#include "si_build_pm4.h"
struct si_fine_fence {
- struct r600_resource *buf;
+ struct si_resource *buf;
unsigned offset;
};
*
* \param event EVENT_TYPE_*
* \param event_flags Optional cache flush flags (TC)
- * \param data_sel 1 = fence, 3 = timestamp
+ * \param dst_sel MEM or TC_L2
+ * \param int_sel NONE or SEND_DATA_AFTER_WR_CONFIRM
+ * \param data_sel DISCARD, VALUE_32BIT, TIMESTAMP, or GDS
* \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)
+void si_cp_release_mem(struct si_context *ctx, struct radeon_cmdbuf *cs,
+ unsigned event, unsigned event_flags,
+ unsigned dst_sel, unsigned int_sel, unsigned data_sel,
+ struct si_resource *buf, uint64_t va,
+ uint32_t new_fence, unsigned query_type)
{
- struct radeon_winsys_cs *cs = ctx->gfx_cs;
unsigned op = EVENT_TYPE(event) |
- EVENT_INDEX(5) |
+ EVENT_INDEX(event == V_028A90_CS_DONE ||
+ event == V_028A90_PS_DONE ? 6 : 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->chip_class >= GFX9) {
+ unsigned sel = EOP_DST_SEL(dst_sel) |
+ EOP_INT_SEL(int_sel) |
+ EOP_DATA_SEL(data_sel);
+ bool compute_ib = !ctx->has_graphics ||
+ cs == ctx->prim_discard_compute_cs;
+
+ if (ctx->chip_class >= GFX9 ||
+ (compute_ib && ctx->chip_class >= GFX7)) {
/* 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->chip_class == GFX9 &&
+ if (ctx->chip_class == GFX9 && !compute_ib &&
query_type != PIPE_QUERY_OCCLUSION_COUNTER &&
query_type != PIPE_QUERY_OCCLUSION_PREDICATE &&
query_type != PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
- struct r600_resource *scratch = ctx->eop_bug_scratch;
+ struct si_resource *scratch = ctx->eop_bug_scratch;
assert(16 * ctx->screen->info.num_render_backends <=
scratch->b.b.width0);
RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
}
- radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, 6, 0));
+ radeon_emit(cs, PKT3(PKT3_RELEASE_MEM, ctx->chip_class >= GFX9 ? 6 : 5, 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 */
+ if (ctx->chip_class >= GFX9)
+ radeon_emit(cs, 0); /* unused */
} else {
- if (ctx->chip_class == CIK ||
- ctx->chip_class == VI) {
- struct r600_resource *scratch = ctx->eop_bug_scratch;
+ if (ctx->chip_class == GFX7 ||
+ ctx->chip_class == GFX8) {
+ struct si_resource *scratch = ctx->eop_bug_scratch;
uint64_t va = scratch->gpu_address;
/* Two EOP events are required to make all engines go idle
}
}
-unsigned si_gfx_write_fence_dwords(struct si_screen *screen)
+unsigned si_cp_write_fence_dwords(struct si_screen *screen)
{
unsigned dwords = 6;
- if (screen->info.chip_class == CIK ||
- screen->info.chip_class == VI)
+ if (screen->info.chip_class == GFX7 ||
+ screen->info.chip_class == GFX8)
dwords *= 2;
return dwords;
}
-void si_gfx_wait_fence(struct si_context *ctx,
- uint64_t va, uint32_t ref, uint32_t mask)
+void si_cp_wait_mem(struct si_context *ctx, struct radeon_cmdbuf *cs,
+ uint64_t va, uint32_t ref, uint32_t mask, unsigned flags)
{
- struct radeon_winsys_cs *cs = ctx->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, WAIT_REG_MEM_MEM_SPACE(1) | flags);
radeon_emit(cs, va);
radeon_emit(cs, va >> 32);
radeon_emit(cs, ref); /* reference value */
{
struct radeon_winsys *ws = sctx->ws;
- if (sctx->dma_cs)
- ws->cs_add_fence_dependency(sctx->dma_cs, fence);
- ws->cs_add_fence_dependency(sctx->gfx_cs, fence);
+ if (sctx->sdma_cs)
+ ws->cs_add_fence_dependency(sctx->sdma_cs, fence, 0);
+ ws->cs_add_fence_dependency(sctx->gfx_cs, fence, 0);
}
static void si_add_syncobj_signal(struct si_context *sctx,
struct pipe_fence_handle *src)
{
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;
-
- if (pipe_reference(&(*rdst)->reference, &rsrc->reference)) {
- 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);
+ struct si_multi_fence **sdst = (struct si_multi_fence **)dst;
