radeonsi: set COMPUTE_DISPATCH_INITIATOR.ORDER_MODE = 1

[mesa.git] / src / gallium / drivers / freedreno / freedreno_context.c

diff --git a/src/gallium/drivers/freedreno/freedreno_context.c b/src/gallium/drivers/freedreno/freedreno_context.c
index f8604f132ad056e0d3685d12a2eb4267144a43d1..1cf366b0c6a62e219b440af2594e6fd756e52e67 100644 (file)
--- a/src/gallium/drivers/freedreno/freedreno_context.c
+++ b/src/gallium/drivers/freedreno/freedreno_context.c
@@ -37,27 +37,33 @@
 #include "freedreno_query.h"
 #include "freedreno_query_hw.h"
 #include "freedreno_util.h"
+#include "util/u_upload_mgr.h"
 
 static void
 fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fence,
                unsigned flags)
 {
        struct fd_context *ctx = fd_context(pctx);
-       uint32_t timestamp;
+
+       if (flags & PIPE_FLUSH_FENCE_FD)
+               ctx->batch->needs_out_fence_fd = true;
 
        if (!ctx->screen->reorder) {
-               struct fd_batch *batch = NULL;
-               fd_batch_reference(&batch, ctx->batch);
-               fd_batch_flush(batch, true);
-               timestamp = fd_ringbuffer_timestamp(batch->gmem);
-               fd_batch_reference(&batch, NULL);
+               fd_batch_flush(ctx->batch, true);
        } else {
-               timestamp = fd_bc_flush(&ctx->screen->batch_cache, ctx);
+               fd_bc_flush(&ctx->screen->batch_cache, ctx);
        }
 
        if (fence) {
-               fd_screen_fence_ref(pctx->screen, fence, NULL);
-               *fence = fd_fence_create(pctx, timestamp);
+               /* if there hasn't been any rendering submitted yet, we might not
+                * have actually created a fence
+                */
+               if (!ctx->last_fence || ctx->batch->needs_out_fence_fd) {
+                       ctx->batch->needs_flush = true;
+                       fd_gmem_render_noop(ctx->batch);
+                       fd_batch_reset(ctx->batch);
+               }
+               fd_fence_ref(pctx->screen, fence, ctx->last_fence);
        }
 }
 
@@ -80,7 +86,10 @@ fd_emit_string_marker(struct pipe_context *pctx, const char *string, int len)
        /* max packet size is 0x3fff dwords: */
        len = MIN2(len, 0x3fff * 4);
 
-       OUT_PKT3(ring, CP_NOP, align(len, 4) / 4);
+       if (ctx->screen->gpu_id >= 500)
+               OUT_PKT7(ring, CP_NOP, align(len, 4) / 4);
+       else
+               OUT_PKT3(ring, CP_NOP, align(len, 4) / 4);
        while (len >= 4) {
                OUT_RING(ring, *buf);
                buf++;
@@ -103,25 +112,29 @@ fd_context_destroy(struct pipe_context *pctx)
 
        DBG("");
 
-       if (ctx->screen->reorder)
+       if (ctx->screen->reorder && util_queue_is_initialized(&ctx->flush_queue))
                util_queue_destroy(&ctx->flush_queue);
 
        fd_batch_reference(&ctx->batch, NULL);  /* unref current batch */
        fd_bc_invalidate_context(ctx);
 
+       fd_fence_ref(pctx->screen, &ctx->last_fence, NULL);
+
        fd_prog_fini(pctx);
-       fd_hw_query_fini(pctx);
 
        if (ctx->blitter)
                util_blitter_destroy(ctx->blitter);
 
+       if (pctx->stream_uploader)
+               u_upload_destroy(pctx->stream_uploader);
+
        if (ctx->clear_rs_state)
                pctx->delete_rasterizer_state(pctx, ctx->clear_rs_state);
 
        if (ctx->primconvert)
                util_primconvert_destroy(ctx->primconvert);
 
-       slab_destroy(&ctx->transfer_pool);
+       slab_destroy_child(&ctx->transfer_pool);
 
        for (i = 0; i < ARRAY_SIZE(ctx->pipe); i++) {
                struct fd_vsc_pipe *pipe = &ctx->pipe[i];
@@ -197,7 +210,7 @@ fd_context_setup_common_vbos(struct fd_context *ctx)
                        }});
        ctx->solid_vbuf_state.vertexbuf.count = 1;
        ctx->solid_vbuf_state.vertexbuf.vb[0].stride = 12;
-       ctx->solid_vbuf_state.vertexbuf.vb[0].buffer = ctx->solid_vbuf;
+       ctx->solid_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->solid_vbuf;
 
        /* setup blit_vbuf_state: */
        ctx->blit_vbuf_state.vtx = pctx->create_vertex_elements_state(
@@ -212,9 +225,9 @@ fd_context_setup_common_vbos(struct fd_context *ctx)
                        }});
        ctx->blit_vbuf_state.vertexbuf.count = 2;
        ctx->blit_vbuf_state.vertexbuf.vb[0].stride = 8;
-       ctx->blit_vbuf_state.vertexbuf.vb[0].buffer = ctx->blit_texcoord_vbuf;
+       ctx->blit_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->blit_texcoord_vbuf;
        ctx->blit_vbuf_state.vertexbuf.vb[1].stride = 12;
-       ctx->blit_vbuf_state.vertexbuf.vb[1].buffer = ctx->solid_vbuf;
+       ctx->blit_vbuf_state.vertexbuf.vb[1].buffer.resource = ctx->solid_vbuf;
 }
 
 void
@@ -256,24 +269,23 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
        pctx->flush = fd_context_flush;
        pctx->emit_string_marker = fd_emit_string_marker;
        pctx->set_debug_callback = fd_set_debug_callback;
+       pctx->create_fence_fd = fd_create_fence_fd;
+       pctx->fence_server_sync = fd_fence_server_sync;
 
-       /* TODO what about compute?  Ideally it creates it's own independent
-        * batches per compute job (since it isn't using tiling, so no point
-        * in getting involved with the re-ordering madness)..
-        */
-       if (!screen->reorder) {
-               ctx->batch = fd_bc_alloc_batch(&screen->batch_cache, ctx);
-       }
+       pctx->stream_uploader = u_upload_create_default(pctx);
+       if (!pctx->stream_uploader)
+               goto fail;
+       pctx->const_uploader = pctx->stream_uploader;
 
-       slab_create(&ctx->transfer_pool, sizeof(struct fd_transfer),
-                       16);
+       ctx->batch = fd_bc_alloc_batch(&screen->batch_cache, ctx);
+
+       slab_create_child(&ctx->transfer_pool, &screen->transfer_pool);
 
        fd_draw_init(pctx);
        fd_resource_context_init(pctx);
        fd_query_context_init(pctx);
        fd_texture_init(pctx);
        fd_state_init(pctx);
-       fd_hw_query_init(pctx);
 
        ctx->blitter = util_blitter_create(pctx);
        if (!ctx->blitter)
@@ -283,6 +295,9 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
        if (!ctx->primconvert)
                goto fail;
 
+       list_inithead(&ctx->hw_active_queries);
+       list_inithead(&ctx->acc_active_queries);
+
        return pctx;
 
 fail: