#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)
{
struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
- int i;
+ struct radeon_winsys_cs *dma = ctx->b.rings.dma.cs;
+
+ /* Flush the DMA IB if it's not empty. */
+ if (dma && dma->cdw)
+ ctx->b.rings.dma.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
/* There are two memory usage counters in the winsys for all buffers
- * that have been added (cs_add_reloc) and two counters in the pipe
+ * that have been added (cs_add_buffer) and two counters in the pipe
* driver for those that haven't been added yet.
- * */
- if (!ctx->b.ws->cs_memory_below_limit(ctx->b.rings.gfx.cs, ctx->b.vram, ctx->b.gtt)) {
+ */
+ 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);
ctx->b.vram = 0;
/* If the CS is sufficiently large, don't count the space needed
- * and just flush if there is less than 8096 dwords left.
+ * and just flush if there is not enough space left.
*/
- if (cs->max_dw >= 24 * 1024) {
- if (cs->cdw > cs->max_dw - 8 * 1024)
- ctx->b.rings.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
- return;
- }
-
- /* The number of dwords we already used in the CS so far. */
- num_dw += cs->cdw;
-
- if (count_draw_in) {
- for (i = 0; i < SI_NUM_ATOMS; 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 +
- ctx->b.num_cs_dw_timer_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;
-
- if (ctx->screen->b.trace_bo)
- num_dw += SI_TRACE_CS_DWORDS * 2;
-
- /* Flush if there's not enough space. */
- if (num_dw > cs->max_dw) {
+ 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,
r600_preflush_suspend_features(&ctx->b);
ctx->b.flags |= SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER |
- SI_CONTEXT_INV_TC_L1 |
- SI_CONTEXT_INV_TC_L2 |
+ SI_CONTEXT_INV_VMEM_L1 |
+ SI_CONTEXT_INV_GLOBAL_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) {
+ unsigned i;
+
+ /* Save the IB for debug contexts. */
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);
+
+ /* Save the buffer list. */
+ if (ctx->last_bo_list) {
+ for (i = 0; i < ctx->last_bo_count; i++)
+ pb_reference(&ctx->last_bo_list[i].buf, NULL);
+ free(ctx->last_bo_list);
+ }
+ ctx->last_bo_count = ws->cs_get_buffer_list(cs, NULL);
+ ctx->last_bo_list = calloc(ctx->last_bo_count,
+ sizeof(ctx->last_bo_list[0]));
+ ws->cs_get_buffer_list(cs, ctx->last_bo_list);
}
/* Flush the CS. */
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);
}
/* Flush read caches at the beginning of CS. */
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_VMEM_L1 |
+ SI_CONTEXT_INV_GLOBAL_L2 |
+ SI_CONTEXT_INV_SMEM_L1 |
SI_CONTEXT_INV_ICACHE;
/* set all valid group as dirty so they get reemited on
/* The CS initialization should be emitted before everything else. */
si_pm4_emit(ctx, ctx->init_config);
- si_mark_atom_dirty(ctx, &ctx->clip_regs);
+ 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->spi_ps_input);
si_mark_atom_dirty(ctx, &ctx->b.streamout.enable_atom);
si_all_descriptors_begin_new_cs(ctx);