+ struct si_multi_fence *ssrc = (struct si_multi_fence *)src;
+
+ if (pipe_reference(&(*sdst)->reference, &ssrc->reference)) {
+ ws->fence_reference(&(*sdst)->gfx, NULL);
+ ws->fence_reference(&(*sdst)->sdma, NULL);
+ tc_unflushed_batch_token_reference(&(*sdst)->tc_token, NULL);
+ si_resource_reference(&(*sdst)->fine.buf, NULL);
+ FREE(*sdst);
}
- *rdst = rsrc;
+ *sdst = ssrc;
}
static struct si_multi_fence *si_create_multi_fence()
assert(util_bitcount(flags & (PIPE_FLUSH_TOP_OF_PIPE | PIPE_FLUSH_BOTTOM_OF_PIPE)) == 1);
- /* Use uncached system memory for the fence. */
+ /* Use cached system memory for the fence. */
u_upload_alloc(ctx->cached_gtt_allocator, 0, 4, 4,
&fine->offset, (struct pipe_resource **)&fine->buf, (void **)&fence_ptr);
if (!fine->buf)
*fence_ptr = 0;
- uint64_t fence_va = fine->buf->gpu_address + fine->offset;
-
- radeon_add_to_buffer_list(ctx, ctx->gfx_cs, fine->buf,
- RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
if (flags & PIPE_FLUSH_TOP_OF_PIPE) {
- struct radeon_winsys_cs *cs = ctx->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);
+ uint32_t value = 0x80000000;
+
+ si_cp_write_data(ctx, fine->buf, fine->offset, 4,
+ V_370_MEM, V_370_PFP, &value);
} 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);
+ uint64_t fence_va = fine->buf->gpu_address + fine->offset;
+
+ radeon_add_to_buffer_list(ctx, ctx->gfx_cs, fine->buf,
+ RADEON_USAGE_WRITE, RADEON_PRIO_QUERY);
+ si_cp_release_mem(ctx, ctx->gfx_cs,
+ V_028A90_BOTTOM_OF_PIPE_TS, 0,
+ EOP_DST_SEL_MEM, EOP_INT_SEL_NONE,
+ 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_context *ctx,
- struct pipe_fence_handle *fence,
- uint64_t timeout)
+static bool si_fence_finish(struct pipe_screen *screen,
+ struct pipe_context *ctx,
+ struct pipe_fence_handle *fence,
+ uint64_t timeout)
{
struct radeon_winsys *rws = ((struct si_screen*)screen)->ws;
- struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+ struct si_multi_fence *sfence = (struct si_multi_fence *)fence;
+ struct si_context *sctx;
int64_t abs_timeout = os_time_get_absolute_timeout(timeout);
- if (!util_queue_fence_is_signalled(&rfence->ready)) {
- if (rfence->tc_token) {
+ ctx = threaded_context_unwrap_sync(ctx);
+ sctx = (struct si_context*)(ctx ? ctx : NULL);
+
+ if (!util_queue_fence_is_signalled(&sfence->ready)) {
+ if (sfence->tc_token) {
/* Ensure that si_flush_from_st will be called for
* this fence, but only if we're in the API thread
* where the context is current.
* 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, sfence->tc_token,
timeout == 0);
}
return false;
if (timeout == PIPE_TIMEOUT_INFINITE) {
- util_queue_fence_wait(&rfence->ready);
+ util_queue_fence_wait(&sfence->ready);
} else {
- if (!util_queue_fence_wait_timeout(&rfence->ready, abs_timeout))
+ if (!util_queue_fence_wait_timeout(&sfence->ready, abs_timeout))
return false;
}
}
}
- if (rfence->sdma) {
- if (!rws->fence_wait(rws, rfence->sdma, timeout))
+ if (sfence->sdma) {
+ if (!rws->fence_wait(rws, sfence->sdma, timeout))
return false;
/* Recompute the timeout after waiting. */
}
}
- if (!rfence->gfx)
+ if (!sfence->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);
+ if (sfence->fine.buf &&
+ si_fine_fence_signaled(rws, &sfence->fine)) {
+ rws->fence_reference(&sfence->gfx, NULL);
+ si_resource_reference(&sfence->fine.buf, NULL);
return true;
}
/* Flush the gfx IB if it hasn't been flushed yet. */
- 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->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) |
- RADEON_FLUSH_START_NEXT_GFX_IB_NOW,
- NULL);
- rfence->gfx_unflushed.ctx = NULL;
-
- if (!timeout)
- return false;
+ if (sctx && sfence->gfx_unflushed.ctx == sctx &&
+ sfence->gfx_unflushed.ib_index == sctx->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.
