t_vertex_sse.c

Mesa is an open-source implementation of the OpenGL specification - a system for rendering interacti

/*
 * Copyright 2003 Tungsten Graphics, inc.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * on the rights to use, copy, modify, merge, publish, distribute, sub
 * license, and/or sell copies of the Software, and to permit persons to whom
 * the Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.  IN NO EVENT SHALL
 * TUNGSTEN GRAPHICS AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 * Authors:
 *    Keith Whitwell <[e-mail]>
 */

#include "glheader.h"
#include "context.h"
#include "colormac.h"
#include "t_context.h"
#include "t_vertex.h"
#include "simple_list.h"
#include "enums.h"

#if defined(USE_SSE_ASM)

#include "x86/rtasm/x86sse.h"
#include "x86/common_x86_asm.h"


/**
 * Number of bytes to allocate for generated SSE functions
 */
#define MAX_SSE_CODE_SIZE 1024


#define X    0
#define Y    1
#define Z    2
#define W    3


struct x86_program {
   struct x86_function func;

   GLcontext *ctx;
   GLboolean inputs_safe;
   GLboolean outputs_safe;
   GLboolean have_sse2;
   
   struct x86_reg identity;
   struct x86_reg chan0;
};


static struct x86_reg get_identity( struct x86_program *p )
{
   return p->identity;
}

static void emit_load4f_4( struct x86_program *p, 			   
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   sse_movups(&p->func, dest, arg0);
}

static void emit_load4f_3( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   /* Have to jump through some hoops:
    *
    * c 0 0 0
    * c 0 0 1
    * 0 0 c 1
    * a b c 1
    */
   sse_movss(&p->func, dest, x86_make_disp(arg0, 8));
   sse_shufps(&p->func, dest, get_identity(p), SHUF(X,Y,Z,W) );
   sse_shufps(&p->func, dest, dest, SHUF(Y,Z,X,W) );
   sse_movlps(&p->func, dest, arg0);
}

static void emit_load4f_2( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   /* Initialize from identity, then pull in low two words:
    */
   sse_movups(&p->func, dest, get_identity(p));
   sse_movlps(&p->func, dest, arg0);
}

static void emit_load4f_1( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   /* Pull in low word, then swizzle in identity */
   sse_movss(&p->func, dest, arg0);
   sse_shufps(&p->func, dest, get_identity(p), SHUF(X,Y,Z,W) );
}



static void emit_load3f_3( struct x86_program *p, 			   
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   /* Over-reads by 1 dword - potential SEGV if input is a vertex
    * array.
    */
   if (p->inputs_safe) {
      sse_movups(&p->func, dest, arg0);
   } 
   else {
      /* c 0 0 0
       * c c c c
       * a b c c 
       */
      sse_movss(&p->func, dest, x86_make_disp(arg0, 8));
      sse_shufps(&p->func, dest, dest, SHUF(X,X,X,X));
      sse_movlps(&p->func, dest, arg0);
   }
}

static void emit_load3f_2( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   emit_load4f_2(p, dest, arg0);
}

static void emit_load3f_1( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   emit_load4f_1(p, dest, arg0);
}

static void emit_load2f_2( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   sse_movlps(&p->func, dest, arg0);
}

static void emit_load2f_1( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   emit_load4f_1(p, dest, arg0);
}

static void emit_load1f_1( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   sse_movss(&p->func, dest, arg0);
}

static void (*load[4][4])( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 ) = {
   { emit_load1f_1, 
     emit_load1f_1, 
     emit_load1f_1, 
     emit_load1f_1 },

   { emit_load2f_1, 
     emit_load2f_2, 
     emit_load2f_2, 
     emit_load2f_2 },

   { emit_load3f_1, 
     emit_load3f_2, 
     emit_load3f_3, 
     emit_load3f_3 },

