static void precalc_lit( struct brw_wm_compile *c,
const struct prog_instruction *inst )
{
struct prog_src_register src0 = inst->SrcReg[0];
struct prog_dst_register dst = inst->DstReg;
if (dst.WriteMask & WRITEMASK_XW) {
struct prog_instruction *swz;
/* dst.xw = swz src0.1111
*/
swz = emit_op(c,
OPCODE_SWZ,
dst_mask(dst, WRITEMASK_XW),
0,
src_swizzle1(src0, SWIZZLE_ONE),
src_undef(),
src_undef());
/* Avoid letting the negation flag of src0 affect our 1 constant. */
swz->SrcReg[0].Negate = NEGATE_NONE;
}
if (dst.WriteMask & WRITEMASK_YZ) {
emit_op(c,
OPCODE_LIT,
dst_mask(dst, WRITEMASK_YZ),
inst->SaturateMode,
src0,
src_undef(),
src_undef());
}
}
static void precalc_lit( struct brw_wm_compile *c,
const struct prog_instruction *inst )
{
struct prog_src_register src0 = inst->SrcReg[0];
struct prog_dst_register dst = inst->DstReg;
if (dst.WriteMask & WRITEMASK_XW) {
/* dst.xw = swz src0.1111
*/
emit_op(c,
OPCODE_SWZ,
dst_mask(dst, WRITEMASK_XW),
0, 0, 0,
src_swizzle1(src0, SWIZZLE_ONE),
src_undef(),
src_undef());
}
if (dst.WriteMask & WRITEMASK_YZ) {
emit_op(c,
OPCODE_LIT,
dst_mask(dst, WRITEMASK_YZ),
inst->SaturateMode, 0, 0,
src0,
src_undef(),
src_undef());
}
}
/* Many Mesa opcodes produce the same value across all the result channels.
* We'd rather not have to support that splatting in the opcode implementations,
* and brw_wm_pass*.c wants to optimize them out by shuffling references around
* anyway. We can easily get both by emitting the opcode to one channel, and
* then MOVing it to the others, which brw_wm_pass*.c already understands.
*/
static struct prog_instruction *emit_scalar_insn(struct brw_wm_compile *c,
const struct prog_instruction *inst0)
{
struct prog_instruction *inst;
unsigned int dst_chan;
unsigned int other_channel_mask;
if (inst0->DstReg.WriteMask == 0)
return NULL;
dst_chan = _mesa_ffs(inst0->DstReg.WriteMask) - 1;
inst = get_fp_inst(c);
*inst = *inst0;
inst->DstReg.WriteMask = 1 << dst_chan;
other_channel_mask = inst0->DstReg.WriteMask & ~(1 << dst_chan);
if (other_channel_mask != 0) {
inst = emit_op(c,
OPCODE_MOV,
dst_mask(inst0->DstReg, other_channel_mask),
0,
src_swizzle1(src_reg_from_dst(inst0->DstReg), dst_chan),
src_undef(),
src_undef());
}
return inst;
}
static struct prog_src_register get_pixel_xy( struct brw_wm_compile *c )
{
if (src_is_undef(c->pixel_xy)) {
struct prog_dst_register pixel_xy = get_temp(c);
struct prog_src_register payload_r0_depth = src_reg(PROGRAM_PAYLOAD, PAYLOAD_DEPTH);
/* Emit the out calculations, and hold onto the results. Use
* two instructions as a temporary is required.
*/
/* pixel_xy.xy = PIXELXY payload[0];
*/
emit_op(c,
WM_PIXELXY,
dst_mask(pixel_xy, WRITEMASK_XY),
0,
payload_r0_depth,
src_undef(),
src_undef());
c->pixel_xy = src_reg_from_dst(pixel_xy);
}
return c->pixel_xy;
}
static INLINE void emit_op1(struct brw_wm_compile *c,
GLuint op,
struct brw_fp_dst dest,
struct brw_fp_src src0)
{
emit_tex_op(c, op, dest, 0, 0, 0, src0, src_undef(), src_undef());
}
static struct brw_fp_src get_pixel_w( struct brw_wm_compile *c )
{
if (src_is_undef(c->fp_pixel_w)) {
struct brw_fp_dst pixel_w = get_temp(c);
struct brw_fp_src deltas = get_delta_xy(c);
/* XXX: assuming position is always first -- valid?
