char *string, int len)
 {
-	DBG("fence=%p", fence);
+	struct fd_context *ctx = fd_context(pctx);
 
-#if 0
-	if (fence) {
-		fd_fence_ref(ctx->screen->fence.current,
-				(struct fd_fence **)fence);
-	}
-#endif
+	if (!ctx->batch)
+		return;
+
+	ctx->batch->needs_flush = true;
 
-	fd_context_render(pctx);
+	if (ctx->screen->gpu_id >= 500) {
+		fd_emit_string5(ctx->batch->draw, string, len);
+	} else {
+		fd_emit_string(ctx->batch->draw, string, len);
+	}
 }
 
 void
@@ -142,43 +211,199 @@ fd_context_destroy(struct pipe_context *pctx)
 
 	DBG("");
 
-	util_slab_destroy(&ctx->transfer_pool);
+	fd_screen_lock(ctx->screen);
+	list_del(&ctx->node);
+	fd_screen_unlock(ctx->screen);
+
+	fd_log_process(ctx, true);
+	assert(list_is_empty(&ctx->log_chunks));
+
+	fd_fence_ref(&ctx->last_fence, NULL);
+
+	util_copy_framebuffer_state(&ctx->framebuffer, NULL);
+	fd_batch_reference(&ctx->batch, NULL);  /* unref current batch */
+	fd_bc_invalidate_context(ctx);
 
-	util_dynarray_fini(&ctx->draw_patches);
+	fd_prog_fini(pctx);
 
 	if (ctx->blitter)
 		util_blitter_destroy(ctx->blitter);
 
+	if (pctx->stream_uploader)
+		u_upload_destroy(pctx->stream_uploader);
+
+	if (ctx->clear_rs_state)
+		pctx->delete_rasterizer_state(pctx, ctx->clear_rs_state);
+
 	if (ctx->primconvert)
 		util_primconvert_destroy(ctx->primconvert);
 
-	fd_ringmarker_del(ctx->draw_start);
-	fd_ringmarker_del(ctx->draw_end);
-	fd_ringmarker_del(ctx->binning_start);
-	fd_ringmarker_del(ctx->binning_end);
-
-	for (i = 0; i < ARRAY_SIZE(ctx->rings); i++)
-		fd_ringbuffer_del(ctx->rings[i]);
+	slab_destroy_child(&ctx->transfer_pool);
 
-	for (i = 0; i < ARRAY_SIZE(ctx->pipe); i++) {
-		struct fd_vsc_pipe *pipe = &ctx->pipe[i];
-		if (!pipe->bo)
+	for (i = 0; i < ARRAY_SIZE(ctx->vsc_pipe_bo); i++) {
+		if (!ctx->vsc_pipe_bo[i])
 			break;
-		fd_bo_del(pipe->bo);
+		fd_bo_del(ctx->vsc_pipe_bo[i]);
 	}
 