   { emit_load4f_1, 
     emit_load4f_2, 
     emit_load4f_3, 
     emit_load4f_4 } 
};

static void emit_load( struct x86_program *p,
		       struct x86_reg dest,
		       GLuint sz,
		       struct x86_reg src,
		       GLuint src_sz)
{
   load[sz-1][src_sz-1](p, dest, src);
}

static void emit_store4f( struct x86_program *p, 			   
			  struct x86_reg dest,
			  struct x86_reg arg0 )
{
   sse_movups(&p->func, dest, arg0);
}

static void emit_store3f( struct x86_program *p, 
			  struct x86_reg dest,
			  struct x86_reg arg0 )
{
   if (p->outputs_safe) {
      /* Emit the extra dword anyway.  This may hurt writecombining,
       * may cause other problems.
       */
      sse_movups(&p->func, dest, arg0);
   }
   else {
      /* Alternate strategy - emit two, shuffle, emit one.
       */
      sse_movlps(&p->func, dest, arg0);
      sse_shufps(&p->func, arg0, arg0, SHUF(Z,Z,Z,Z) ); /* NOTE! destructive */
      sse_movss(&p->func, x86_make_disp(dest,8), arg0);
   }
}

static void emit_store2f( struct x86_program *p, 
			   struct x86_reg dest,
			   struct x86_reg arg0 )
{
   sse_movlps(&p->func, dest, arg0);
}

static void emit_store1f( struct x86_program *p, 
			  struct x86_reg dest,
			  struct x86_reg arg0 )
{
   sse_movss(&p->func, dest, arg0);
}


static void (*store[4])( struct x86_program *p, 
			 struct x86_reg dest,
			 struct x86_reg arg0 ) = 
{
   emit_store1f, 
   emit_store2f, 
   emit_store3f, 
   emit_store4f 
};

static void emit_store( struct x86_program *p,
			struct x86_reg dest,
			GLuint sz,
			struct x86_reg temp )

{
   store[sz-1](p, dest, temp);
}

static void emit_pack_store_4ub( struct x86_program *p,
				 struct x86_reg dest,
				 struct x86_reg temp )
{
   /* Scale by 255.0
    */
   sse_mulps(&p->func, temp, p->chan0);

   if (p->have_sse2) {
      sse2_cvtps2dq(&p->func, temp, temp);
      sse2_packssdw(&p->func, temp, temp);
      sse2_packuswb(&p->func, temp, temp);
      sse_movss(&p->func, dest, temp);
   }
   else {
      struct x86_reg mmx0 = x86_make_reg(file_MMX, 0);
      struct x86_reg mmx1 = x86_make_reg(file_MMX, 1);
      sse_cvtps2pi(&p->func, mmx0, temp);
      sse_movhlps(&p->func, temp, temp);
      sse_cvtps2pi(&p->func, mmx1, temp);
      mmx_packssdw(&p->func, mmx0, mmx1);
      mmx_packuswb(&p->func, mmx0, mmx0);
      mmx_movd(&p->func, dest, mmx0);
   }
}

static GLint get_offset( const void *a, const void *b )
{
   return (const char *)b - (const char *)a;
}

/* Not much happens here.  Eventually use this function to try and
 * avoid saving/reloading the source pointers each vertex (if some of
 * them can fit in registers).
 */
static void get_src_ptr( struct x86_program *p,
			 struct x86_reg srcREG,
			 struct x86_reg vtxREG,
			 struct tnl_clipspace_attr *a )
{
   struct tnl_clipspace *vtx = GET_VERTEX_STATE(p->ctx);
   struct x86_reg ptr_to_src = x86_make_disp(vtxREG, get_offset(vtx, &a->inputptr));

   /* Load current a[j].inputptr
    */
   x86_mov(&p->func, srcREG, ptr_to_src);
}

static void update_src_ptr( struct x86_program *p,
			 struct x86_reg srcREG,
			 struct x86_reg vtxREG,
			 struct tnl_clipspace_attr *a )
{
   if (a->inputstride) {
      struct tnl_clipspace *vtx = GET_VERTEX_STATE(p->ctx);
      struct x86_reg ptr_to_src = x86_make_disp(vtxREG, get_offset(vtx, &a->inputptr));

      /* add a[j].inputstride (hardcoded value - could just as easily
       * pull the stride value from memory each time).
       */
      x86_lea(&p->func, srcREG, x86_make_disp(srcREG, a->inputstride));
      
      /* save new value of a[j].inputptr 
       */
      x86_mov(&p->func, ptr_to_src, srcREG);
   }
}


/* Lots of hardcoding
 *
 * EAX -- pointer to current output vertex
 * ECX -- pointer to current attribute 
 * 
 */
static GLboolean build_vertex_emit( struct x86_program *p )