nv40: delay all state emit until before draw
[mesa.git] / src / mesa / pipe / nv40 / nv40_vbo.c
index 4bd6c44e78bd2dac74da2a18f24f2a841c54c2bd..fd1d884193b8d2ecbfae8b649fe3f93813875e18 100644 (file)
@@ -3,9 +3,11 @@
 #include "pipe/p_util.h"
 
 #include "nv40_context.h"
-#include "nv40_dma.h"
 #include "nv40_state.h"
 
+#include "pipe/nouveau/nouveau_channel.h"
+#include "pipe/nouveau/nouveau_pushbuf.h"
+
 static INLINE int
 nv40_vbo_ncomp(uint format)
 {
@@ -44,7 +46,7 @@ nv40_vbo_static_attrib(struct nv40_context *nv40, int attrib,
        type = nv40_vbo_type(ve->src_format);
        ncomp = nv40_vbo_ncomp(ve->src_format);
 
-       map  = ws->buffer_map(ws, vb->buffer, PIPE_BUFFER_FLAG_READ);
+       map  = ws->buffer_map(ws, vb->buffer, PIPE_BUFFER_USAGE_CPU_READ);
        map += vb->buffer_offset + ve->src_offset;
 
        switch (type) {
@@ -95,11 +97,13 @@ nv40_vbo_static_attrib(struct nv40_context *nv40, int attrib,
 }
 
 static void
-nv40_vbo_arrays_update(struct nv40_context *nv40)
+nv40_vbo_arrays_update(struct nv40_context *nv40, struct pipe_buffer *ib,
+                      unsigned ib_format)
 {
        struct nv40_vertex_program *vp = nv40->vertprog.active;
-       uint32_t inputs, vtxfmt[16];
-       int hw, num_hw;
+       struct nouveau_stateobj *vtxbuf, *vtxfmt;
+       unsigned inputs, hw, num_hw;
+       unsigned vb_flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD;
 
        inputs = vp->ir;
        for (hw = 0; hw < 16 && inputs; hw++) {
@@ -110,53 +114,71 @@ nv40_vbo_arrays_update(struct nv40_context *nv40)
        }
        num_hw++;
 
+       vtxbuf = so_new(20, 18);
+       so_method(vtxbuf, nv40->curie, NV40TCL_VTXBUF_ADDRESS(0), num_hw);
+       vtxfmt = so_new(17, 0);
+       so_method(vtxfmt, nv40->curie, NV40TCL_VTXFMT(0), num_hw);
+
        inputs = vp->ir;
        for (hw = 0; hw < num_hw; hw++) {
                struct pipe_vertex_element *ve;
                struct pipe_vertex_buffer *vb;
 
                if (!(inputs & (1 << hw))) {
-                       vtxfmt[hw] = NV40TCL_VTXFMT_TYPE_FLOAT;
+                       so_data(vtxbuf, 0);
+                       so_data(vtxfmt, NV40TCL_VTXFMT_TYPE_FLOAT);
                        continue;
                }
 
                ve = &nv40->vtxelt[hw];
                vb = &nv40->vtxbuf[ve->vertex_buffer_index];
 
-               if (vb->pitch == 0) {
-                       vtxfmt[hw] = NV40TCL_VTXFMT_TYPE_FLOAT;
-                       if (nv40_vbo_static_attrib(nv40, hw, ve, vb) == TRUE)
-                               continue;
+               if (!vb->pitch && nv40_vbo_static_attrib(nv40, hw, ve, vb)) {
+                       so_data(vtxbuf, 0);
+                       so_data(vtxfmt, NV40TCL_VTXFMT_TYPE_FLOAT);
+                       continue;
                }
 