-	FREE(ctx);
+	fd_device_del(ctx->dev);
+	fd_pipe_del(ctx->pipe);
+
+	mtx_destroy(&ctx->gmem_lock);
+
+	if (fd_mesa_debug & (FD_DBG_BSTAT | FD_DBG_MSGS)) {
+		printf("batch_total=%u, batch_sysmem=%u, batch_gmem=%u, batch_nondraw=%u, batch_restore=%u\n",
+			(uint32_t)ctx->stats.batch_total, (uint32_t)ctx->stats.batch_sysmem,
+			(uint32_t)ctx->stats.batch_gmem, (uint32_t)ctx->stats.batch_nondraw,
+			(uint32_t)ctx->stats.batch_restore);
+	}
+}
+
+static void
+fd_set_debug_callback(struct pipe_context *pctx,
+		const struct pipe_debug_callback *cb)
+{
+	struct fd_context *ctx = fd_context(pctx);
+
+	if (cb)
+		ctx->debug = *cb;
+	else
+		memset(&ctx->debug, 0, sizeof(ctx->debug));
+}
+
+static uint32_t
+fd_get_reset_count(struct fd_context *ctx, bool per_context)
+{
+	uint64_t val;
+	enum fd_param_id param =
+		per_context ? FD_CTX_FAULTS : FD_GLOBAL_FAULTS;
+	int ret = fd_pipe_get_param(ctx->pipe, param, &val);
+	debug_assert(!ret);
+	return val;
+}
+
+static enum pipe_reset_status
+fd_get_device_reset_status(struct pipe_context *pctx)
+{
+	struct fd_context *ctx = fd_context(pctx);
+	int context_faults = fd_get_reset_count(ctx, true);
+	int global_faults  = fd_get_reset_count(ctx, false);
+	enum pipe_reset_status status;
+
+	if (context_faults != ctx->context_reset_count) {
+		status = PIPE_GUILTY_CONTEXT_RESET;
+	} else if (global_faults != ctx->global_reset_count) {
+		status = PIPE_INNOCENT_CONTEXT_RESET;
+	} else {
+		status = PIPE_NO_RESET;
+	}
+
+	ctx->context_reset_count = context_faults;
+	ctx->global_reset_count = global_faults;
+
+	return status;
+}
+
+/* TODO we could combine a few of these small buffers (solid_vbuf,
+ * blit_texcoord_vbuf, and vsc_size_mem, into a single buffer and
+ * save a tiny bit of memory
+ */
+
+static struct pipe_resource *
+create_solid_vertexbuf(struct pipe_context *pctx)
+{
+	static const float init_shader_const[] = {
+			-1.000000, +1.000000, +1.000000,
+			+1.000000, -1.000000, +1.000000,
+	};
+	struct pipe_resource *prsc = pipe_buffer_create(pctx->screen,
+			PIPE_BIND_CUSTOM, PIPE_USAGE_IMMUTABLE, sizeof(init_shader_const));
+	pipe_buffer_write(pctx, prsc, 0,
+			sizeof(init_shader_const), init_shader_const);
+	return prsc;
+}
+
+static struct pipe_resource *
+create_blit_texcoord_vertexbuf(struct pipe_context *pctx)
+{
+	struct pipe_resource *prsc = pipe_buffer_create(pctx->screen,
+			PIPE_BIND_CUSTOM, PIPE_USAGE_DYNAMIC, 16);
+	return prsc;
+}
+
+void
+fd_context_setup_common_vbos(struct fd_context *ctx)
+{
+	struct pipe_context *pctx = &ctx->base;
+
+	ctx->solid_vbuf = create_solid_vertexbuf(pctx);
+	ctx->blit_texcoord_vbuf = create_blit_texcoord_vertexbuf(pctx);
+
+	/* setup solid_vbuf_state: */
+	ctx->solid_vbuf_state.vtx = pctx->create_vertex_elements_state(
+			pctx, 1, (struct pipe_vertex_element[]){{
+				.vertex_buffer_index = 0,
+				.src_offset = 0,
+				.src_format = PIPE_FORMAT_R32G32B32_FLOAT,
+			}});
+	ctx->solid_vbuf_state.vertexbuf.count = 1;
+	ctx->solid_vbuf_state.vertexbuf.vb[0].stride = 12;
+	ctx->solid_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->solid_vbuf;
+
+	/* setup blit_vbuf_state: */
+	ctx->blit_vbuf_state.vtx = pctx->create_vertex_elements_state(
+			pctx, 2, (struct pipe_vertex_element[]){{
+				.vertex_buffer_index = 0,
+				.src_offset = 0,
+				.src_format = PIPE_FORMAT_R32G32_FLOAT,
+			}, {
+				.vertex_buffer_index = 1,
+				.src_offset = 0,
+				.src_format = PIPE_FORMAT_R32G32B32_FLOAT,
+			}});
+	ctx->blit_vbuf_state.vertexbuf.count = 2;
+	ctx->blit_vbuf_state.vertexbuf.vb[0].stride = 8;
+	ctx->blit_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->blit_texcoord_vbuf;
+	ctx->blit_vbuf_state.vertexbuf.vb[1].stride = 12;
+	ctx->blit_vbuf_state.vertexbuf.vb[1].buffer.resource = ctx->solid_vbuf;
+}
+
+void
+fd_context_cleanup_common_vbos(struct fd_context *ctx)
+{
+	struct pipe_context *pctx = &ctx->base;
+
+	pctx->delete_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
+	pctx->delete_vertex_elements_state(pctx, ctx->blit_vbuf_state.vtx);
+
+	pipe_resource_reference(&ctx->solid_vbuf, NULL);
+	pipe_resource_reference(&ctx->blit_texcoord_vbuf, NULL);
 }
 
 struct pipe_context *
 fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
-		const uint8_t *primtypes, void *priv)
+		const uint8_t *primtypes, void *priv, unsigned flags)
 {
 	struct fd_screen *screen = fd_screen(pscreen);
 	struct pipe_context *pctx;
+	unsigned prio = 1;
 	int i;
 
