radeonsi: move si_shader_binary_upload out of si_compile_llvm
[mesa.git] / src / gallium / drivers / radeonsi / si_hw_context.c
index 1d5d42657e4e80c2fcdb8da96a13f753bc5db9c2..baa02293c41ce928e7c7e769b6b16fe99d8af640 100644 (file)
 /* initialize */
 void si_need_cs_space(struct si_context *ctx)
 {
-       struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
+       struct radeon_winsys_cs *cs = ctx->b.gfx.cs;
+       struct radeon_winsys_cs *dma = ctx->b.dma.cs;
+
+       /* Flush the DMA IB if it's not empty. */
+       if (dma && dma->cdw)
+               ctx->b.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 (unlikely(!ctx->b.ws->cs_memory_below_limit(ctx->b.rings.gfx.cs,
+       if (unlikely(!ctx->b.ws->cs_memory_below_limit(ctx->b.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.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
                return;
        }
        ctx->b.gtt = 0;
@@ -49,32 +54,36 @@ void si_need_cs_space(struct si_context *ctx)
         * 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);
+               ctx->b.gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
 }
 
 void si_context_gfx_flush(void *context, unsigned flags,
                          struct pipe_fence_handle **fence)
 {
        struct si_context *ctx = context;
-       struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
+       struct radeon_winsys_cs *cs = ctx->b.gfx.cs;
        struct radeon_winsys *ws = ctx->b.ws;
 
+       if (ctx->gfx_flush_in_progress)
+               return;
+
+       ctx->gfx_flush_in_progress = true;
+
        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);
+               ctx->gfx_flush_in_progress = false;
                return;
        }
 
-       ctx->b.rings.gfx.flushing = true;
-
        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, NULL);
@@ -85,25 +94,42 @@ void si_context_gfx_flush(void *context, unsigned 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. */
        ws->cs_flush(cs, flags, &ctx->last_gfx_fence,
                     ctx->screen->b.cs_count++);
-       ctx->b.rings.gfx.flushing = false;
 
        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);
+       ctx->gfx_flush_in_progress = false;
 }
 
 void si_begin_new_cs(struct si_context *ctx)
@@ -127,9 +153,9 @@ void si_begin_new_cs(struct si_context *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
@@ -139,6 +165,8 @@ void si_begin_new_cs(struct si_context *ctx)
 
        /* The CS initialization should be emitted before everything else. */
        si_pm4_emit(ctx, ctx->init_config);
+       if (ctx->init_config_gs_rings)
+               si_pm4_emit(ctx, ctx->init_config_gs_rings);
 
        ctx->framebuffer.dirty_cbufs = (1 << 8) - 1;
        ctx->framebuffer.dirty_zsbuf = true;
@@ -154,7 +182,9 @@ void si_begin_new_cs(struct si_context *ctx)
        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_mark_atom_dirty(ctx, &ctx->b.render_cond_atom);
        si_all_descriptors_begin_new_cs(ctx);
 
        ctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
@@ -164,7 +194,7 @@ void si_begin_new_cs(struct si_context *ctx)
 
        r600_postflush_resume_features(&ctx->b);
 
-       ctx->b.initial_gfx_cs_size = ctx->b.rings.gfx.cs->cdw;
+       ctx->b.initial_gfx_cs_size = ctx->b.gfx.cs->cdw;
 
        /* Invalidate various draw states so that they are emitted before
         * the first draw call. */