1 files changed, 101 insertions, 33 deletions

diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c
index 6c891029a5..93bf388776 100644
--- a/src/gallium/drivers/r300/r300_render.c
+++ b/src/gallium/drivers/r300/r300_render.c
@@ -30,18 +30,17 @@
 
 #include "util/u_format.h"
 #include "util/u_memory.h"
+#include "util/u_upload_mgr.h"
 #include "util/u_prim.h"
 
 #include "r300_cs.h"
 #include "r300_context.h"
+#include "r300_screen_buffer.h"
 #include "r300_emit.h"
 #include "r300_reg.h"
 #include "r300_render.h"
 #include "r300_state_derived.h"
 
-/* r300_render: Vertex and index buffer primitive emission. */
-#define R300_MAX_VBO_SIZE  (1024 * 1024)
-
 /* XXX The DRM rejects VAP_ALT_NUM_VERTICES.. */
 //#define ENABLE_ALT_NUM_VERTS
 
@@ -123,7 +122,7 @@ static uint32_t r300_provoking_vertex_fixes(struct r300_context *r300,
 static boolean r300_reserve_cs_space(struct r300_context *r300,
                                      unsigned dwords)
 {
-    if (!r300->winsys->check_cs(r300->winsys, dwords)) {
+    if (!r300->rws->check_cs(r300->rws, dwords)) {
         r300->context.flush(&r300->context, 0, NULL);
         return TRUE;
     }
@@ -131,9 +130,36 @@ static boolean r300_reserve_cs_space(struct r300_context *r300,
 }
 
 static boolean immd_is_good_idea(struct r300_context *r300,
-                                      unsigned count)
+                                 unsigned count)
 {
-    return count <= 4;
+    struct pipe_vertex_element* velem;
+    struct pipe_vertex_buffer* vbuf;
+    boolean checked[PIPE_MAX_ATTRIBS] = {0};
+    unsigned vertex_element_count = r300->velems->count;
+    unsigned i, vbi;
+
+    if (count > 10) {
+        return FALSE;
+    }
+
+    /* We shouldn't map buffers referenced by CS, busy buffers,
+     * and ones placed in VRAM. */
+    /* XXX Check for VRAM buffers. */
+    for (i = 0; i < vertex_element_count; i++) {
+        velem = &r300->velems->velem[i];
+        vbi = velem->vertex_buffer_index;
+
+        if (!checked[vbi]) {
+            vbuf = &r300->vertex_buffer[vbi];
+
+            if (r300_buffer_is_referenced(r300, vbuf->buffer)) {
+                /* It's a very bad idea to map it... */
+                return FALSE;
+            }
+            checked[vbi] = TRUE;
+        }
+    }
+    return TRUE;
 }
 
 static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
@@ -143,7 +169,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 {
     struct pipe_vertex_element* velem;
     struct pipe_vertex_buffer* vbuf;
-    unsigned vertex_element_count = r300->vertex_element_count;
+    unsigned vertex_element_count = r300->velems->count;
     unsigned i, v, vbi, dw, elem_offset, dwords;
 
     /* Size of the vertex, in dwords. */
@@ -166,7 +192,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 
     /* Calculate the vertex size, offsets, strides etc. and map the buffers. */
     for (i = 0; i < vertex_element_count; i++) {
-        velem = &r300->vertex_element[i];
+        velem = &r300->velems->velem[i];
         offset[i] = velem->src_offset / 4;
         size[i] = util_format_get_blocksize(velem->src_format) / 4;
         vertex_size += size[i];
@@ -183,18 +209,19 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
         }
     }
 
-    dwords = 10 + count * vertex_size;
+    dwords = 9 + count * vertex_size;
 
     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
-    r300_emit_buffer_validate(r300, FALSE, 0);
+    r300_emit_buffer_validate(r300, FALSE, NULL);
     r300_emit_dirty_state(r300);
 
     BEGIN_CS(dwords);
     OUT_CS_REG(R300_GA_COLOR_CONTROL,
             r300_provoking_vertex_fixes(r300, mode));
     OUT_CS_REG(R300_VAP_VTX_SIZE, vertex_size);
-    OUT_CS_REG(R300_VAP_VF_MIN_VTX_INDX, 0);
-    OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, count - 1);
+    OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
+    OUT_CS(count - 1);
+    OUT_CS(0);
     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_IMMD_2, count * vertex_size);
     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_EMBEDDED | (count << 16) |
             r300_translate_primitive(mode));
@@ -202,7 +229,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
     /* Emit vertices. */
     for (v = 0; v < count; v++) {
         for (i = 0; i < vertex_element_count; i++) {
-            velem = &r300->vertex_element[i];
+            velem = &r300->velems->velem[i];
             vbi = velem->vertex_buffer_index;
             elem_offset = offset[i] + stride[vbi] * (v + start);
 
