void _tnl_generate_sse_emit( struct gl_context *ctx )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
struct x86_program p;
if (!cpu_has_xmm) {
vtx->codegen_emit = NULL;
return;
}
memset(&p, 0, sizeof(p));
p.ctx = ctx;
p.inputs_safe = 0; /* for now */
p.outputs_safe = 0; /* for now */
p.have_sse2 = cpu_has_xmm2;
p.identity = x86_make_reg(file_XMM, 6);
p.chan0 = x86_make_reg(file_XMM, 7);
if (!x86_init_func_size(&p.func, MAX_SSE_CODE_SIZE)) {
vtx->emit = NULL;
return;
}
if (build_vertex_emit(&p)) {
_tnl_register_fastpath( vtx, GL_TRUE );
}
else {
/* Note the failure so that we don't keep trying to codegen an
* impossible state:
*/
_tnl_register_fastpath( vtx, GL_FALSE );
x86_release_func(&p.func);
}
}
GLuint _tnl_install_attrs( GLcontext *ctx, const struct tnl_attr_map *map,
GLuint nr, const GLfloat *vp,
GLuint unpacked_size )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
GLuint offset = 0;
GLuint i, j;
assert(nr < _TNL_ATTRIB_MAX);
assert(nr == 0 || map[0].attrib == VERT_ATTRIB_POS);
vtx->emit = 0;
vtx->interp = choose_interp_func;
vtx->copy_pv = choose_copy_pv_func;
vtx->new_inputs = ~0;
for (j = 0, i = 0; i < nr; i++) {
const GLuint format = map[i].format;
if (format == EMIT_PAD) {
/*
fprintf(stderr, "%d: pad %d, offset %d\n", i,
map[i].offset, offset);
*/
offset += map[i].offset;
}
else {
vtx->attr[j].attrib = map[i].attrib;
vtx->attr[j].format = format;
vtx->attr[j].vp = vp;
vtx->attr[j].insert = format_info[format].insert;
vtx->attr[j].extract = format_info[format].extract;
vtx->attr[j].vertattrsize = format_info[format].attrsize;
if (unpacked_size)
vtx->attr[j].vertoffset = map[i].offset;
else
vtx->attr[j].vertoffset = offset;
/*
fprintf(stderr, "%d: %s, vp %p, offset %d\n", i,
format_info[format].name, (void *)vp,
vtx->attr[j].vertoffset);
*/
offset += format_info[format].attrsize;
j++;
}
}
vtx->attr_count = j;
if (unpacked_size)
vtx->vertex_size = unpacked_size;
else
vtx->vertex_size = offset;
assert(vtx->vertex_size <= vtx->max_vertex_size);
return vtx->vertex_size;
}
static void generic_emit( GLcontext *ctx,
GLuint start, GLuint end,
void *dest )
{
struct vertex_buffer *VB = &TNL_CONTEXT(ctx)->vb;
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
struct tnl_clipspace_attr *a = vtx->attr;
GLubyte *v = (GLubyte *)dest;
GLuint i, j;
const GLuint count = vtx->attr_count;
GLuint stride;
for (j = 0; j < count; j++) {
GLvector4f *vptr = VB->AttribPtr[a[j].attrib];
a[j].inputstride = vptr->stride;
a[j].inputptr = ((GLubyte *)vptr->data) + start * vptr->stride;
a[j].emit = a[j].insert[vptr->size - 1];
}
end -= start;
stride = vtx->vertex_size;
for (i = 0 ; i < end ; i++, v += stride) {
for (j = 0; j < count; j++) {
GLfloat *in = (GLfloat *)a[j].inputptr;
a[j].inputptr += a[j].inputstride;
a[j].emit( &a[j], v + a[j].vertoffset, in );
}
}
}
/* Extract a named attribute from a hardware vertex. Will have to
* reverse any viewport transformation, swizzling or other conversions
* which may have been applied:
*/
void _tnl_get_attr( struct gl_context *ctx, const void *vin,
GLenum attr, GLfloat *dest )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
const struct tnl_clipspace_attr *a = vtx->attr;
const GLuint attr_count = vtx->attr_count;
GLuint j;
for (j = 0; j < attr_count; j++) {
if (a[j].attrib == attr) {
a[j].extract( &a[j], dest, (GLubyte *)vin + a[j].vertoffset );
return;
}
}
/* Else return the value from ctx->Current.
