nvfx: emit bo relocations only when needed

Should improve performance, possibly significantly.

diff --git a/src/gallium/drivers/nvfx/nvfx_context.c b/src/gallium/drivers/nvfx/nvfx_context.c
index 2f775f92cf5b68e7ae181887b1285b8ddba0041a..5a2fa14c887b773070ae1de4c8af2840b2f39d88 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_context.c
+++ b/src/gallium/drivers/nvfx/nvfx_context.c
@@ -46,6 +46,9 @@ nvfx_destroy(struct pipe_context *pipe)
        if (nvfx->draw)
                draw_destroy(nvfx->draw);
 
+       if(nvfx->screen->cur_ctx == nvfx)
+               nvfx->screen->cur_ctx = NULL;
+
        FREE(nvfx);
 }
 
@@ -72,8 +75,6 @@ nvfx_create(struct pipe_screen *pscreen, void *priv)
        nvfx->pipe.clear = nvfx_clear;
        nvfx->pipe.flush = nvfx_flush;
 
-       screen->base.channel->user_private = nvfx;
-
        nvfx->is_nv4x = screen->is_nv4x;
        /* TODO: it seems that nv30 might have fixed function clipping usable with vertex programs
         * However, my code for that doesn't work, so use vp clipping for all cards, which works.
@@ -103,6 +104,7 @@ nvfx_create(struct pipe_screen *pscreen, void *priv)
        nvfx->hw_pointsprite_control = -1;
        nvfx->hw_vp_output = -1;
        nvfx->use_vertex_buffers = -1;
+       nvfx->relocs_needed = NVFX_RELOCATE_ALL;
 
        LIST_INITHEAD(&nvfx->render_cache);
 

diff --git a/src/gallium/drivers/nvfx/nvfx_context.h b/src/gallium/drivers/nvfx/nvfx_context.h
index 680f4c6ce0f45f02e0ddb01b352a047bd021a8c4..4c654bfa8bab1ca4af6b104f21e0c6c24a7b38ec 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_context.h
+++ b/src/gallium/drivers/nvfx/nvfx_context.h
@@ -47,6 +47,13 @@
 #define NVFX_NEW_INDEX (1 << 16)
 #define NVFX_NEW_SPRITE  (1 << 17)
 
+#define NVFX_RELOCATE_FRAMEBUFFER (1 << 0)
+#define NVFX_RELOCATE_FRAGTEX (1 << 1)
+#define NVFX_RELOCATE_FRAGPROG (1 << 2)
+#define NVFX_RELOCATE_VTXBUF (1 << 3)
+#define NVFX_RELOCATE_IDXBUF (1 << 4)
+#define NVFX_RELOCATE_ALL 0x1f
+
 struct nvfx_rasterizer_state {
        struct pipe_rasterizer_state pipe;
        unsigned sb_len;
@@ -199,6 +206,8 @@ struct nvfx_context {
        int hw_pointsprite_control;
        int hw_vp_output;
        struct nvfx_fragment_program* hw_fragprog;
+
+       unsigned relocs_needed;
 };
 
 static INLINE struct nvfx_context *
@@ -290,10 +299,25 @@ extern void nvfx_state_sr_validate(struct nvfx_context *nvfx);
 extern void nvfx_state_zsa_validate(struct nvfx_context *nvfx);
 
 /* nvfx_state_emit.c */
-extern void nvfx_state_relocate(struct nvfx_context *nvfx);
+extern void nvfx_state_relocate(struct nvfx_context *nvfx, unsigned relocs);
 extern boolean nvfx_state_validate(struct nvfx_context *nvfx);
 extern boolean nvfx_state_validate_swtnl(struct nvfx_context *nvfx);
-extern void nvfx_state_emit(struct nvfx_context *nvfx);
+
+static inline void
+nvfx_state_emit(struct nvfx_context *nvfx)
+{
+        unsigned relocs = NVFX_RELOCATE_FRAMEBUFFER | NVFX_RELOCATE_FRAGTEX | NVFX_RELOCATE_FRAGPROG;
+        if (nvfx->render_mode == HW)
+        {
+                relocs |= NVFX_RELOCATE_VTXBUF;
+                if(nvfx->use_index_buffer)
+                        relocs |= NVFX_RELOCATE_IDXBUF;
+        }
+
+        relocs &= nvfx->relocs_needed;
+        if(relocs)
+                nvfx_state_relocate(nvfx, relocs);
+}
 
