draw: add streamlined paths for fetching linear verts

3 files changed, 134 insertions, 66 deletions

diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c
index a556477a767..4c794e0e23c 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.c
@@ -92,9 +92,9 @@ struct x86_reg aos_get_x86( struct aos_compilation *cp,
          assert(which_reg == 1);
          offset = Offset(struct aos_machine, constants);
          break;
-      case X86_ATTRIBS:
+      case X86_BUFFERS:
          assert(which_reg == 0);
-         offset = Offset(struct aos_machine, attrib);
+         offset = Offset(struct aos_machine, buffer);
          break;
       default:
          assert(0);
@@ -1939,6 +1939,8 @@ static boolean build_vertex_program( struct draw_vs_varient_aos_sse *varient,
    save_fpu_state( &cp );
    set_fpu_round_nearest( &cp );
 
+   aos_init_inputs( &cp, linear );
+
    /* Note address for loop jump 
     */
    label = x86_get_label(cp.func);
@@ -2018,13 +2020,7 @@ static boolean build_vertex_program( struct draw_vs_varient_aos_sse *varient,
 
       /* Incr index
        */   
-      if (linear) {
-         x86_inc(cp.func, cp.idx_EBX);
-      } 
-      else {
-         x86_lea(cp.func, cp.idx_EBX, x86_make_disp(cp.idx_EBX, 4));
-      }
-
+      aos_incr_inputs( &cp, linear );
    }
    /* decr count, loop if not zero
     */
@@ -2065,14 +2061,10 @@ static void vaos_set_buffer( struct draw_vs_varient *varient,
                              unsigned stride )
 {
    struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
-   unsigned i;
 
-   for (i = 0; i < vaos->base.key.nr_inputs; i++) {
-      if (vaos->base.key.element[i].in.buffer == buf) {
-         vaos->attrib[i].input_ptr = ((char *)ptr +
-                                      vaos->base.key.element[i].in.offset);
-         vaos->attrib[i].input_stride = stride;
-      }
+   if (buf < vaos->nr_vb) {
+      vaos->buffer[buf].base_ptr = (char *)ptr;
+      vaos->buffer[buf].stride = stride;
    }
 }
 
@@ -2089,7 +2081,7 @@ static void PIPE_CDECL vaos_run_elts( struct draw_vs_varient *varient,
    machine->internal[IMM_PSIZE][0] = vaos->draw->rasterizer->point_size;
    machine->constants = vaos->draw->vs.aligned_constants;
    machine->immediates = vaos->base.vs->immediates;
-   machine->attrib = vaos->attrib;
+   machine->buffer = vaos->buffer;
 
    vaos->gen_run_elts( machine,
                        elts,
@@ -2108,7 +2100,7 @@ static void PIPE_CDECL vaos_run_linear( struct draw_vs_varient *varient,
    machine->internal[IMM_PSIZE][0] = vaos->draw->rasterizer->point_size;
    machine->constants = vaos->draw->vs.aligned_constants;
    machine->immediates = vaos->base.vs->immediates;
-   machine->attrib = vaos->attrib;
+   machine->buffer = vaos->buffer;
 
    vaos->gen_run_linear( machine,
                          start,
@@ -2127,7 +2119,7 @@ static void vaos_destroy( struct draw_vs_varient *varient )
 {
    struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
 
-   FREE( vaos->attrib );
+   FREE( vaos->buffer );
 
    x86_release_func( &vaos->func[0] );
    x86_release_func( &vaos->func[1] );
@@ -2140,6 +2132,7 @@ static void vaos_destroy( struct draw_vs_varient *varient )
 static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs,
                                                  const struct draw_vs_varient_key *key )
 {
+   unsigned i;
    struct draw_vs_varient_aos_sse *vaos = CALLOC_STRUCT(draw_vs_varient_aos_sse);
 
    if (!vaos)
@@ -2154,10 +2147,15 @@ static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs,
 
    vaos->draw = vs->draw;
 
-   vaos->attrib = MALLOC( key->nr_inputs * sizeof(vaos->attrib[0]) );
-   if (!vaos->attrib)
+   for (i = 0; i < key->nr_inputs; i++) 
+      vaos->nr_vb = MAX2( vaos->nr_vb, key->element[i].in.buffer + 1 );
+
+   vaos->buffer = MALLOC( vaos->nr_vb * sizeof(vaos->buffer[0]) );
+   if (!vaos->buffer)
       goto fail;
 