*/
if (attr == _TNL_ATTRIB_POINTSIZE) {
/* If the hardware vertex doesn't have point size then use size from
* struct gl_context. XXX this will be wrong if drawing attenuated points!
*/
dest[0] = ctx->Point.Size;
}
else {
memcpy( dest, ctx->Current.Attrib[attr], 4*sizeof(GLfloat));
}
}
static void update_input_ptrs( struct gl_context *ctx, GLuint start )
{
struct vertex_buffer *VB = &TNL_CONTEXT(ctx)->vb;
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
struct tnl_clipspace_attr *a = vtx->attr;
const GLuint count = vtx->attr_count;
GLuint j;
for (j = 0; j < count; j++) {
GLvector4f *vptr = VB->AttribPtr[a[j].attrib];
if (vtx->emit != choose_emit_func) {
assert(a[j].inputstride == vptr->stride);
assert(a[j].inputsize == vptr->size);
}
a[j].inputptr = ((GLubyte *)vptr->data) + start * vptr->stride;
}
if (a->vp) {
vtx->vp_scale[0] = a->vp[MAT_SX];
vtx->vp_scale[1] = a->vp[MAT_SY];
vtx->vp_scale[2] = a->vp[MAT_SZ];
vtx->vp_scale[3] = 1.0;
vtx->vp_xlate[0] = a->vp[MAT_TX];
vtx->vp_xlate[1] = a->vp[MAT_TY];
vtx->vp_xlate[2] = a->vp[MAT_TZ];
vtx->vp_xlate[3] = 0.0;
}
}
void _tnl_free_vertices( struct gl_context *ctx )
{
TNLcontext *tnl = TNL_CONTEXT(ctx);
if (tnl) {
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
struct tnl_clipspace_fastpath *fp, *tmp;
_mesa_align_free(vtx->vertex_buf);
vtx->vertex_buf = NULL;
for (fp = vtx->fastpath ; fp ; fp = tmp) {
tmp = fp->next;
free(fp->attr);
/* KW: At the moment, fp->func is constrained to be allocated by
* _mesa_exec_alloc(), as the hardwired fastpaths in
* t_vertex_generic.c are handled specially. It would be nice
* to unify them, but this probably won't change until this
* module gets another overhaul.
*/
_mesa_exec_free((void *) fp->func);
free(fp);
}
vtx->fastpath = NULL;
}
}
static void do_emit( GLcontext *ctx, GLuint start, GLuint end,
void *dest)
{
TNLcontext *tnl = TNL_CONTEXT(ctx);
struct vertex_buffer *VB = &tnl->vb;
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
struct tnl_clipspace_attr *a = vtx->attr;
const GLuint count = vtx->attr_count;
GLuint j;
for (j = 0; j < count; j++) {
GLvector4f *vptr = VB->AttribPtr[a[j].attrib];
a[j].inputstride = vptr->stride;
a[j].inputptr = ((GLubyte *)vptr->data) + start * vptr->stride;
a[j].emit = a[j].insert[vptr->size - 1];
}
vtx->emit = 0;
if (0)
vtx->emit = _tnl_codegen_emit(ctx);
if (!vtx->emit)
vtx->emit = generic_emit;
vtx->emit( ctx, start, end, dest );
}
void _tnl_free_vertices( GLcontext *ctx )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
if (vtx->vertex_buf) {
ALIGN_FREE(vtx->vertex_buf);
vtx->vertex_buf = 0;
}
}
/* Interpolate between two vertices to produce a third:
*/
void _tnl_interp( struct gl_context *ctx,
GLfloat t,
GLuint edst, GLuint eout, GLuint ein,
GLboolean force_boundary )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
vtx->interp( ctx, t, edst, eout, ein, force_boundary );
}
void *_tnl_emit_vertices_to_buffer( GLcontext *ctx,
GLuint start,
GLuint end,
void *dest )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
do_emit( ctx, start, end, dest );
return (void *)((GLubyte *)dest + vtx->vertex_size * (end - start));
}
void _tnl_invalidate_vertex_state( GLcontext *ctx, GLuint new_state )
{
if (new_state & (_DD_NEW_TRI_LIGHT_TWOSIDE|_DD_NEW_TRI_UNFILLED) ) {
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
vtx->new_inputs = ~0;
vtx->interp = choose_interp_func;
vtx->copy_pv = choose_copy_pv_func;
}
}
void _tnl_invalidate_vertex_state( struct gl_context *ctx, GLuint new_state )
{
/* if two-sided lighting changes or filled/unfilled polygon state changes */
if (new_state & (_NEW_LIGHT | _NEW_POLYGON) ) {
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
vtx->new_inputs = ~0;
vtx->interp = choose_interp_func;
vtx->copy_pv = choose_copy_pv_func;
}
}
void _tnl_build_vertices( struct gl_context *ctx,
GLuint start,
GLuint end,
GLuint newinputs )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
update_input_ptrs( ctx, start );
vtx->emit( ctx, end - start,
(GLubyte *)(vtx->vertex_buf +
start * vtx->vertex_size));
}
/* 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);
}
/* Use the codegen paths to select one of a number of hardwired
* fastpaths.