-               BEGIN_RING(curie, NV40TCL_VTXBUF_ADDRESS(hw), 1);
-               OUT_RELOC(vb->buffer, vb->buffer_offset + ve->src_offset,
-                         NOUVEAU_BO_GART | NOUVEAU_BO_VRAM | NOUVEAU_BO_LOW |
-                         NOUVEAU_BO_OR | NOUVEAU_BO_RD, 0,
-                         NV40TCL_VTXBUF_ADDRESS_DMA1);
-               vtxfmt[hw] = ((vb->pitch << NV40TCL_VTXFMT_STRIDE_SHIFT) |
-                             (nv40_vbo_ncomp(ve->src_format) <<
-                              NV40TCL_VTXFMT_SIZE_SHIFT) |
-                             nv40_vbo_type(ve->src_format));
+               so_reloc(vtxbuf, vb->buffer, vb->buffer_offset + ve->src_offset,
+                        vb_flags | NOUVEAU_BO_LOW | NOUVEAU_BO_OR,
+                        0, NV40TCL_VTXBUF_ADDRESS_DMA1);
+               so_data (vtxfmt, ((vb->pitch << NV40TCL_VTXFMT_STRIDE_SHIFT) |
+                                 (nv40_vbo_ncomp(ve->src_format) <<
+                                  NV40TCL_VTXFMT_SIZE_SHIFT) |
+                                 nv40_vbo_type(ve->src_format)));
        }
 
-       BEGIN_RING(curie, 0x1710, 1);
-       OUT_RING  (0); /* vtx cache flush */
-       BEGIN_RING(curie, NV40TCL_VTXFMT(0), num_hw);
-       OUT_RINGp (vtxfmt, num_hw);
+       if (ib) {
+               so_method(vtxbuf, nv40->curie, NV40TCL_IDXBUF_ADDRESS, 2);
+               so_reloc (vtxbuf, ib, 0, vb_flags | NOUVEAU_BO_LOW, 0, 0);
+               so_reloc (vtxbuf, ib, ib_format, vb_flags | NOUVEAU_BO_OR,
+                         0, NV40TCL_IDXBUF_FORMAT_DMA1);
+       }
+
+       so_emit(nv40->nvws, vtxfmt);
+       so_emit(nv40->nvws, vtxbuf);
+       so_ref (vtxbuf, &nv40->so_vtxbuf);
+       so_ref (NULL, &vtxbuf);
+       so_ref (NULL, &vtxfmt);
 }
 
 static boolean
-nv40_vbo_validate_state(struct nv40_context *nv40)
+nv40_vbo_validate_state(struct nv40_context *nv40,
+                       struct pipe_buffer *ib, unsigned ib_format)
 {
-       if (nv40->dirty & ~NV40_NEW_ARRAYS)
-               nv40_emit_hw_state(nv40);
+       unsigned vdn = nv40->dirty & NV40_NEW_ARRAYS;
 
-       if (nv40->dirty & NV40_NEW_ARRAYS) {
-               nv40_vbo_arrays_update(nv40);
+       nv40_emit_hw_state(nv40);
+       if (vdn || ib) {
+               nv40_vbo_arrays_update(nv40, ib, ib_format);
                nv40->dirty &= ~NV40_NEW_ARRAYS;
        }
 
+       so_emit_reloc_markers(nv40->nvws, nv40->so_vtxbuf);
+
+       BEGIN_RING(curie, 0x1710, 1);
+       OUT_RING  (0); /* vtx cache flush */
+
        return TRUE;
 }
 
@@ -164,10 +186,10 @@ boolean
 nv40_draw_arrays(struct pipe_context *pipe, unsigned mode, unsigned start,
                 unsigned count)
 {
-       struct nv40_context *nv40 = (struct nv40_context *)pipe;
+       struct nv40_context *nv40 = nv40_context(pipe);
        unsigned nr;
 
-       assert(nv40_vbo_validate_state(nv40));
+       assert(nv40_vbo_validate_state(nv40, NULL, 0));
 
        BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
        OUT_RING  (nvgl_primitive(mode));
@@ -267,45 +289,115 @@ nv40_draw_elements_u32(struct nv40_context *nv40, void *ib,
        }
 }
 
-boolean
-nv40_draw_elements(struct pipe_context *pipe,
-                  struct pipe_buffer_handle *indexBuffer, unsigned indexSize,
-                  unsigned mode, unsigned start, unsigned count)
+static boolean
+nv40_draw_elements_inline(struct pipe_context *pipe,
+                         struct pipe_buffer *ib, unsigned ib_size,
+                         unsigned mode, unsigned start, unsigned count)
 {
-       struct nv40_context *nv40 = (struct nv40_context *)pipe;
-       void *ib;
+       struct nv40_context *nv40 = nv40_context(pipe);
+       struct pipe_winsys *ws = pipe->winsys;
+       void *map;
 