@@ -215,7 +242,7 @@ static void r300_emit_draw_arrays_immediate(struct r300_context *r300,
 
     /* Unmap buffers. */
     for (i = 0; i < vertex_element_count; i++) {
-        vbi = r300->vertex_element[i].vertex_buffer_index;
+        vbi = r300->velems->velem[i].vertex_buffer_index;
 
         if (map[vbi]) {
             vbuf = &r300->vertex_buffer[vbi];
@@ -238,15 +265,16 @@ static void r300_emit_draw_arrays(struct r300_context *r300,
 
     if (alt_num_verts) {
         assert(count < (1 << 24));
-        BEGIN_CS(10);
+        BEGIN_CS(9);
         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
     } else {
-        BEGIN_CS(8);
+        BEGIN_CS(7);
     }
     OUT_CS_REG(R300_GA_COLOR_CONTROL,
             r300_provoking_vertex_fixes(r300, mode));
-    OUT_CS_REG(R300_VAP_VF_MIN_VTX_INDX, 0);
-    OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, count - 1);
+    OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
+    OUT_CS(count - 1);
+    OUT_CS(0);
     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
     OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (count << 16) |
            r300_translate_primitive(mode) |
@@ -272,24 +300,24 @@ static void r300_emit_draw_elements(struct r300_context *r300,
 #endif
     CS_LOCALS(r300);
 
-    assert((start * indexSize)  % 4 == 0);
     assert(count < (1 << 24));
 
+    maxIndex = MIN2(maxIndex, r300->vertex_buffer_max_index);
+
     DBG(r300, DBG_DRAW, "r300: Indexbuf of %u indices, min %u max %u\n",
         count, minIndex, maxIndex);
 
-    maxIndex = MIN2(maxIndex, ((1 << 24) - 1));
-
     if (alt_num_verts) {
-        BEGIN_CS(16);
+        BEGIN_CS(15);
         OUT_CS_REG(R500_VAP_ALT_NUM_VERTICES, count);
     } else {
-        BEGIN_CS(14);
+        BEGIN_CS(13);
     }
     OUT_CS_REG(R300_GA_COLOR_CONTROL,
             r300_provoking_vertex_fixes(r300, mode));
-    OUT_CS_REG(R300_VAP_VF_MIN_VTX_INDX, minIndex);
-    OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, maxIndex);
+    OUT_CS_REG_SEQ(R300_VAP_VF_MAX_VTX_INDX, 2);
+    OUT_CS(maxIndex);
+    OUT_CS(minIndex);
     OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
     if (indexSize == 4) {
         count_dwords = count;
@@ -313,14 +341,15 @@ static void r300_emit_draw_elements(struct r300_context *r300,
     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
     OUT_CS(offset_dwords << 2);
-    OUT_CS_RELOC(indexBuffer, count_dwords,
-        RADEON_GEM_DOMAIN_GTT, 0, 0);
+    OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
+		     RADEON_GEM_DOMAIN_GTT, 0, 0);
 
     END_CS;
 }
 
 static void r300_shorten_ubyte_elts(struct r300_context* r300,
                                     struct pipe_buffer** elts,
+                                    unsigned start,
                                     unsigned count)
 {
     struct pipe_screen* screen = r300->context.screen;
@@ -338,6 +367,8 @@ static void r300_shorten_ubyte_elts(struct r300_context* r300,
     in_map = pipe_buffer_map(screen, *elts, PIPE_BUFFER_USAGE_CPU_READ);
     out_map = pipe_buffer_map(screen, new_elts, PIPE_BUFFER_USAGE_CPU_WRITE);
 
+    in_map += start;
+
     for (i = 0; i < count; i++) {
         *out_map = (unsigned short)*in_map;
         in_map++;
@@ -350,6 +381,32 @@ static void r300_shorten_ubyte_elts(struct r300_context* r300,
     *elts = new_elts;
 }
 