*/
void _tnl_generate_hardwired_emit( struct gl_context *ctx )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
tnl_emit_func func = NULL;
/* Does it fit a hardwired fastpath? Help! this is growing out of
* control!
*/
switch (vtx->attr_count) {
case 2:
if (vtx->attr[0].emit == insert_3f_viewport_3) {
if (vtx->attr[1].emit == insert_4ub_4f_bgra_4)
func = emit_viewport3_bgra4;
else if (vtx->attr[1].emit == insert_4ub_4f_rgba_4)
func = emit_viewport3_rgba4;
}
else if (vtx->attr[0].emit == insert_3f_3 &&
vtx->attr[1].emit == insert_4ub_4f_rgba_4) {
func = emit_xyz3_rgba4;
}
break;
case 3:
if (vtx->attr[2].emit == insert_2f_2) {
if (vtx->attr[1].emit == insert_4ub_4f_rgba_4) {
if (vtx->attr[0].emit == insert_4f_viewport_4)
func = emit_viewport4_rgba4_st2;
else if (vtx->attr[0].emit == insert_4f_4)
func = emit_xyzw4_rgba4_st2;
}
else if (vtx->attr[1].emit == insert_4ub_4f_bgra_4 &&
vtx->attr[0].emit == insert_4f_viewport_4)
func = emit_viewport4_bgra4_st2;
}
break;
case 4:
if (vtx->attr[2].emit == insert_2f_2 &&
vtx->attr[3].emit == insert_2f_2) {
if (vtx->attr[1].emit == insert_4ub_4f_rgba_4) {
if (vtx->attr[0].emit == insert_4f_viewport_4)
func = emit_viewport4_rgba4_st2_st2;
else if (vtx->attr[0].emit == insert_4f_4)
func = emit_xyzw4_rgba4_st2_st2;
}
else if (vtx->attr[1].emit == insert_4ub_4f_bgra_4 &&
vtx->attr[0].emit == insert_4f_viewport_4)
func = emit_viewport4_bgra4_st2_st2;
}
break;
}
vtx->emit = func;
}
static void choose_copy_pv_func( GLcontext *ctx, GLuint edst, GLuint esrc )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
if (vtx->need_extras &&
(ctx->_TriangleCaps & (DD_TRI_LIGHT_TWOSIDE|DD_TRI_UNFILLED))) {
vtx->copy_pv = generic_copy_pv_extras;
} else {
vtx->copy_pv = generic_copy_pv;
}
vtx->copy_pv( ctx, edst, esrc );
}
/* Emit VB vertices start..end to dest. Note that VB vertex at
* postion start will be emitted to dest at position zero.