-       assert(nv40_vbo_validate_state(nv40));
+       assert(nv40_vbo_validate_state(nv40, NULL, 0));
 
-       ib = pipe->winsys->buffer_map(pipe->winsys, indexBuffer,
-                                     PIPE_BUFFER_FLAG_READ);
-       if (!ib) {
-               NOUVEAU_ERR("Couldn't map index buffer!!\n");
-               return FALSE;
-       }
+       map = ws->buffer_map(ws, ib, PIPE_BUFFER_USAGE_CPU_READ);
+       if (!ib)
+               assert(0);
 
        BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
        OUT_RING  (nvgl_primitive(mode));
 
-       switch (indexSize) {
+       switch (ib_size) {
        case 1:
-               nv40_draw_elements_u08(nv40, ib, start, count);
+               nv40_draw_elements_u08(nv40, map, start, count);
                break;
        case 2:
-               nv40_draw_elements_u16(nv40, ib, start, count);
+               nv40_draw_elements_u16(nv40, map, start, count);
                break;
        case 4:
-               nv40_draw_elements_u32(nv40, ib, start, count);
+               nv40_draw_elements_u32(nv40, map, start, count);
                break;
        default:
-               NOUVEAU_ERR("unsupported elt size %d\n", indexSize);
+               assert(0);
                break;
        }
 
        BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
        OUT_RING  (0);
 
-       pipe->winsys->buffer_unmap(pipe->winsys, indexBuffer);
+       ws->buffer_unmap(ws, ib);
+
+       return TRUE;
+}
+
+static boolean
+nv40_draw_elements_vbo(struct pipe_context *pipe,
+                      struct pipe_buffer *ib, unsigned ib_size,
+                      unsigned mode, unsigned start, unsigned count)
+{
+       struct nv40_context *nv40 = nv40_context(pipe);
+       unsigned nr, type;
+
+       switch (ib_size) {
+       case 2:
+               type = NV40TCL_IDXBUF_FORMAT_TYPE_U16;
+               break;
+       case 4:
+               type = NV40TCL_IDXBUF_FORMAT_TYPE_U32;
+               break;
+       default:
+               assert(0);
+       }
+
+       assert(nv40_vbo_validate_state(nv40, ib, type));
+
+       BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
+       OUT_RING  (nvgl_primitive(mode));
+
+       nr = (count & 0xff);
+       if (nr) {
+               BEGIN_RING(curie, NV40TCL_VB_INDEX_BATCH, 1);
+               OUT_RING  (((nr - 1) << 24) | start);
+               start += nr;
+       }
+
+       nr = count >> 8;
+       while (nr) {
+               unsigned push = nr > 2047 ? 2047 : nr;
+
+               nr -= push;
+
+               BEGIN_RING_NI(curie, NV40TCL_VB_INDEX_BATCH, push);
+               while (push--) {
+                       OUT_RING(((0x100 - 1) << 24) | start);
+                       start += 0x100;
+               }
+       }
+
+       BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
+       OUT_RING  (0);
+
+       return TRUE;
+}
+
+boolean
+nv40_draw_elements(struct pipe_context *pipe,
+                  struct pipe_buffer *indexBuffer, unsigned indexSize,
+                  unsigned mode, unsigned start, unsigned count)
+{
+       struct nv40_context *nv40 = nv40_context(pipe);
+
+       /* 0x4497 doesn't support real index buffers, and there doesn't appear
+        * to be support on any chipset for 8-bit indices.
+        */
+       if (nv40->curie->grclass == NV44TCL || indexSize == 1) {
+               nv40_draw_elements_inline(pipe, indexBuffer, indexSize,
+                                         mode, start, count);
+       } else {
+               nv40_draw_elements_vbo(pipe, indexBuffer, indexSize,
+                                      mode, start, count);
+       }
+
        pipe->flush(pipe, 0);
        return TRUE;
 }