* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
- *
- * Authors:
- * Jerome Glisse
*/
#include "si_pipe.h"
#include "radeon/r600_cs.h"
-static unsigned si_descriptor_list_cs_space(unsigned count, unsigned element_size)
-{
- /* Ensure we have enough space to start a new range in a hole */
- assert(element_size >= 3);
-
- /* 5 dwords for possible load to reinitialize when we have no preamble
- * IB + 5 dwords for write to L2 + 3 bytes for every range written to
- * CE RAM.
- */
- return 5 + 5 + 3 + count * element_size;
-}
+#include "util/os_time.h"
-static unsigned si_ce_needed_cs_space(void)
+void si_destroy_saved_cs(struct si_saved_cs *scs)
{
- unsigned space = 0;
-
- space += si_descriptor_list_cs_space(SI_NUM_CONST_BUFFERS, 4);
- space += si_descriptor_list_cs_space(SI_NUM_SHADER_BUFFERS, 4);
- space += si_descriptor_list_cs_space(SI_NUM_SAMPLERS, 16);
- space += si_descriptor_list_cs_space(SI_NUM_IMAGES, 8);
- space *= SI_NUM_SHADERS;
-
- space += si_descriptor_list_cs_space(SI_NUM_RW_BUFFERS, 4);
-
- /* Increment CE counter packet */
- space += 2;
-
- return space;
+ si_clear_saved_cs(&scs->gfx);
+ r600_resource_reference(&scs->trace_buf, NULL);
+ free(scs);
}
/* initialize */
void si_need_cs_space(struct si_context *ctx)
{
struct radeon_winsys_cs *cs = ctx->b.gfx.cs;
- struct radeon_winsys_cs *ce_ib = ctx->ce_ib;
/* There is no need to flush the DMA IB here, because
* r600_need_dma_space always flushes the GFX IB if there is
ctx->b.vram, ctx->b.gtt))) {
ctx->b.gtt = 0;
ctx->b.vram = 0;
- ctx->b.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
+ ctx->b.gfx.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
return;
}
ctx->b.gtt = 0;
/* If the CS is sufficiently large, don't count the space needed
* and just flush if there is not enough space left.
*/
- if (!ctx->b.ws->cs_check_space(cs, 2048) ||
- (ce_ib && !ctx->b.ws->cs_check_space(ce_ib, si_ce_needed_cs_space())))
- ctx->b.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
+ if (!ctx->b.ws->cs_check_space(cs, 2048))
+ ctx->b.gfx.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
}
void si_context_gfx_flush(void *context, unsigned flags,
if (!radeon_emitted(cs, ctx->b.initial_gfx_cs_size))
return;
- if (r600_check_device_reset(&ctx->b))
+ if (si_check_device_reset(&ctx->b))
return;
- if (ctx->screen->b.debug_flags & DBG_CHECK_VM)
- flags &= ~RADEON_FLUSH_ASYNC;
+ if (ctx->screen->debug_flags & DBG(CHECK_VM))
+ flags &= ~PIPE_FLUSH_ASYNC;
/* If the state tracker is flushing the GFX IB, r600_flush_from_st is
* responsible for flushing the DMA IB and merging the fences from both.
ctx->gfx_flush_in_progress = true;
- r600_preflush_suspend_features(&ctx->b);
+ if (!LIST_IS_EMPTY(&ctx->b.active_queries))
+ si_suspend_queries(&ctx->b);
+
+ ctx->streamout.suspended = false;
+ if (ctx->streamout.begin_emitted) {
+ si_emit_streamout_end(ctx);
+ ctx->streamout.suspended = true;
+ }
ctx->b.flags |= SI_CONTEXT_CS_PARTIAL_FLUSH |
SI_CONTEXT_PS_PARTIAL_FLUSH;
si_emit_cache_flush(ctx);
- if (ctx->trace_buf)
+ if (ctx->current_saved_cs) {
si_trace_emit(ctx);
+ si_log_hw_flush(ctx);
- if (ctx->is_debug) {
/* Save the IB for debug contexts. */
- radeon_clear_saved_cs(&ctx->last_gfx);
- radeon_save_cs(ws, cs, &ctx->last_gfx);
- r600_resource_reference(&ctx->last_trace_buf, ctx->trace_buf);
- r600_resource_reference(&ctx->trace_buf, NULL);
+ si_save_cs(ws, cs, &ctx->current_saved_cs->gfx, true);
+ ctx->current_saved_cs->flushed = true;
+ ctx->current_saved_cs->time_flush = os_time_get_nano();
}
/* Flush the CS. */
ws->cs_flush(cs, flags, &ctx->b.last_gfx_fence);
if (fence)
ws->fence_reference(fence, ctx->b.last_gfx_fence);
+
+ /* This must be after cs_flush returns, since the context's API
+ * thread can concurrently read this value in si_fence_finish. */
ctx->b.num_gfx_cs_flushes++;
/* Check VM faults if needed. */
- if (ctx->screen->b.debug_flags & DBG_CHECK_VM) {
+ if (ctx->screen->debug_flags & DBG(CHECK_VM)) {
/* Use conservative timeout 800ms, after which we won't wait any
* longer and assume the GPU is hung.