 /* nvfx_transfer.c */
 extern void nvfx_init_transfer_functions(struct pipe_context *pipe);

diff --git a/src/gallium/drivers/nvfx/nvfx_fragprog.c b/src/gallium/drivers/nvfx/nvfx_fragprog.c
index 049b814d49ff716a2e774e07c7bdbc22a54cdb70..7caddfab707419742bcced49fb3018faea315280 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_fragprog.c
+++ b/src/gallium/drivers/nvfx/nvfx_fragprog.c
@@ -1471,6 +1471,8 @@ update:
                        nvfx->hw_pointsprite_control = pointsprite_control;
                }
        }
+
+       nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAGPROG;
 }
 
 void
@@ -1487,6 +1489,7 @@ nvfx_fragprog_relocate(struct nvfx_context *nvfx)
        OUT_RELOC(chan, bo, offset, fp_flags | NOUVEAU_BO_LOW |
                      NOUVEAU_BO_OR, NV34TCL_FP_ACTIVE_PROGRAM_DMA0,
                      NV34TCL_FP_ACTIVE_PROGRAM_DMA1);
+       nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAGPROG;
 }
 
 void

diff --git a/src/gallium/drivers/nvfx/nvfx_fragtex.c b/src/gallium/drivers/nvfx/nvfx_fragtex.c
index 00c47be76ab9479e5fb24ed993b284c0e0105116..6503c7afcbf17aaf4f98b5a34ae7182c9b4112c9 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_fragtex.c
+++ b/src/gallium/drivers/nvfx/nvfx_fragtex.c
@@ -205,6 +205,7 @@ nvfx_fragtex_validate(struct nvfx_context *nvfx)
                }
        }
        nvfx->dirty_samplers = 0;
+       nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAGTEX;
 }
 
 void
@@ -231,6 +232,7 @@ nvfx_fragtex_relocate(struct nvfx_context *nvfx)
                OUT_RELOC(chan, bo, nvfx->hw_txf[unit], tex_flags | NOUVEAU_BO_OR | NOUVEAU_BO_DUMMY,
                                NV34TCL_TX_FORMAT_DMA0, NV34TCL_TX_FORMAT_DMA1);
        }
+       nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAGTEX;
 }
 
 void

diff --git a/src/gallium/drivers/nvfx/nvfx_screen.c b/src/gallium/drivers/nvfx/nvfx_screen.c
index 72cb5239b5a9922f466defbc2da09c79372a38c8..99b4d8b58c090e0a009a42c20dc4cef3e6a4745e 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_screen.c
+++ b/src/gallium/drivers/nvfx/nvfx_screen.c
@@ -377,6 +377,14 @@ nvfx_screen_get_vertex_buffer_flags(struct nvfx_screen* screen)
        return vram_hack ? NOUVEAU_BO_VRAM : NOUVEAU_BO_GART;
 }
 
+static void nvfx_channel_flush_notify(struct nouveau_channel* chan)
+{
+       struct nvfx_screen* screen = chan->user_private;
+       struct nvfx_context* nvfx = screen->cur_ctx;
+       if(nvfx)
+               nvfx->relocs_needed = NVFX_RELOCATE_ALL;
+}
+
 struct pipe_screen *
 nvfx_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
 {
@@ -398,6 +406,9 @@ nvfx_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
                return NULL;
        }
        chan = screen->base.channel;
+       screen->cur_ctx = NULL;
+       chan->user_private = screen;
+       chan->flush_notify = nvfx_channel_flush_notify;
 
        pscreen->winsys = ws;
        pscreen->destroy = nvfx_screen_destroy;

diff --git a/src/gallium/drivers/nvfx/nvfx_state_emit.c b/src/gallium/drivers/nvfx/nvfx_state_emit.c
index c43a75aaa213ff162a063eb311d690dde981ce8c..cfcb0f7ef66a5c4c22085652f00a85d25fd95ae7 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_state_emit.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_emit.c
@@ -20,6 +20,7 @@ nvfx_state_validate_common(struct nvfx_context *nvfx)
                nvfx->hw_pointsprite_control = -1;
                nvfx->hw_vp_output = -1;
                nvfx->screen->cur_ctx = nvfx;
+               nvfx->relocs_needed = NVFX_RELOCATE_ALL;
        }
 
        /* These can trigger use the of 3D engine to copy temporaries.
@@ -244,12 +245,12 @@ nvfx_state_validate_common(struct nvfx_context *nvfx)
        return TRUE;
 }
 
