#include "si_pipe.h"
/* initialize */
-void si_need_cs_space(struct si_context *ctx, unsigned num_dw,
- boolean count_draw_in)
+void si_need_cs_space(struct si_context *ctx)
{
- int i;
-
- /* The number of dwords we already used in the CS so far. */
- num_dw += ctx->b.rings.gfx.cs->cdw;
-
- if (count_draw_in) {
- for (i = 0; i < SI_NUM_ATOMS(ctx); i++) {
- if (ctx->atoms.array[i]->dirty) {
- num_dw += ctx->atoms.array[i]->num_dw;
- }
- }
-
- /* The number of dwords all the dirty states would take. */
- num_dw += si_pm4_dirty_dw(ctx);
-
- /* The upper-bound of how much a draw command would take. */
- num_dw += SI_MAX_DRAW_CS_DWORDS;
- }
-
- /* Count in queries_suspend. */
- num_dw += ctx->b.num_cs_dw_nontimer_queries_suspend;
-
- /* Count in streamout_end at the end of CS. */
- if (ctx->b.streamout.begin_emitted) {
- num_dw += ctx->b.streamout.num_dw_for_end;
- }
-
- /* Count in render_condition(NULL) at the end of CS. */
- if (ctx->b.predicate_drawing) {
- num_dw += 3;
- }
-
- /* Count in framebuffer cache flushes at the end of CS. */
- num_dw += ctx->atoms.s.cache_flush->num_dw;
+ struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
-#if SI_TRACE_CS
- if (ctx->screen->b.trace_bo) {
- num_dw += SI_TRACE_CS_DWORDS;
+ /* There are two memory usage counters in the winsys for all buffers
+ * that have been added (cs_add_buffer) and two counters in the pipe
+ * driver for those that haven't been added yet.
+ */
+ if (unlikely(!ctx->b.ws->cs_memory_below_limit(ctx->b.rings.gfx.cs,
+ ctx->b.vram, ctx->b.gtt))) {
+ ctx->b.gtt = 0;
+ ctx->b.vram = 0;
+ ctx->b.rings.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
+ return;
}
-#endif
+ ctx->b.gtt = 0;
+ ctx->b.vram = 0;
- /* Flush if there's not enough space. */
- if (num_dw > RADEON_MAX_CMDBUF_DWORDS) {
+ /* If the CS is sufficiently large, don't count the space needed
+ * and just flush if there is not enough space left.
+ */
+ if (unlikely(cs->cdw > cs->max_dw - 2048))
ctx->b.rings.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
- }
}
void si_context_gfx_flush(void *context, unsigned flags,
{
struct si_context *ctx = context;
struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
-
- if (cs->cdw == ctx->b.initial_gfx_cs_size && !fence)
+ struct radeon_winsys *ws = ctx->b.ws;
+
+ if (cs->cdw == ctx->b.initial_gfx_cs_size &&
+ (!fence || ctx->last_gfx_fence)) {
+ if (fence)
+ ws->fence_reference(fence, ctx->last_gfx_fence);
+ if (!(flags & RADEON_FLUSH_ASYNC))
+ ws->cs_sync_flush(cs);
return;
+ }
ctx->b.rings.gfx.flushing = true;
SI_CONTEXT_INV_TC_L2 |
/* this is probably not needed anymore */
SI_CONTEXT_PS_PARTIAL_FLUSH;
- si_emit_cache_flush(&ctx->b, NULL);
+ si_emit_cache_flush(ctx, NULL);
/* force to keep tiling flags */
flags |= RADEON_FLUSH_KEEP_TILING_FLAGS;
+ if (ctx->trace_buf)
+ si_trace_emit(ctx);
+
+ /* Save the IB for debug contexts. */
+ if (ctx->is_debug) {
+ free(ctx->last_ib);
+ ctx->last_ib_dw_size = cs->cdw;
+ ctx->last_ib = malloc(cs->cdw * 4);
+ memcpy(ctx->last_ib, cs->buf, cs->cdw * 4);
+ r600_resource_reference(&ctx->last_trace_buf, ctx->trace_buf);
+ r600_resource_reference(&ctx->trace_buf, NULL);
+ }
+
/* Flush the CS. */
- ctx->b.ws->cs_flush(cs, flags, fence, ctx->screen->b.cs_count++);