*/
ctx->b.ws->fence_wait(ctx->b.ws, ctx->b.last_gfx_fence, 800*1000*1000);
- si_check_vm_faults(&ctx->b, &ctx->last_gfx, RING_GFX);
+ si_check_vm_faults(&ctx->b, &ctx->current_saved_cs->gfx, RING_GFX);
}
+ if (ctx->current_saved_cs)
+ si_saved_cs_reference(&ctx->current_saved_cs, NULL);
+
si_begin_new_cs(ctx);
ctx->gfx_flush_in_progress = false;
}
-void si_begin_new_cs(struct si_context *ctx)
+static void si_begin_cs_debug(struct si_context *ctx)
{
- if (ctx->is_debug) {
- uint32_t zero = 0;
+ static const uint32_t zeros[1];
+ assert(!ctx->current_saved_cs);
+
+ ctx->current_saved_cs = calloc(1, sizeof(*ctx->current_saved_cs));
+ if (!ctx->current_saved_cs)
+ return;
- /* Create a buffer used for writing trace IDs and initialize it to 0. */
- assert(!ctx->trace_buf);
- ctx->trace_buf = (struct r600_resource*)
+ pipe_reference_init(&ctx->current_saved_cs->reference, 1);
+
+ ctx->current_saved_cs->trace_buf = (struct r600_resource*)
pipe_buffer_create(ctx->b.b.screen, 0,
- PIPE_USAGE_STAGING, 4);
- if (ctx->trace_buf)
- pipe_buffer_write_nooverlap(&ctx->b.b, &ctx->trace_buf->b.b,
- 0, sizeof(zero), &zero);
- ctx->trace_id = 0;
+ PIPE_USAGE_STAGING, 8);
+ if (!ctx->current_saved_cs->trace_buf) {
+ free(ctx->current_saved_cs);
+ ctx->current_saved_cs = NULL;
+ return;
}
- if (ctx->trace_buf)
- si_trace_emit(ctx);
+ pipe_buffer_write_nooverlap(&ctx->b.b, &ctx->current_saved_cs->trace_buf->b.b,
+ 0, sizeof(zeros), zeros);
+ ctx->current_saved_cs->trace_id = 0;
+
+ si_trace_emit(ctx);
+
+ radeon_add_to_buffer_list(&ctx->b, &ctx->b.gfx, ctx->current_saved_cs->trace_buf,
+ RADEON_USAGE_READWRITE, RADEON_PRIO_TRACE);
+}
+
+void si_begin_new_cs(struct si_context *ctx)
+{
+ if (ctx->is_debug)
+ si_begin_cs_debug(ctx);
/* Flush read caches at the beginning of CS not flushed by the kernel. */
if (ctx->b.chip_class >= CIK)
ctx->b.flags |= SI_CONTEXT_INV_SMEM_L1 |
SI_CONTEXT_INV_ICACHE;
- ctx->b.flags |= R600_CONTEXT_START_PIPELINE_STATS;
+ ctx->b.flags |= SI_CONTEXT_START_PIPELINE_STATS;
/* set all valid group as dirty so they get reemited on
* next draw command
if (ctx->init_config_gs_rings)
si_pm4_emit(ctx, ctx->init_config_gs_rings);
- if (ctx->ce_preamble_ib)
- si_ce_enable_loads(ctx->ce_preamble_ib);
- else if (ctx->ce_ib)
- si_ce_enable_loads(ctx->ce_ib);
-
- if (ctx->ce_preamble_ib)
- si_ce_reinitialize_all_descriptors(ctx);
-
- if (ctx->b.chip_class >= CIK)
- si_mark_atom_dirty(ctx, &ctx->prefetch_L2);
-
- ctx->framebuffer.dirty_cbufs = (1 << 8) - 1;
- ctx->framebuffer.dirty_zsbuf = true;
+ if (ctx->queued.named.ls)
+ ctx->prefetch_L2_mask |= SI_PREFETCH_LS;
+ if (ctx->queued.named.hs)
+ ctx->prefetch_L2_mask |= SI_PREFETCH_HS;
+ if (ctx->queued.named.es)
+ ctx->prefetch_L2_mask |= SI_PREFETCH_ES;
+ if (ctx->queued.named.gs)
+ ctx->prefetch_L2_mask |= SI_PREFETCH_GS;
+ if (ctx->queued.named.vs)
+ ctx->prefetch_L2_mask |= SI_PREFETCH_VS;
+ if (ctx->queued.named.ps)
+ ctx->prefetch_L2_mask |= SI_PREFETCH_PS;
+ if (ctx->vb_descriptors_buffer && ctx->vertex_elements)
+ ctx->prefetch_L2_mask |= SI_PREFETCH_VBO_DESCRIPTORS;
+
+ /* CLEAR_STATE disables all colorbuffers, so only enable bound ones. */