-void
-nvfx_state_emit(struct nvfx_context *nvfx)
+inline void
+nvfx_state_relocate(struct nvfx_context *nvfx, unsigned relocs)
 {
        struct nouveau_channel* chan = nvfx->screen->base.channel;
        /* we need to ensure there is enough space to output relocations in one go */
-       unsigned max_relocs = 0
+       const unsigned max_relocs = 0
              + 16 /* vertex buffers, incl. dma flag */
              + 2 /* index buffer plus format+dma flag */
              + 2 * 5 /* 4 cbufs + zsbuf, plus dma objects */
@@ -257,22 +258,19 @@ nvfx_state_emit(struct nvfx_context *nvfx)
              + 2 * 4 /* vertex textures plus format+dma flag */
              + 1 /* fragprog incl dma flag */
              ;
+
        MARK_RING(chan, max_relocs * 2, max_relocs * 2);
-       nvfx_state_relocate(nvfx);
-}
 
-void
-nvfx_state_relocate(struct nvfx_context *nvfx)
-{
-       nvfx_framebuffer_relocate(nvfx);
-       nvfx_fragtex_relocate(nvfx);
-       nvfx_fragprog_relocate(nvfx);
-       if (nvfx->render_mode == HW)
-       {
+       if(relocs & NVFX_RELOCATE_FRAMEBUFFER)
+               nvfx_framebuffer_relocate(nvfx);
+       if(relocs & NVFX_RELOCATE_FRAGTEX)
+               nvfx_fragtex_relocate(nvfx);
+       if(relocs & NVFX_RELOCATE_FRAGPROG)
+               nvfx_fragprog_relocate(nvfx);
+       if(relocs & NVFX_RELOCATE_VTXBUF)
                nvfx_vbo_relocate(nvfx);
-               if(nvfx->use_index_buffer)
-                       nvfx_idxbuf_relocate(nvfx);
-       }
+       if(relocs & NVFX_RELOCATE_IDXBUF)
+               nvfx_idxbuf_relocate(nvfx);
 }
 
 boolean

diff --git a/src/gallium/drivers/nvfx/nvfx_state_fb.c b/src/gallium/drivers/nvfx/nvfx_state_fb.c
index 3db9cec905454b150f3d30b33c8e8749f1fe8025..3b869d43a15a2ce62dc39214067e442a1c8c7d7b 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_state_fb.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_fb.c
@@ -278,6 +278,7 @@ nvfx_framebuffer_validate(struct nvfx_context *nvfx, unsigned prepare_result)
                OUT_RING(chan, RING_3D(NV34TCL_VIEWPORT_TX_ORIGIN, 1));
                OUT_RING(chan, 0);
        }
+       nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAMEBUFFER;
 }
 
 void
@@ -307,4 +308,5 @@ nvfx_framebuffer_relocate(struct nvfx_context *nvfx)
        DO(NV40, 3);
 
        DO_(nvfx->hw_zeta, NV34, ZETA);
+       nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAMEBUFFER;
 }

diff --git a/src/gallium/drivers/nvfx/nvfx_vbo.c b/src/gallium/drivers/nvfx/nvfx_vbo.c
index 21d6e0e6f84d4751cf8edad4c9ed5b05564ee740..e6e9a8f2e4019ea3907e69b172067e3551e0b893 100644 (file)
--- a/src/gallium/drivers/nvfx/nvfx_vbo.c
+++ b/src/gallium/drivers/nvfx/nvfx_vbo.c
@@ -334,6 +334,7 @@ nvfx_vbo_validate(struct nvfx_context *nvfx)
        OUT_RING(chan, 0);
 
        nvfx->hw_vtxelt_nr = nvfx->vtxelt->num_elements;
+       nvfx->relocs_needed &=~ NVFX_RELOCATE_VTXBUF;
        return TRUE;
 }
 
@@ -362,6 +363,7 @@ nvfx_vbo_relocate(struct nvfx_context *nvfx)
                                vb_flags | NOUVEAU_BO_LOW | NOUVEAU_BO_OR,
                                0, NV34TCL_VTXBUF_ADDRESS_DMA1);
        }
+        nvfx->relocs_needed &=~ NVFX_RELOCATE_VTXBUF;
 }
 
 static void
@@ -382,6 +384,7 @@ nvfx_idxbuf_emit(struct nvfx_context* nvfx, unsigned ib_flags)
        OUT_RELOC(chan, bo, nvfx->idxbuf.offset + 1, ib_flags | NOUVEAU_BO_LOW, 0, 0);
        OUT_RELOC(chan, bo, ib_format, ib_flags | NOUVEAU_BO_OR,
                        0, NV34TCL_IDXBUF_FORMAT_DMA1);
+       nvfx->relocs_needed &=~ NVFX_RELOCATE_IDXBUF;
 }
 
 void