+ ws->cs_flush(cs, flags, &ctx->last_gfx_fence,
+ ctx->screen->b.cs_count++);
ctx->b.rings.gfx.flushing = false;
-#if SI_TRACE_CS
- if (ctx->screen->b.trace_bo) {
- struct si_screen *sscreen = ctx->screen;
- unsigned i;
-
- for (i = 0; i < 10; i++) {
- usleep(5);
- if (!ctx->b.ws->buffer_is_busy(sscreen->b.trace_bo->buf, RADEON_USAGE_READWRITE)) {
- break;
- }
- }
- if (i == 10) {
- fprintf(stderr, "timeout on cs lockup likely happen at cs %d dw %d\n",
- sscreen->b.trace_ptr[1], sscreen->b.trace_ptr[0]);
- } else {
- fprintf(stderr, "cs %d executed in %dms\n", sscreen->b.trace_ptr[1], i * 5);
- }
- }
-#endif
+ if (fence)
+ ws->fence_reference(fence, ctx->last_gfx_fence);
+
+ /* Check VM faults if needed. */
+ if (ctx->screen->b.debug_flags & DBG_CHECK_VM)
+ si_check_vm_faults(ctx);
si_begin_new_cs(ctx);
}
void si_begin_new_cs(struct si_context *ctx)
{
+ if (ctx->is_debug) {
+ uint32_t zero = 0;
+
+ /* Create a buffer used for writing trace IDs and initialize it to 0. */
+ assert(!ctx->trace_buf);
+ ctx->trace_buf = (struct r600_resource*)
+ pipe_buffer_create(ctx->b.b.screen, PIPE_BIND_CUSTOM,
+ 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;
+ }
+
+ if (ctx->trace_buf)
+ si_trace_emit(ctx);
+
/* Flush read caches at the beginning of CS. */
- ctx->b.flags |= SI_CONTEXT_INV_TC_L1 |
+ ctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER |
+ SI_CONTEXT_INV_TC_L1 |
SI_CONTEXT_INV_TC_L2 |
SI_CONTEXT_INV_KCACHE |
SI_CONTEXT_INV_ICACHE;
/* The CS initialization should be emitted before everything else. */
si_pm4_emit(ctx, ctx->init_config);
- ctx->clip_regs.dirty = true;
- ctx->framebuffer.atom.dirty = true;
- ctx->msaa_sample_locs.dirty = true;
- ctx->msaa_config.dirty = true;
- ctx->db_render_state.dirty = true;
- ctx->b.streamout.enable_atom.dirty = true;
+ ctx->framebuffer.dirty_cbufs = (1 << 8) - 1;
+ ctx->framebuffer.dirty_zsbuf = true;
+ 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);
+ si_mark_atom_dirty(ctx, &ctx->msaa_sample_locs);
+ si_mark_atom_dirty(ctx, &ctx->msaa_config);
+ si_mark_atom_dirty(ctx, &ctx->sample_mask.atom);
+ si_mark_atom_dirty(ctx, &ctx->cb_target_mask);
+ si_mark_atom_dirty(ctx, &ctx->blend_color.atom);
+ si_mark_atom_dirty(ctx, &ctx->db_render_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_all_descriptors_begin_new_cs(ctx);
+ ctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
+ ctx->viewports.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
+ si_mark_atom_dirty(ctx, &ctx->scissors.atom);
+ si_mark_atom_dirty(ctx, &ctx->viewports.atom);
+
r600_postflush_resume_features(&ctx->b);
ctx->b.initial_gfx_cs_size = ctx->b.rings.gfx.cs->cdw;
+
+ /* Invalidate various draw states so that they are emitted before
+ * the first draw call. */
si_invalidate_draw_sh_constants(ctx);
ctx->last_primitive_restart_en = -1;
ctx->last_restart_index = SI_RESTART_INDEX_UNKNOWN;
ctx->last_gs_out_prim = -1;
ctx->last_prim = -1;
ctx->last_multi_vgt_param = -1;
+ ctx->last_ls_hs_config = -1;
ctx->last_rast_prim = -1;
ctx->last_sc_line_stipple = ~0;
ctx->emit_scratch_reloc = true;
+ ctx->last_ls = NULL;
+ ctx->last_tcs = NULL;
+ ctx->last_tes_sh_base = -1;
+ ctx->last_num_tcs_input_cp = -1;
}