+   debug_printf("nr_vb: %d\n", vaos->nr_vb);
+
 #if 0
    tgsi_dump(vs->state.tokens, 0);
 #endif
@@ -2179,8 +2177,8 @@ static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs,
    return &vaos->base;
 
  fail:
-   if (vaos && vaos->attrib)
-      FREE(vaos->attrib);
+   if (vaos && vaos->buffer)
+      FREE(vaos->buffer);
 
    if (vaos)
       x86_release_func( &vaos->func[0] );
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.h b/src/gallium/auxiliary/draw/draw_vs_aos.h
index 7fe6f79db0d..306392e5d6d 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.h
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.h
@@ -87,9 +87,10 @@ struct lit_info {
 #define MAX_SHINE_TAB    4
 #define MAX_LIT_INFO     16
 
-struct aos_attrib {
-   const void *input_ptr;
-   unsigned input_stride;
+struct aos_buffer {
+   const void *base_ptr;
+   unsigned stride;
+   void *ptr;                   /* updated per vertex */
 };
 
 
@@ -123,7 +124,7 @@ struct aos_machine {
    const float (*immediates)[4];     /* points to shader data */
    const float (*constants)[4];      /* points to draw data */
 
-   const struct aos_attrib *attrib; /* points to ? */
+   const struct aos_buffer *buffer; /* points to ? */
 };
 
 
@@ -179,8 +180,9 @@ struct x86_reg aos_get_shader_reg( struct aos_compilation *cp,
                                    unsigned file,
                                    unsigned idx );
 
-boolean aos_fetch_inputs( struct aos_compilation *cp,
-                          boolean linear );
+boolean aos_init_inputs( struct aos_compilation *cp, boolean linear );
+boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear );
+boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear );
 
 boolean aos_emit_outputs( struct aos_compilation *cp );
 
@@ -210,7 +212,7 @@ do {                                                                    \
 #define X86_NULL       0
 #define X86_IMMEDIATES 1
 #define X86_CONSTANTS  2
-#define X86_ATTRIBS    3
+#define X86_BUFFERS    3
 
 struct x86_reg aos_get_x86( struct aos_compilation *cp,
                             unsigned which_reg,
@@ -232,7 +234,8 @@ struct draw_vs_varient_aos_sse {
    struct draw_vs_varient base;
    struct draw_context *draw;
 
-   struct aos_attrib *attrib;
+   struct aos_buffer *buffer;
+   unsigned nr_vb;
 
    vaos_run_linear_func gen_run_linear;
    vaos_run_elts_func gen_run_elts;
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c
index 26297c74f82..8e08b9285f9 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c
@@ -95,28 +95,6 @@ static void emit_load_R8G8B8A8_UNORM( struct aos_compilation *cp,
 
 
 
-static void get_src_ptr( struct aos_compilation *cp,
-                         struct x86_reg src,
-                         struct x86_reg elt,
-                         unsigned a )
-{
-   struct x86_reg attrib = x86_make_disp(aos_get_x86( cp, 0, X86_ATTRIBS ), 
-                                         a * sizeof(struct aos_attrib));
-
-   struct x86_reg input_ptr = x86_make_disp(attrib, 
-                                            Offset(struct aos_attrib, input_ptr));
-
-   struct x86_reg input_stride = x86_make_disp(attrib, 
-                                               Offset(struct aos_attrib, input_stride));
-
-   /* Calculate pointer to current attrib:
-    */
-   x86_mov(cp->func, src, input_stride);
-   x86_imul(cp->func, src, elt);
-   x86_add(cp->func, src, input_ptr);
-}
-
-
 /* Extended swizzles?  Maybe later.
  */  
 static void emit_swizzle( struct aos_compilation *cp,
@@ -128,22 +106,44 @@ static void emit_swizzle( struct aos_compilation *cp,
 }
 