+ */
+ si_flush_gfx_cs(sctx,
+ (timeout ? 0 : PIPE_FLUSH_ASYNC) |
+ RADEON_FLUSH_START_NEXT_GFX_IB_NOW,
+ NULL);
+ sfence->gfx_unflushed.ctx = NULL;
- /* 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;
- }
+ 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;
}
}
- if (rws->fence_wait(rws, rfence->gfx, timeout))
+ if (rws->fence_wait(rws, sfence->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))
+ if (sfence->fine.buf &&
+ si_fine_fence_signaled(rws, &sfence->fine))
return true;
return false;
{
struct si_screen *sscreen = (struct si_screen*)ctx->screen;
struct radeon_winsys *ws = sscreen->ws;
- struct si_multi_fence *rfence;
+ struct si_multi_fence *sfence;
*pfence = NULL;
- rfence = si_create_multi_fence();
- if (!rfence)
+ sfence = si_create_multi_fence();
+ if (!sfence)
return;
switch (type) {
if (!sscreen->info.has_fence_to_handle)
goto finish;
- rfence->gfx = ws->fence_import_sync_file(ws, fd);
+ sfence->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);
+ sfence->gfx = ws->fence_import_syncobj(ws, fd);
break;
default:
}
finish:
- if (!rfence->gfx) {
- FREE(rfence);
+ if (!sfence->gfx) {
+ FREE(sfence);
return;
}
- *pfence = (struct pipe_fence_handle*)rfence;
+ *pfence = (struct pipe_fence_handle*)sfence;
}
static int si_fence_get_fd(struct pipe_screen *screen,
{
struct si_screen *sscreen = (struct si_screen*)screen;
struct radeon_winsys *ws = sscreen->ws;
- struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+ struct si_multi_fence *sfence = (struct si_multi_fence *)fence;
int gfx_fd = -1, sdma_fd = -1;
if (!sscreen->info.has_fence_to_handle)
return -1;
- util_queue_fence_wait(&rfence->ready);
+ util_queue_fence_wait(&sfence->ready);
/* Deferred fences aren't supported. */
- assert(!rfence->gfx_unflushed.ctx);
- if (rfence->gfx_unflushed.ctx)
+ assert(!sfence->gfx_unflushed.ctx);
+ if (sfence->gfx_unflushed.ctx)
return -1;
- if (rfence->sdma) {
- sdma_fd = ws->fence_export_sync_file(ws, rfence->sdma);
+ if (sfence->sdma) {
+ sdma_fd = ws->fence_export_sync_file(ws, sfence->sdma);
if (sdma_fd == -1)
return -1;
}
- if (rfence->gfx) {
- gfx_fd = ws->fence_export_sync_file(ws, rfence->gfx);
+ if (sfence->gfx) {
+ gfx_fd = ws->fence_export_sync_file(ws, sfence->gfx);
if (gfx_fd == -1) {
if (sdma_fd != -1)
close(sdma_fd);
}
/* DMA IBs are preambles to gfx IBs, therefore must be flushed first. */
- if (sctx->dma_cs)
+ if (sctx->sdma_cs)
si_flush_dma_cs(sctx, rflags, fence ? &sdma_fence : NULL);
if (!radeon_emitted(sctx->gfx_cs, sctx->initial_gfx_cs_size)) {
}
assert(!fine.buf);
finish:
- if (!(flags & PIPE_FLUSH_DEFERRED)) {
- if (sctx->dma_cs)
- ws->cs_sync_flush(sctx->dma_cs);
+ if (!(flags & (PIPE_FLUSH_DEFERRED | PIPE_FLUSH_ASYNC))) {
+ if (sctx->sdma_cs)
+ ws->cs_sync_flush(sctx->sdma_cs);
ws->cs_sync_flush(sctx->gfx_cs);
}
}
struct pipe_fence_handle *fence)
{
struct si_context *sctx = (struct si_context *)ctx;
- struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+ struct si_multi_fence *sfence = (struct si_multi_fence *)fence;
/* We should have at least one syncobj to signal */
- assert(rfence->sdma || rfence->gfx);
+ assert(sfence->sdma || sfence->gfx);
- if (rfence->sdma)
- si_add_syncobj_signal(sctx, rfence->sdma);
- if (rfence->gfx)
- si_add_syncobj_signal(sctx, rfence->gfx);
+ if (sfence->sdma)
+ si_add_syncobj_signal(sctx, sfence->sdma);
+ if (sfence->gfx)
+ si_add_syncobj_signal(sctx, sfence->gfx);
/**
* The spec does not require a flush here. We insert a flush
struct pipe_fence_handle *fence)
{
struct si_context *sctx = (struct si_context *)ctx;
- struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
+ struct si_multi_fence *sfence = (struct si_multi_fence *)fence;
- util_queue_fence_wait(&rfence->ready);
+ util_queue_fence_wait(&sfence->ready);
/* Unflushed fences from the same context are no-ops. */
- if (rfence->gfx_unflushed.ctx &&
- rfence->gfx_unflushed.ctx == sctx)
+ if (sfence->gfx_unflushed.ctx &&
+ sfence->gfx_unflushed.ctx == sctx)
return;
/* All unflushed commands will not start execution before
*/
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);
+ if (sfence->sdma)
+ si_add_fence_dependency(sctx, sfence->sdma);
+ if (sfence->gfx)
+ si_add_fence_dependency(sctx, sfence->gfx);
}
void si_init_fence_functions(struct si_context *ctx)
projects / mesa.git / blobdiff