X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fauxiliary%2Fdraw%2Fdraw_vs_aos.c;h=6817f29c2a9007ff252f6d8b34b9011aaccb358e;hb=27e5097e2aa6ee2b9cc8dbc5129fa7f4c8eb283c;hp=a556477a7677fc9c4eed17c3136b1142fe3c7a5a;hpb=2cb3d2b7a0d8f505131e98053d3e19982012fccd;p=mesa.git

diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c
index a556477a767..6817f29c2a9 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);
@@ -196,6 +196,18 @@ static void spill( struct aos_compilation *cp, unsigned idx )
 }
 
 
+void aos_spill_all( struct aos_compilation *cp )
+{
+   unsigned i;
+
+   for (i = 0; i < 8; i++) {
+      if (cp->xmm[i].dirty) 
+         spill(cp, i);
+      aos_release_xmm_reg(cp, i);
+   }
+}
+
+
 static struct x86_reg get_xmm_writable( struct aos_compilation *cp,
                                         struct x86_reg reg )
 {
@@ -859,7 +871,7 @@ static void set_fpu_round_nearest( struct aos_compilation *cp )
    }
 }
 
-
+#if 0
 static void x87_emit_ex2( struct aos_compilation *cp )
 {
    struct x86_reg st0 = x86_make_reg(file_x87, 0);
@@ -882,13 +894,17 @@ static void x87_emit_ex2( struct aos_compilation *cp )
    assert( stack == cp->func->x87_stack);
       
 }
+#endif
 
+#if 0
 static void PIPE_CDECL print_reg( const char *msg,
                                   const float *reg )
 {
    debug_printf("%s: %f %f %f %f\n", msg, reg[0], reg[1], reg[2], reg[3]);
 }
+#endif
 
+#if 0
 static void emit_print( struct aos_compilation *cp,
                         const char *message, /* must point to a static string! */
                         unsigned file,
@@ -940,6 +956,7 @@ static void emit_print( struct aos_compilation *cp,
    /* Done... 
     */
 }
+#endif
 
 /**
  * The traditional instructions.  All operate on internal registers
@@ -1078,7 +1095,7 @@ static boolean emit_LG2( struct aos_compilation *cp, const struct tgsi_full_inst
    return TRUE;
 }
 
-
+#if 0
 static boolean emit_EX2( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
 {
    x87_fld_src(cp, &op->FullSrcRegisters[0], 0);
@@ -1086,6 +1103,7 @@ static boolean emit_EX2( struct aos_compilation *cp, const struct tgsi_full_inst
    x87_fstp_dest4(cp, &op->FullDstRegisters[0]);
    return TRUE;
 }
+#endif
 
 
 static boolean emit_FLR( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
@@ -1620,6 +1638,17 @@ static boolean emit_SUB( struct aos_compilation *cp, const struct tgsi_full_inst
    return TRUE;
 }
 
+static boolean emit_TRUNC( struct aos_compilation *cp, const struct tgsi_full_instruction *op )
+{
+   struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]);
+   struct x86_reg tmp0 = aos_get_xmm_reg(cp);
+
+   sse2_cvttps2dq(cp->func, tmp0, arg0);
+   sse2_cvtdq2ps(cp->func, tmp0, tmp0);
+
+   store_dest(cp, &op->FullDstRegisters[0], tmp0);
+   return TRUE;
+}
 
 static boolean emit_XPD( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) 
 {
@@ -1758,6 +1787,9 @@ emit_instruction( struct aos_compilation *cp,
    case TGSI_OPCODE_SIN:
       return emit_SIN(cp, inst);
 
+   case TGSI_OPCODE_TRUNC:
+      return emit_TRUNC(cp, inst);
+
    case TGSI_OPCODE_END:
       return TRUE;
 
@@ -1939,6 +1971,11 @@ 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 );
+
+   cp.x86_reg[0] = 0;
+   cp.x86_reg[1] = 0;
+   
    /* Note address for loop jump 
     */
    label = x86_get_label(cp.func);
@@ -2018,13 +2055,7 @@ 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
     */
@@ -2065,15 +2096,13 @@ 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;
    }
+
+   if (0) debug_printf("%s %d/%d: %p %d\n", __FUNCTION__, buf, vaos->nr_vb, ptr, stride);
 }
 
 
@@ -2086,10 +2115,12 @@ static void PIPE_CDECL vaos_run_elts( struct draw_vs_varient *varient,
    struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
    struct aos_machine *machine = vaos->draw->vs.aos_machine;
 
+   if (0) debug_printf("%s %d\n", __FUNCTION__, count);
+
    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,
@@ -2105,10 +2136,13 @@ static void PIPE_CDECL vaos_run_linear( struct draw_vs_varient *varient,
    struct draw_vs_varient_aos_sse *vaos = (struct draw_vs_varient_aos_sse *)varient;
    struct aos_machine *machine = vaos->draw->vs.aos_machine;
 
+   if (0) debug_printf("%s %d %d const: %x\n", __FUNCTION__, start, count, 
+                       vaos->base.key.const_vbuffers);
+
    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 +2161,7 @@ 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 +2174,7 @@ 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)
@@ -2147,17 +2182,23 @@ static struct draw_vs_varient *varient_aos_sse( struct draw_vertex_shader *vs,
    
    vaos->base.key = *key;
    vaos->base.vs = vs;
-   vaos->base.set_input = vaos_set_buffer;
+   vaos->base.set_buffer = vaos_set_buffer;
    vaos->base.destroy = vaos_destroy;
    vaos->base.run_linear = vaos_run_linear;
    vaos->base.run_elts = vaos_run_elts;
 
    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;
 
+   if (0)
+      debug_printf("nr_vb: %d const: %x\n", vaos->nr_vb, vaos->base.key.const_vbuffers);
+
 #if 0
    tgsi_dump(vs->state.tokens, 0);
 #endif
@@ -2179,8 +2220,8 @@ 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] );