+ bool has_clear_state = ctx->screen->has_clear_state;
+ if (has_clear_state) {
+ ctx->framebuffer.dirty_cbufs =
+ u_bit_consecutive(0, ctx->framebuffer.state.nr_cbufs);
+ /* CLEAR_STATE disables the zbuffer, so only enable it if it's bound. */
+ ctx->framebuffer.dirty_zsbuf = ctx->framebuffer.state.zsbuf != NULL;
+ } else {
+ ctx->framebuffer.dirty_cbufs = u_bit_consecutive(0, 8);
+ ctx->framebuffer.dirty_zsbuf = true;
+ }
+ /* This should always be marked as dirty to set the framebuffer scissor
+ * at least. */
si_mark_atom_dirty(ctx, &ctx->framebuffer.atom);
si_mark_atom_dirty(ctx, &ctx->clip_regs);
- si_mark_atom_dirty(ctx, &ctx->clip_state.atom);
+ /* CLEAR_STATE sets zeros. */
+ if (!has_clear_state || ctx->clip_state.any_nonzeros)
+ si_mark_atom_dirty(ctx, &ctx->clip_state.atom);
ctx->msaa_sample_locs.nr_samples = 0;
si_mark_atom_dirty(ctx, &ctx->msaa_sample_locs.atom);
si_mark_atom_dirty(ctx, &ctx->msaa_config);
- si_mark_atom_dirty(ctx, &ctx->sample_mask.atom);
+ /* CLEAR_STATE sets 0xffff. */
+ if (!has_clear_state || ctx->sample_mask.sample_mask != 0xffff)
+ si_mark_atom_dirty(ctx, &ctx->sample_mask.atom);
si_mark_atom_dirty(ctx, &ctx->cb_render_state);
- si_mark_atom_dirty(ctx, &ctx->blend_color.atom);
+ /* CLEAR_STATE sets zeros. */
+ if (!has_clear_state || ctx->blend_color.any_nonzeros)
+ si_mark_atom_dirty(ctx, &ctx->blend_color.atom);
si_mark_atom_dirty(ctx, &ctx->db_render_state);
+ if (ctx->b.chip_class >= GFX9)
+ si_mark_atom_dirty(ctx, &ctx->dpbb_state);
si_mark_atom_dirty(ctx, &ctx->stencil_ref.atom);
si_mark_atom_dirty(ctx, &ctx->spi_map);
- si_mark_atom_dirty(ctx, &ctx->b.streamout.enable_atom);
+ si_mark_atom_dirty(ctx, &ctx->streamout.enable_atom);
si_mark_atom_dirty(ctx, &ctx->b.render_cond_atom);
si_all_descriptors_begin_new_cs(ctx);
+ si_all_resident_buffers_begin_new_cs(ctx);
- ctx->b.scissors.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
- ctx->b.viewports.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
- ctx->b.viewports.depth_range_dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
- si_mark_atom_dirty(ctx, &ctx->b.scissors.atom);
- si_mark_atom_dirty(ctx, &ctx->b.viewports.atom);
+ ctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
+ ctx->viewports.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
+ ctx->viewports.depth_range_dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
+ si_mark_atom_dirty(ctx, &ctx->scissors.atom);
+ si_mark_atom_dirty(ctx, &ctx->viewports.atom);
si_mark_atom_dirty(ctx, &ctx->scratch_state);
if (ctx->scratch_buffer) {
- r600_context_add_resource_size(&ctx->b.b,
- &ctx->scratch_buffer->b.b);
+ si_context_add_resource_size(&ctx->b.b,
+ &ctx->scratch_buffer->b.b);
+ }
+
+ if (ctx->streamout.suspended) {
+ ctx->streamout.append_bitmask = ctx->streamout.enabled_mask;
+ si_streamout_buffers_dirty(ctx);
}
- r600_postflush_resume_features(&ctx->b);
+ if (!LIST_IS_EMPTY(&ctx->b.active_queries))
+ si_resume_queries(&ctx->b);
assert(!ctx->b.gfx.cs->prev_dw);
ctx->b.initial_gfx_cs_size = ctx->b.gfx.cs->current.cdw;
projects / mesa.git / blobdiff