*/
void *_tnl_emit_vertices_to_buffer( struct gl_context *ctx,
GLuint start,
GLuint end,
void *dest )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
update_input_ptrs(ctx, start);
/* Note: dest should not be adjusted for non-zero 'start' values:
*/
vtx->emit( ctx, end - start, (GLubyte*) dest );
return (void *)((GLubyte *)dest + vtx->vertex_size * (end - start));
}
static void adjust_input_ptrs( struct gl_context *ctx, GLint diff)
{
struct vertex_buffer *VB = &TNL_CONTEXT(ctx)->vb;
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
struct tnl_clipspace_attr *a = vtx->attr;
const GLuint count = vtx->attr_count;
GLuint j;
diff -= 1;
for (j=0; j<count; ++j) {
register GLvector4f *vptr = VB->AttribPtr[a->attrib];
(a++)->inputptr += diff*vptr->stride;
}
}
void _tnl_generic_interp( struct gl_context *ctx,
GLfloat t,
GLuint edst, GLuint eout, GLuint ein,
GLboolean force_boundary )
{
TNLcontext *tnl = TNL_CONTEXT(ctx);
struct vertex_buffer *VB = &tnl->vb;
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
const GLubyte *vin = vtx->vertex_buf + ein * vtx->vertex_size;
const GLubyte *vout = vtx->vertex_buf + eout * vtx->vertex_size;
GLubyte *vdst = vtx->vertex_buf + edst * vtx->vertex_size;
const struct tnl_clipspace_attr *a = vtx->attr;
const GLuint attr_count = vtx->attr_count;
GLuint j;
(void) force_boundary;
if (tnl->NeedNdcCoords) {
const GLfloat *dstclip = VB->ClipPtr->data[edst];
if (dstclip[3] != 0.0) {
const GLfloat w = 1.0f / dstclip[3];
GLfloat pos[4];
pos[0] = dstclip[0] * w;
pos[1] = dstclip[1] * w;
pos[2] = dstclip[2] * w;
pos[3] = w;
a[0].insert[4-1]( &a[0], vdst, pos );
}
}
else {
a[0].insert[4-1]( &a[0], vdst, VB->ClipPtr->data[edst] );
}
for (j = 1; j < attr_count; j++) {
GLfloat fin[4], fout[4], fdst[4];
a[j].extract( &a[j], fin, vin + a[j].vertoffset );
a[j].extract( &a[j], fout, vout + a[j].vertoffset );
INTERP_F( t, fdst[3], fout[3], fin[3] );
INTERP_F( t, fdst[2], fout[2], fin[2] );
INTERP_F( t, fdst[1], fout[1], fin[1] );
INTERP_F( t, fdst[0], fout[0], fin[0] );
a[j].insert[4-1]( &a[j], vdst + a[j].vertoffset, fdst );
}
}
void _tnl_init_vertices( struct gl_context *ctx,
GLuint vb_size,
GLuint max_vertex_size )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
_tnl_install_attrs( ctx, NULL, 0, NULL, 0 );
vtx->need_extras = GL_TRUE;
if (max_vertex_size > vtx->max_vertex_size) {
_tnl_free_vertices( ctx );
vtx->max_vertex_size = max_vertex_size;
vtx->vertex_buf = _mesa_align_calloc(vb_size * max_vertex_size, 32 );
invalidate_funcs(vtx);
}
switch(CHAN_TYPE) {
case GL_UNSIGNED_BYTE:
vtx->chan_scale[0] = 255.0;
vtx->chan_scale[1] = 255.0;
vtx->chan_scale[2] = 255.0;
vtx->chan_scale[3] = 255.0;
break;
case GL_UNSIGNED_SHORT:
vtx->chan_scale[0] = 65535.0;
vtx->chan_scale[1] = 65535.0;
vtx->chan_scale[2] = 65535.0;
vtx->chan_scale[3] = 65535.0;
break;
default:
vtx->chan_scale[0] = 1.0;
vtx->chan_scale[1] = 1.0;
vtx->chan_scale[2] = 1.0;
vtx->chan_scale[3] = 1.0;
break;
}
vtx->identity[0] = 0.0;
vtx->identity[1] = 0.0;
vtx->identity[2] = 0.0;
vtx->identity[3] = 1.0;
vtx->codegen_emit = NULL;
#ifdef USE_SSE_ASM
if (!getenv("MESA_NO_CODEGEN"))
vtx->codegen_emit = _tnl_generate_sse_emit;
#endif
}
/* Complementary operation to the above.