 
+
+static boolean get_buffer_ptr( struct aos_compilation *cp,
+                            unsigned buf_idx,
+                            struct x86_reg elt,
+                            struct x86_reg ptr)
+{
+   struct x86_reg buf = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ), 
+                                      buf_idx * sizeof(struct aos_buffer));
+
+   struct x86_reg buf_base_ptr = x86_make_disp(buf, 
+                                               Offset(struct aos_buffer, base_ptr));
+
+   struct x86_reg buf_stride = x86_make_disp(buf, 
+                                             Offset(struct aos_buffer, stride));
+
+   /* Calculate pointer to current attrib:
+    */
+   x86_mov(cp->func, ptr, buf_stride);
+   x86_imul(cp->func, ptr, elt);
+   x86_add(cp->func, ptr, buf_base_ptr);
+
+   return TRUE;
+}
+
+
+
+
 static boolean load_input( struct aos_compilation *cp,
                            unsigned idx,
-                           boolean linear )
+                           struct x86_reg bufptr )
 {
    unsigned format = cp->vaos->base.key.element[idx].in.format;
-   struct x86_reg src = cp->tmp_EAX;
+   unsigned offset = cp->vaos->base.key.element[idx].in.offset;
    struct x86_reg dataXMM = aos_get_xmm_reg(cp);
 
    /* Figure out source pointer address:
     */
-   get_src_ptr(cp, 
-               src, 
-               linear ? cp->idx_EBX : x86_deref(cp->idx_EBX),
-               idx);
-
-   src = x86_deref(src);
+   struct x86_reg src = x86_make_disp(bufptr, offset);
 
    aos_adopt_xmm_reg( cp,
                       dataXMM,
@@ -179,20 +179,87 @@ static boolean load_input( struct aos_compilation *cp,
    return TRUE;
 }
 
-
-boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear )
+static boolean load_inputs( struct aos_compilation *cp,
+                            unsigned buffer,
+                            struct x86_reg ptr )
 {
    unsigned i;
-   
+
    for (i = 0; i < cp->vaos->base.key.nr_inputs; i++) {
-      if (!load_input( cp, i, linear ))
+      if (cp->vaos->base.key.element[i].in.buffer == buffer) {
+
+         if (!load_input( cp, i, ptr ))
+            return FALSE;
+
+         cp->insn_counter++;
+      }
+   }
+   
+   return TRUE;
+}
+
+boolean aos_init_inputs( struct aos_compilation *cp, boolean linear )
+{
+   if (linear && cp->vaos->nr_vb == 1) {
+
+      struct x86_reg elt = cp->idx_EBX;
+      struct x86_reg ptr = cp->tmp_EAX;
+
+      if (!get_buffer_ptr( cp, 0, elt, ptr ))
          return FALSE;
-      cp->insn_counter++;
+
+      /* In the linear, single buffer case, keep the buffer pointer
+       * instead of the index number.
+       */
+      x86_mov( cp->func, elt, ptr );
+   }
+
+   return TRUE;
+}
+
+boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear )
+{
+   if (linear && cp->vaos->nr_vb == 1) {
+      
+      load_inputs( cp, 0, cp->idx_EBX );
+
+   }
+   else {
+      struct x86_reg elt = linear ? cp->idx_EBX : x86_deref(cp->idx_EBX);
+      unsigned j;
+   
+      for (j = 0; j < cp->vaos->nr_vb; j++) {
+         struct x86_reg ptr = cp->tmp_EAX;
+
+         if (!get_buffer_ptr( cp, j, elt, ptr ))
+            return FALSE;
+
+         cp->insn_counter++;
+
+         if (!load_inputs( cp, j, ptr ))
+            return FALSE;
+      }
    }
 
    return TRUE;
 }
 
+boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear )
+{
+   if (linear && cp->vaos->nr_vb == 1) {
+      struct x86_reg stride = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ), 
+                                            (0 * sizeof(struct aos_buffer) + 
+                                             Offset(struct aos_buffer, stride)));
+
+      x86_add(cp->func, cp->idx_EBX, stride);
+   }
+   else if (linear) {
+      x86_inc(cp->func, cp->idx_EBX);
+   } 
+   else {
+      x86_lea(cp->func, cp->idx_EBX, x86_make_disp(cp->idx_EBX, 4));
+   }
+}