+static void r300_align_ushort_elts(struct r300_context *r300,
+                                   struct pipe_buffer **elts,
+                                   unsigned start, unsigned count)
+{
+    struct pipe_screen* screen = r300->context.screen;
+    struct pipe_buffer* new_elts;
+    unsigned short *in_map;
+    unsigned short *out_map;
+
+    new_elts = screen->buffer_create(screen, 32,
+                                     PIPE_BUFFER_USAGE_INDEX |
+                                     PIPE_BUFFER_USAGE_CPU_WRITE |
+                                     PIPE_BUFFER_USAGE_GPU_READ,
+                                     2 * count);
+
+    in_map = pipe_buffer_map(screen, *elts, PIPE_BUFFER_USAGE_CPU_READ);
+    out_map = pipe_buffer_map(screen, new_elts, PIPE_BUFFER_USAGE_CPU_WRITE);
+
+    memcpy(out_map, in_map+start, 2 * count);
+
+    pipe_buffer_unmap(screen, *elts);
+    pipe_buffer_unmap(screen, new_elts);
+
+    *elts = new_elts;
+}
+
 /* This is the fast-path drawing & emission for HW TCL. */
 void r300_draw_range_elements(struct pipe_context* pipe,
                               struct pipe_buffer* indexBuffer,
@@ -375,18 +432,26 @@ void r300_draw_range_elements(struct pipe_context* pipe,
     }
 
     if (indexSize == 1) {
-        r300_shorten_ubyte_elts(r300, &indexBuffer, count);
+        r300_shorten_ubyte_elts(r300, &indexBuffer, start, count);
         indexSize = 2;
+        start = 0;
+    } else if (indexSize == 2 && start % 2 != 0) {
+        r300_align_ushort_elts(r300, &indexBuffer, start, count);
+        start = 0;
     }
 
     r300_update_derived_state(r300);
 
+    r300_upload_index_buffer(r300, &indexBuffer, indexSize, start, count);
+
     /* 128 dwords for emit_aos and emit_draw_elements */
     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
     r300_emit_buffer_validate(r300, TRUE, indexBuffer);
     r300_emit_dirty_state(r300);
     r300_emit_aos(r300, 0);
 
+    u_upload_flush(r300->upload_vb);
+    u_upload_flush(r300->upload_ib);
     if (alt_num_verts || count <= 65535) {
         r300_emit_draw_elements(r300, indexBuffer, indexSize, minIndex,
                                 maxIndex, mode, start, count);
@@ -409,7 +474,7 @@ void r300_draw_range_elements(struct pipe_context* pipe,
     }
 
     if (indexBuffer != orgIndexBuffer) {
-        pipe->screen->buffer_destroy(indexBuffer);
+        pipe_buffer_reference( &indexBuffer, NULL );
     }
 }
 
@@ -450,7 +515,7 @@ void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
         /* Make sure there are at least 128 spare dwords in the command buffer.
          * (most of it being consumed by emit_aos) */
         r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 128);
-        r300_emit_buffer_validate(r300, TRUE, 0);
+        r300_emit_buffer_validate(r300, TRUE, NULL);
         r300_emit_dirty_state(r300);
 
         if (alt_num_verts || count <= 65535) {
@@ -468,11 +533,12 @@ void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
                 /* Again, we emit both AOS and draw_arrays so there should be
                  * at least 128 spare dwords. */
                 if (count && r300_reserve_cs_space(r300, 128)) {
-                    r300_emit_buffer_validate(r300, TRUE, 0);
+                    r300_emit_buffer_validate(r300, TRUE, NULL);
                     r300_emit_dirty_state(r300);
                 }
             } while (count);
         }
+	u_upload_flush(r300->upload_vb);
     }
 }
 
@@ -620,9 +686,9 @@ static boolean r300_render_allocate_vertices(struct vbuf_render* render,
         r300render->vbo = pipe_buffer_create(screen,
                                              64,
                                              PIPE_BUFFER_USAGE_VERTEX,
-                                             R300_MAX_VBO_SIZE);
+                                             R300_MAX_DRAW_VBO_SIZE);
         r300render->vbo_offset = 0;
-        r300render->vbo_size = R300_MAX_VBO_SIZE;
+        r300render->vbo_size = R300_MAX_DRAW_VBO_SIZE;
     }
 
     r300render->vertex_size = vertex_size;
@@ -688,6 +754,7 @@ static void r300_render_draw_arrays(struct vbuf_render* render,
     CS_LOCALS(r300);
 
     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + 2);
+    r300_emit_buffer_validate(r300, FALSE, NULL);
     r300_emit_dirty_state(r300);
 
     DBG(r300, DBG_DRAW, "r300: Doing vbuf render, count %d\n", count);
@@ -711,6 +778,7 @@ static void r300_render_draw(struct vbuf_render* render,
     CS_LOCALS(r300);
 
     r300_reserve_cs_space(r300, r300_get_num_dirty_dwords(r300) + dwords);
+    r300_emit_buffer_validate(r300, FALSE, NULL);
     r300_emit_dirty_state(r300);
 
     BEGIN_CS(dwords);