+	/* lower numerical value == higher priority: */
+	if (fd_mesa_debug & FD_DBG_HIPRIO)
+		prio = 0;
+	else if (flags & PIPE_CONTEXT_HIGH_PRIORITY)
+		prio = 0;
+	else if (flags & PIPE_CONTEXT_LOW_PRIORITY)
+		prio = 2;
+
 	ctx->screen = screen;
+	ctx->pipe = fd_pipe_new2(screen->dev, FD_PIPE_3D, prio);
+
+	if (fd_device_version(screen->dev) >= FD_VERSION_ROBUSTNESS) {
+		ctx->context_reset_count = fd_get_reset_count(ctx, true);
+		ctx->global_reset_count = fd_get_reset_count(ctx, false);
+	}
 
 	ctx->primtypes = primtypes;
 	ctx->primtype_mask = 0;
@@ -186,32 +411,37 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
 		if (primtypes[i])
 			ctx->primtype_mask |= (1 << i);
 
-	/* need some sane default in case state tracker doesn't
+	(void) mtx_init(&ctx->gmem_lock, mtx_plain);
+
+	/* need some sane default in case gallium frontends don't
 	 * set some state:
 	 */
 	ctx->sample_mask = 0xffff;
+	ctx->active_queries = true;
 
 	pctx = &ctx->base;
 	pctx->screen = pscreen;
 	pctx->priv = priv;
 	pctx->flush = fd_context_flush;
+	pctx->emit_string_marker = fd_emit_string_marker;
+	pctx->set_debug_callback = fd_set_debug_callback;
+	pctx->get_device_reset_status = fd_get_device_reset_status;
+	pctx->create_fence_fd = fd_create_fence_fd;
+	pctx->fence_server_sync = fd_fence_server_sync;
+	pctx->fence_server_signal = fd_fence_server_signal;
+	pctx->texture_barrier = fd_texture_barrier;
+	pctx->memory_barrier = fd_memory_barrier;
+
+	pctx->stream_uploader = u_upload_create_default(pctx);
+	if (!pctx->stream_uploader)
+		goto fail;
+	pctx->const_uploader = pctx->stream_uploader;
 
-	for (i = 0; i < ARRAY_SIZE(ctx->rings); i++) {
-		ctx->rings[i] = fd_ringbuffer_new(screen->pipe, 0x400000);
-		if (!ctx->rings[i])
-			goto fail;
-	}
-
-	fd_context_next_rb(pctx);
-	fd_reset_rmw_state(ctx);
-
-	util_dynarray_init(&ctx->draw_patches);
-
-	util_slab_create(&ctx->transfer_pool, sizeof(struct pipe_transfer),
-			16, UTIL_SLAB_SINGLETHREADED);
+	slab_create_child(&ctx->transfer_pool, &screen->transfer_pool);
 
 	fd_draw_init(pctx);
 	fd_resource_context_init(pctx);
+	fd_query_context_init(pctx);
 	fd_texture_init(pctx);
 	fd_state_init(pctx);
 
@@ -223,6 +453,36 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
 	if (!ctx->primconvert)
 		goto fail;
 
+	list_inithead(&ctx->hw_active_queries);
+	list_inithead(&ctx->acc_active_queries);
+	list_inithead(&ctx->log_chunks);
+
+	fd_screen_lock(ctx->screen);
+	list_add(&ctx->node, &ctx->screen->context_list);
+	fd_screen_unlock(ctx->screen);
+
+	ctx->current_scissor = &ctx->disabled_scissor;
+
+	ctx->log_out = stdout;
+
+	if ((fd_mesa_debug & FD_DBG_LOG) &&
+			!(ctx->record_timestamp && ctx->ts_to_ns)) {
+		printf("logging not supported!\n");
+		fd_mesa_debug &= ~FD_DBG_LOG;
+	}
+
+#if DETECT_OS_ANDROID
+	if (fd_mesa_debug & FD_DBG_LOG) {
+		static unsigned idx = 0;
+		char *p;
+		asprintf(&p, "/data/fdlog/%s-%d.log", util_get_process_name(), idx++);
+
+		FILE *f = fopen(p, "w");
+		if (f)
+			ctx->log_out = f;
+	}
+#endif
+
 	return pctx;
 
 fail: