#include "util/os_time.h"
#include "util/u_memory.h"
#include "util/u_queue.h"
+#include "util/u_upload_mgr.h"
#include "si_pipe.h"
+#include "radeon/r600_cs.h"
+
+struct si_fine_fence {
+ struct r600_resource *buf;
+ unsigned offset;
+};
struct si_multi_fence {
struct pipe_reference reference;
/* If the context wasn't flushed at fence creation, this is non-NULL. */
struct {
- struct r600_common_context *ctx;
+ struct si_context *ctx;
unsigned ib_index;
} gfx_unflushed;
+
+ struct si_fine_fence fine;
};
-static void si_add_fence_dependency(struct r600_common_context *rctx,
+/**
+ * Write an EOP event.
+ *
+ * \param event EVENT_TYPE_*
+ * \param event_flags Optional cache flush flags (TC)
+ * \param data_sel 1 = fence, 3 = timestamp
+ * \param buf Buffer
+ * \param va GPU address
+ * \param old_value Previous fence value (for a bug workaround)
+ * \param new_value Fence value to write for this event.
+ */
+void si_gfx_write_event_eop(struct si_context *ctx,
+ unsigned event, unsigned event_flags,
+ unsigned data_sel,
+ struct r600_resource *buf, uint64_t va,
+ uint32_t new_fence, unsigned query_type)
+{
+ struct radeon_winsys_cs *cs = ctx->b.gfx_cs;
+ unsigned op = EVENT_TYPE(event) |
+ EVENT_INDEX(5) |
+ event_flags;
+ unsigned sel = EOP_DATA_SEL(data_sel);
+
+ /* Wait for write confirmation before writing data, but don't send
+ * an interrupt. */
+ if (data_sel != EOP_DATA_SEL_DISCARD)
+ sel |= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM);
+
+ if (ctx->b.chip_class >= GFX9) {
+ /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
+ * counters) must immediately precede every timestamp event to
+ * prevent a GPU hang on GFX9.
+ *
+ * Occlusion queries don't need to do it here, because they
+ * always do ZPASS_DONE before the timestamp.
+ */
+ if (ctx->b.chip_class == GFX9 &&
+ query_type != PIPE_QUERY_OCCLUSION_COUNTER &&
+ query_type != PIPE_QUERY_OCCLUSION_PREDICATE &&
+ query_type != PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
+ struct r600_resource *scratch = ctx->b.eop_bug_scratch;
+
+ assert(16 * ctx->b.screen->info.num_render_backends <=
+ scratch->b.b.width0);
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
+ radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
+ radeon_emit(cs, scratch->gpu_address);
+ radeon_emit(cs, scratch->gpu_address >> 32);
+
+ radeon_add_to_buffer_list(&ctx->b, ctx->b.gfx_cs, scratch,
+ RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
+ }
+
+ radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, 6, 0));
+ radeon_emit(cs, op);
+ radeon_emit(cs, sel);
+ radeon_emit(cs, va); /* address lo */
+ radeon_emit(cs, va >> 32); /* address hi */
+ radeon_emit(cs, new_fence); /* immediate data lo */
+ radeon_emit(cs, 0); /* immediate data hi */
+ radeon_emit(cs, 0); /* unused */
+ } else {
+ if (ctx->b.chip_class == CIK ||
+ ctx->b.chip_class == VI) {
+ struct r600_resource *scratch = ctx->b.eop_bug_scratch;
+ uint64_t va = scratch->gpu_address;
+
+ /* Two EOP events are required to make all engines go idle
+ * (and optional cache flushes executed) before the timestamp
+ * is written.
+ */
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+ radeon_emit(cs, op);
+ radeon_emit(cs, va);
+ radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
+ radeon_emit(cs, 0); /* immediate data */
+ radeon_emit(cs, 0); /* unused */
+
+ radeon_add_to_buffer_list(&ctx->b, ctx->b.gfx_cs, scratch,
+ RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
+ }
+
+ radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
+ radeon_emit(cs, op);
+ radeon_emit(cs, va);
+ radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
+ radeon_emit(cs, new_fence); /* immediate data */
+ radeon_emit(cs, 0); /* unused */
+ }
+
+ if (buf) {
+ radeon_add_to_buffer_list(&ctx->b, ctx->b.gfx_cs, buf, RADEON_USAGE_WRITE,
+ RADEON_PRIO_QUERY);
+ }
+}
+
+unsigned si_gfx_write_fence_dwords(struct si_screen *screen)
+{
+ unsigned dwords = 6;
+
+ if (screen->info.chip_class == CIK ||
+ screen->info.chip_class == VI)
+ dwords *= 2;
+
+ return dwords;
+}
+
+void si_gfx_wait_fence(struct si_context *ctx,
+ uint64_t va, uint32_t ref, uint32_t mask)
+{
+ struct radeon_winsys_cs *cs = ctx->b.gfx_cs;
+
+ radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
+ radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
+ radeon_emit(cs, va);
+ radeon_emit(cs, va >> 32);
+ radeon_emit(cs, ref); /* reference value */
+ radeon_emit(cs, mask); /* mask */
+ radeon_emit(cs, 4); /* poll interval */
+}
+
+static void si_add_fence_dependency(struct si_context *sctx,
struct pipe_fence_handle *fence)
{
- struct radeon_winsys *ws = rctx->ws;
+ struct radeon_winsys *ws = sctx->b.ws;
- if (rctx->dma.cs)
- ws->cs_add_fence_dependency(rctx->dma.cs, fence);
- ws->cs_add_fence_dependency(rctx->gfx.cs, fence);
+ if (sctx->b.dma_cs)
+ ws->cs_add_fence_dependency(sctx->b.dma_cs, fence);
+ ws->cs_add_fence_dependency(sctx->b.gfx_cs, fence);
+}
+
+static void si_add_syncobj_signal(struct si_context *sctx,
+ struct pipe_fence_handle *fence)
+{
+ sctx->b.ws->cs_add_syncobj_signal(sctx->b.gfx_cs, fence);
}
static void si_fence_reference(struct pipe_screen *screen,
struct pipe_fence_handle **dst,
struct pipe_fence_handle *src)
{
- struct radeon_winsys *ws = ((struct r600_common_screen*)screen)->ws;
+ struct radeon_winsys *ws = ((struct si_screen*)screen)->ws;
struct si_multi_fence **rdst = (struct si_multi_fence **)dst;
struct si_multi_fence *rsrc = (struct si_multi_fence *)src;
ws->fence_reference(&(*rdst)->gfx, NULL);
ws->fence_reference(&(*rdst)->sdma, NULL);
tc_unflushed_batch_token_reference(&(*rdst)->tc_token, NULL);
+ r600_resource_reference(&(*rdst)->fine.buf, NULL);
FREE(*rdst);
}
*rdst = rsrc;
return (struct pipe_fence_handle *)fence;
}
-static void si_fence_server_sync(struct pipe_context *ctx,
- struct pipe_fence_handle *fence)
+static bool si_fine_fence_signaled(struct radeon_winsys *rws,
+ const struct si_fine_fence *fine)
{
- struct r600_common_context *rctx = (struct r600_common_context *)ctx;
- struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+ char *map = rws->buffer_map(fine->buf->buf, NULL, PIPE_TRANSFER_READ |
+ PIPE_TRANSFER_UNSYNCHRONIZED);
+ if (!map)
+ return false;
- util_queue_fence_wait(&rfence->ready);
+ uint32_t *fence = (uint32_t*)(map + fine->offset);
+ return *fence != 0;
+}
- /* Unflushed fences from the same context are no-ops. */
- if (rfence->gfx_unflushed.ctx &&
- rfence->gfx_unflushed.ctx == rctx)
+static void si_fine_fence_set(struct si_context *ctx,
+ struct si_fine_fence *fine,
+ unsigned flags)
+{
+ uint32_t *fence_ptr;
+
+ assert(util_bitcount(flags & (PIPE_FLUSH_TOP_OF_PIPE | PIPE_FLUSH_BOTTOM_OF_PIPE)) == 1);
+
+ /* Use uncached system memory for the fence. */
+ u_upload_alloc(ctx->b.cached_gtt_allocator, 0, 4, 4,
+ &fine->offset, (struct pipe_resource **)&fine->buf, (void **)&fence_ptr);
+ if (!fine->buf)
return;
- /* All unflushed commands will not start execution before
- * this fence dependency is signalled.
- *
- * Should we flush the context to allow more GPU parallelism?
- */
- if (rfence->sdma)
- si_add_fence_dependency(rctx, rfence->sdma);
- if (rfence->gfx)
- si_add_fence_dependency(rctx, rfence->gfx);
+ *fence_ptr = 0;
+
+ uint64_t fence_va = fine->buf->gpu_address + fine->offset;
+
+ radeon_add_to_buffer_list(&ctx->b, ctx->b.gfx_cs, fine->buf,
+ RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
+ if (flags & PIPE_FLUSH_TOP_OF_PIPE) {
+ struct radeon_winsys_cs *cs = ctx->b.gfx_cs;
+ radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
+ radeon_emit(cs, S_370_DST_SEL(V_370_MEM_ASYNC) |
+ S_370_WR_CONFIRM(1) |
+ S_370_ENGINE_SEL(V_370_PFP));
+ radeon_emit(cs, fence_va);
+ radeon_emit(cs, fence_va >> 32);
+ radeon_emit(cs, 0x80000000);
+ } else if (flags & PIPE_FLUSH_BOTTOM_OF_PIPE) {
+ si_gfx_write_event_eop(ctx, V_028A90_BOTTOM_OF_PIPE_TS, 0,
+ EOP_DATA_SEL_VALUE_32BIT,
+ NULL, fence_va, 0x80000000,
+ PIPE_QUERY_GPU_FINISHED);
+ } else {
+ assert(false);
+ }
}
static boolean si_fence_finish(struct pipe_screen *screen,
struct pipe_fence_handle *fence,
uint64_t timeout)
{
- struct radeon_winsys *rws = ((struct r600_common_screen*)screen)->ws;
+ struct radeon_winsys *rws = ((struct si_screen*)screen)->ws;
struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
- struct r600_common_context *rctx;
int64_t abs_timeout = os_time_get_absolute_timeout(timeout);
- ctx = threaded_context_unwrap_sync(ctx);
- rctx = ctx ? (struct r600_common_context*)ctx : NULL;
-
if (!util_queue_fence_is_signalled(&rfence->ready)) {
- if (!timeout)
- return false;
-
if (rfence->tc_token) {
/* Ensure that si_flush_from_st will be called for
* this fence, but only if we're in the API thread
* be in flight in the driver thread, so the fence
* may not be ready yet when this call returns.
*/
- threaded_context_flush(ctx, rfence->tc_token);
+ threaded_context_flush(ctx, rfence->tc_token,
+ timeout == 0);
}
+ if (!timeout)
+ return false;
+
if (timeout == PIPE_TIMEOUT_INFINITE) {
util_queue_fence_wait(&rfence->ready);
} else {
if (!util_queue_fence_wait_timeout(&rfence->ready, abs_timeout))
return false;
}
+
+ if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
+ int64_t time = os_time_get_nano();
+ timeout = abs_timeout > time ? abs_timeout - time : 0;
+ }
}
if (rfence->sdma) {
if (!rfence->gfx)
return true;
+ if (rfence->fine.buf &&
+ si_fine_fence_signaled(rws, &rfence->fine)) {
+ rws->fence_reference(&rfence->gfx, NULL);
+ r600_resource_reference(&rfence->fine.buf, NULL);
+ return true;
+ }
+
/* Flush the gfx IB if it hasn't been flushed yet. */
- if (rctx &&
- rfence->gfx_unflushed.ctx == rctx &&
- rfence->gfx_unflushed.ib_index == rctx->num_gfx_cs_flushes) {
- rctx->gfx.flush(rctx, timeout ? 0 : RADEON_FLUSH_ASYNC, NULL);
- rfence->gfx_unflushed.ctx = NULL;
+ if (ctx && rfence->gfx_unflushed.ctx) {
+ struct si_context *sctx;
+
+ sctx = (struct si_context *)threaded_context_unwrap_unsync(ctx);
+ if (rfence->gfx_unflushed.ctx == sctx &&
+ rfence->gfx_unflushed.ib_index == sctx->b.num_gfx_cs_flushes) {
+ /* Section 4.1.2 (Signaling) of the OpenGL 4.6 (Core profile)
+ * spec says:
+ *
+ * "If the sync object being blocked upon will not be
+ * signaled in finite time (for example, by an associated
+ * fence command issued previously, but not yet flushed to
+ * the graphics pipeline), then ClientWaitSync may hang
+ * forever. To help prevent this behavior, if
+ * ClientWaitSync is called and all of the following are
+ * true:
+ *
+ * * the SYNC_FLUSH_COMMANDS_BIT bit is set in flags,
+ * * sync is unsignaled when ClientWaitSync is called,
+ * * and the calls to ClientWaitSync and FenceSync were
+ * issued from the same context,
+ *
+ * then the GL will behave as if the equivalent of Flush
+ * were inserted immediately after the creation of sync."
+ *
+ * This means we need to flush for such fences even when we're
+ * not going to wait.
+ */
+ threaded_context_unwrap_sync(ctx);
+ si_flush_gfx_cs(sctx, timeout ? 0 : PIPE_FLUSH_ASYNC, NULL);
+ rfence->gfx_unflushed.ctx = NULL;
- if (!timeout)
- return false;
+ if (!timeout)
+ return false;
- /* Recompute the timeout after all that. */
- if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
- int64_t time = os_time_get_nano();
- timeout = abs_timeout > time ? abs_timeout - time : 0;
+ /* Recompute the timeout after all that. */
+ if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
+ int64_t time = os_time_get_nano();
+ timeout = abs_timeout > time ? abs_timeout - time : 0;
+ }
}
}
- return rws->fence_wait(rws, rfence->gfx, timeout);
+ if (rws->fence_wait(rws, rfence->gfx, timeout))
+ return true;
+
+ /* Re-check in case the GPU is slow or hangs, but the commands before
+ * the fine-grained fence have completed. */
+ if (rfence->fine.buf &&
+ si_fine_fence_signaled(rws, &rfence->fine))
+ return true;
+
+ return false;
}
static void si_create_fence_fd(struct pipe_context *ctx,
- struct pipe_fence_handle **pfence, int fd)
+ struct pipe_fence_handle **pfence, int fd,
+ enum pipe_fd_type type)
{
- struct r600_common_screen *rscreen = (struct r600_common_screen*)ctx->screen;
- struct radeon_winsys *ws = rscreen->ws;
+ struct si_screen *sscreen = (struct si_screen*)ctx->screen;
+ struct radeon_winsys *ws = sscreen->ws;
struct si_multi_fence *rfence;
*pfence = NULL;
- if (!rscreen->info.has_sync_file)
- return;
-
rfence = si_create_multi_fence();
if (!rfence)
return;
- rfence->gfx = ws->fence_import_sync_file(ws, fd);
+ switch (type) {
+ case PIPE_FD_TYPE_NATIVE_SYNC:
+ if (!sscreen->info.has_fence_to_handle)
+ goto finish;
+
+ rfence->gfx = ws->fence_import_sync_file(ws, fd);
+ break;
+
+ case PIPE_FD_TYPE_SYNCOBJ:
+ if (!sscreen->info.has_syncobj)
+ goto finish;
+
+ rfence->gfx = ws->fence_import_syncobj(ws, fd);
+ break;
+
+ default:
+ unreachable("bad fence fd type when importing");
+ }
+
+finish:
if (!rfence->gfx) {
FREE(rfence);
return;
static int si_fence_get_fd(struct pipe_screen *screen,
struct pipe_fence_handle *fence)
{
- struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
- struct radeon_winsys *ws = rscreen->ws;
+ struct si_screen *sscreen = (struct si_screen*)screen;
+ struct radeon_winsys *ws = sscreen->ws;
struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
int gfx_fd = -1, sdma_fd = -1;
- if (!rscreen->info.has_sync_file)
+ if (!sscreen->info.has_fence_to_handle)
return -1;
util_queue_fence_wait(&rfence->ready);
/* If we don't have FDs at this point, it means we don't have fences
* either. */
+ if (sdma_fd == -1 && gfx_fd == -1)
+ return ws->export_signalled_sync_file(ws);
if (sdma_fd == -1)
return gfx_fd;
if (gfx_fd == -1)
unsigned flags)
{
struct pipe_screen *screen = ctx->screen;
- struct r600_common_context *rctx = (struct r600_common_context *)ctx;
- struct radeon_winsys *ws = rctx->ws;
+ struct si_context *sctx = (struct si_context *)ctx;
+ struct radeon_winsys *ws = sctx->b.ws;
struct pipe_fence_handle *gfx_fence = NULL;
struct pipe_fence_handle *sdma_fence = NULL;
bool deferred_fence = false;
- unsigned rflags = RADEON_FLUSH_ASYNC;
+ struct si_fine_fence fine = {};
+ unsigned rflags = PIPE_FLUSH_ASYNC;
if (flags & PIPE_FLUSH_END_OF_FRAME)
- rflags |= RADEON_FLUSH_END_OF_FRAME;
+ rflags |= PIPE_FLUSH_END_OF_FRAME;
+
+ if (flags & (PIPE_FLUSH_TOP_OF_PIPE | PIPE_FLUSH_BOTTOM_OF_PIPE)) {
+ assert(flags & PIPE_FLUSH_DEFERRED);
+ assert(fence);
+
+ si_fine_fence_set(sctx, &fine, flags);
+ }
/* DMA IBs are preambles to gfx IBs, therefore must be flushed first. */
- if (rctx->dma.cs)
- rctx->dma.flush(rctx, rflags, fence ? &sdma_fence : NULL);
+ if (sctx->b.dma_cs)
+ si_flush_dma_cs(sctx, rflags, fence ? &sdma_fence : NULL);
- if (!radeon_emitted(rctx->gfx.cs, rctx->initial_gfx_cs_size)) {
+ if (!radeon_emitted(sctx->b.gfx_cs, sctx->b.initial_gfx_cs_size)) {
if (fence)
- ws->fence_reference(&gfx_fence, rctx->last_gfx_fence);
+ ws->fence_reference(&gfx_fence, sctx->b.last_gfx_fence);
if (!(flags & PIPE_FLUSH_DEFERRED))
- ws->cs_sync_flush(rctx->gfx.cs);
+ ws->cs_sync_flush(sctx->b.gfx_cs);
} else {
/* Instead of flushing, create a deferred fence. Constraints:
* - The state tracker must allow a deferred flush.
if (flags & PIPE_FLUSH_DEFERRED &&
!(flags & PIPE_FLUSH_FENCE_FD) &&
fence) {
- gfx_fence = rctx->ws->cs_get_next_fence(rctx->gfx.cs);
+ gfx_fence = sctx->b.ws->cs_get_next_fence(sctx->b.gfx_cs);
deferred_fence = true;
} else {
- rctx->gfx.flush(rctx, rflags, fence ? &gfx_fence : NULL);
+ si_flush_gfx_cs(sctx, rflags, fence ? &gfx_fence : NULL);
}
}
multi_fence->sdma = sdma_fence;
if (deferred_fence) {
- multi_fence->gfx_unflushed.ctx = rctx;
- multi_fence->gfx_unflushed.ib_index = rctx->num_gfx_cs_flushes;
+ multi_fence->gfx_unflushed.ctx = sctx;
+ multi_fence->gfx_unflushed.ib_index = sctx->b.num_gfx_cs_flushes;
}
+ multi_fence->fine = fine;
+ fine.buf = NULL;
+
if (flags & TC_FLUSH_ASYNC) {
util_queue_fence_signal(&multi_fence->ready);
tc_unflushed_batch_token_reference(&multi_fence->tc_token, NULL);
}
}
+ assert(!fine.buf);
finish:
if (!(flags & PIPE_FLUSH_DEFERRED)) {
- if (rctx->dma.cs)
- ws->cs_sync_flush(rctx->dma.cs);
- ws->cs_sync_flush(rctx->gfx.cs);
+ if (sctx->b.dma_cs)
+ ws->cs_sync_flush(sctx->b.dma_cs);
+ ws->cs_sync_flush(sctx->b.gfx_cs);
}
}
+static void si_fence_server_signal(struct pipe_context *ctx,
+ struct pipe_fence_handle *fence)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+ struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+
+ /* We should have at least one syncobj to signal */
+ assert(rfence->sdma || rfence->gfx);
+
+ if (rfence->sdma)
+ si_add_syncobj_signal(sctx, rfence->sdma);
+ if (rfence->gfx)
+ si_add_syncobj_signal(sctx, rfence->gfx);
+
+ /**
+ * The spec does not require a flush here. We insert a flush
+ * because syncobj based signals are not directly placed into
+ * the command stream. Instead the signal happens when the
+ * submission associated with the syncobj finishes execution.
+ *
+ * Therefore, we must make sure that we flush the pipe to avoid
+ * new work being emitted and getting executed before the signal
+ * operation.
+ */
+ si_flush_from_st(ctx, NULL, PIPE_FLUSH_ASYNC);
+}
+
+static void si_fence_server_sync(struct pipe_context *ctx,
+ struct pipe_fence_handle *fence)
+{
+ struct si_context *sctx = (struct si_context *)ctx;
+ struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+
+ util_queue_fence_wait(&rfence->ready);
+
+ /* Unflushed fences from the same context are no-ops. */
+ if (rfence->gfx_unflushed.ctx &&
+ rfence->gfx_unflushed.ctx == sctx)
+ return;
+
+ /* All unflushed commands will not start execution before
+ * this fence dependency is signalled.
+ *
+ * Therefore we must flush before inserting the dependency
+ */
+ si_flush_from_st(ctx, NULL, PIPE_FLUSH_ASYNC);
+
+ if (rfence->sdma)
+ si_add_fence_dependency(sctx, rfence->sdma);
+ if (rfence->gfx)
+ si_add_fence_dependency(sctx, rfence->gfx);
+}
+
void si_init_fence_functions(struct si_context *ctx)
{
ctx->b.b.flush = si_flush_from_st;
ctx->b.b.create_fence_fd = si_create_fence_fd;
ctx->b.b.fence_server_sync = si_fence_server_sync;
+ ctx->b.b.fence_server_signal = si_fence_server_signal;
}
void si_init_screen_fence_functions(struct si_screen *screen)
{
- screen->b.b.fence_finish = si_fence_finish;
- screen->b.b.fence_reference = si_fence_reference;
- screen->b.b.fence_get_fd = si_fence_get_fd;
+ screen->b.fence_finish = si_fence_finish;
+ screen->b.fence_reference = si_fence_reference;
+ screen->b.fence_get_fd = si_fence_get_fd;
}