*/
void _tnl_set_attr( struct gl_context *ctx, void *vout,
GLenum attr, const GLfloat *src )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
const struct tnl_clipspace_attr *a = vtx->attr;
const GLuint attr_count = vtx->attr_count;
GLuint j;
for (j = 0; j < attr_count; j++) {
if (a[j].attrib == attr) {
a[j].insert[4-1]( &a[j], (GLubyte *)vout + a[j].vertoffset, src );
return;
}
}
}
void _tnl_build_vertices( GLcontext *ctx,
GLuint start,
GLuint end,
GLuint newinputs )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
const GLuint stride = vtx->vertex_size;
GLubyte *vDest = ((GLubyte *)vtx->vertex_buf + (start*stride));
newinputs |= vtx->new_inputs;
vtx->new_inputs = 0;
if (newinputs)
do_emit( ctx, start, end, vDest );
}
static void choose_interp_func( GLcontext *ctx,
GLfloat t,
GLuint edst, GLuint eout, GLuint ein,
GLboolean force_boundary )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
if (vtx->need_extras &&
(ctx->_TriangleCaps & (DD_TRI_LIGHT_TWOSIDE|DD_TRI_UNFILLED))) {
vtx->interp = generic_interp_extras;
} else {
vtx->interp = generic_interp;
}
vtx->interp( ctx, t, edst, eout, ein, force_boundary );
}
static void choose_copy_pv_func( struct gl_context *ctx, GLuint edst, GLuint esrc )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
GLboolean unfilled = (ctx->Polygon.FrontMode != GL_FILL ||
ctx->Polygon.BackMode != GL_FILL);
GLboolean twosided = ctx->Light.Enabled && ctx->Light.Model.TwoSide;
if (vtx->need_extras && (twosided || unfilled)) {
vtx->copy_pv = _tnl_generic_copy_pv_extras;
} else {
vtx->copy_pv = _tnl_generic_copy_pv;
}
vtx->copy_pv( ctx, edst, esrc );
}
void _tnl_init_vertices( GLcontext *ctx,
GLuint vb_size,
GLuint max_vertex_size )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
_tnl_install_attrs( ctx, 0, 0, 0, 0 );
vtx->need_extras = GL_TRUE;
if (max_vertex_size > vtx->max_vertex_size) {
_tnl_free_vertices( ctx );
vtx->max_vertex_size = max_vertex_size;
vtx->vertex_buf = (GLubyte *)ALIGN_CALLOC(vb_size * max_vertex_size, 32 );
}
_tnl_init_c_codegen( &vtx->codegen );
}
void _tnl_generic_emit( struct gl_context *ctx,
GLuint count,
GLubyte *v )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
struct tnl_clipspace_attr *a = vtx->attr;
const GLuint attr_count = vtx->attr_count;
const GLuint stride = vtx->vertex_size;
GLuint i, j;
for (i = 0 ; i < count ; i++, v += stride) {
for (j = 0; j < attr_count; j++) {
GLfloat *in = (GLfloat *)a[j].inputptr;
a[j].inputptr += a[j].inputstride;
a[j].emit( &a[j], v + a[j].vertoffset, in );
}
}
}
/* Extract color attributes from one vertex and insert them into
* another. (Shortcircuit extract/insert with memcpy).
*/
void _tnl_generic_copy_pv( struct gl_context *ctx, GLuint edst, GLuint esrc )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
GLubyte *vsrc = vtx->vertex_buf + esrc * vtx->vertex_size;
GLubyte *vdst = vtx->vertex_buf + edst * vtx->vertex_size;
const struct tnl_clipspace_attr *a = vtx->attr;
const GLuint attr_count = vtx->attr_count;
GLuint j;
for (j = 0; j < attr_count; j++) {
if (a[j].attrib == VERT_ATTRIB_COLOR) {
memcpy( vdst + a[j].vertoffset,
vsrc + a[j].vertoffset,
a[j].vertattrsize );
}
}
}
static void choose_interp_func( struct gl_context *ctx,
GLfloat t,
GLuint edst, GLuint eout, GLuint ein,
GLboolean force_boundary )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
GLboolean unfilled = (ctx->Polygon.FrontMode != GL_FILL ||
ctx->Polygon.BackMode != GL_FILL);
GLboolean twosided = ctx->Light.Enabled && ctx->Light.Model.TwoSide;
if (vtx->need_extras && (twosided || unfilled)) {
vtx->interp = _tnl_generic_interp_extras;
} else {
vtx->interp = _tnl_generic_interp;
}
vtx->interp( ctx, t, edst, eout, ein, force_boundary );
}
/* Extract a named attribute from a hardware vertex. Will have to
* reverse any viewport transformation, swizzling or other conversions
* which may have been applied:
*/
void _tnl_get_attr( GLcontext *ctx, const void *vin,
GLenum attr, GLfloat *dest )
{
struct tnl_clipspace *vtx = GET_VERTEX_STATE(ctx);
const struct tnl_clipspace_attr *a = vtx->attr;
const GLuint attr_count = vtx->attr_count;
GLuint j;
for (j = 0; j < attr_count; j++) {
if (a[j].attrib == attr) {
a[j].extract( &a[j], dest, (GLubyte *)vin + a[j].vertoffset );
return;
}
}
/* Else return the value from ctx->Current -- dangerous???
*/
_mesa_memcpy( dest, ctx->Current.Attrib[attr], 4*sizeof(GLfloat));
}
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);
}
}