*/
struct brw_fp_src interp_wpos = src_reg(BRW_FILE_PAYLOAD, 0);
/* deltas.xyw = DELTAS2 deltas.xy, payload.interp_wpos.x
*/
emit_op3(c,
WM_PIXELW,
dst_mask(pixel_w, BRW_WRITEMASK_W),
interp_wpos,
deltas,
src_undef());
c->fp_pixel_w = src_reg_from_dst(pixel_w);
}
return c->fp_pixel_w;
}
/***********************************************************************
* Expand various instructions here to simpler forms.
*/
static void precalc_dst( struct brw_wm_compile *c,
const struct prog_instruction *inst )
{
struct prog_src_register src0 = inst->SrcReg[0];
struct prog_src_register src1 = inst->SrcReg[1];
struct prog_dst_register dst = inst->DstReg;
if (dst.WriteMask & WRITEMASK_Y) {
/* dst.y = mul src0.y, src1.y
*/
emit_op(c,
OPCODE_MUL,
dst_mask(dst, WRITEMASK_Y),
inst->SaturateMode,
src0,
src1,
src_undef());
}
if (dst.WriteMask & WRITEMASK_XZ) {
struct prog_instruction *swz;
GLuint z = GET_SWZ(src0.Swizzle, Z);
/* dst.xz = swz src0.1zzz
*/
swz = emit_op(c,
OPCODE_SWZ,
dst_mask(dst, WRITEMASK_XZ),
inst->SaturateMode,
src_swizzle(src0, SWIZZLE_ONE, z, z, z),
src_undef(),
src_undef());
/* Avoid letting negation flag of src0 affect our 1 constant. */
swz->SrcReg[0].Negate &= ~NEGATE_X;
}
if (dst.WriteMask & WRITEMASK_W) {
/* dst.w = mov src1.w
*/
emit_op(c,
OPCODE_MOV,
dst_mask(dst, WRITEMASK_W),
inst->SaturateMode,
src1,
src_undef(),
src_undef());
}
}
/**
* Emit code for TXP.
*/
static void precalc_txp( struct brw_wm_compile *c,
const struct prog_instruction *inst )
{
struct prog_src_register src0 = inst->SrcReg[0];
if (projtex(c, inst)) {
struct prog_dst_register tmp = get_temp(c);
struct prog_instruction tmp_inst;
/* tmp0.w = RCP inst.arg[0][3]
*/
emit_op(c,
OPCODE_RCP,
dst_mask(tmp, WRITEMASK_W),
0,
src_swizzle1(src0, GET_SWZ(src0.Swizzle, W)),
src_undef(),
src_undef());
/* tmp0.xyz = MUL inst.arg[0], tmp0.wwww
*/
emit_op(c,
OPCODE_MUL,
dst_mask(tmp, WRITEMASK_XYZ),
0,
src0,
src_swizzle1(src_reg_from_dst(tmp), W),
src_undef());
/* dst = precalc(TEX tmp0)
*/
tmp_inst = *inst;
tmp_inst.SrcReg[0] = src_reg_from_dst(tmp);
precalc_tex(c, &tmp_inst);
release_temp(c, tmp);
}
else
{
/* dst = precalc(TEX src0)
*/
precalc_tex(c, inst);
}
}
/* Internally generated immediates: overkill...
*/
static struct brw_fp_src src_imm( struct brw_wm_compile *c,
const GLfloat *v,
unsigned nr)
{
unsigned i, j;
unsigned swizzle;
/* Could do a first pass where we examine all existing immediates
* without expanding.
*/
for (i = 0; i < c->nr_immediates; i++) {
if (match_or_expand_immediate( v,
nr,
c->immediate[i].v,
&c->immediate[i].nr,
&swizzle ))
goto out;
}
if (c->nr_immediates < Elements(c->immediate)) {
i = c->nr_immediates++;
if (match_or_expand_immediate( v,
nr,
c->immediate[i].v,
&c->immediate[i].nr,
&swizzle ))
goto out;
}
c->error = 1;
return src_undef();
out:
/* Make sure that all referenced elements are from this immediate.
* Has the effect of making size-one immediates into scalars.
*/
for (j = nr; j < 4; j++)
swizzle |= (swizzle & 0x3) << (j * 2);
return src_swizzle( src_reg( TGSI_FILE_IMMEDIATE, i ),
BRW_GET_SWZ(swizzle, X),
BRW_GET_SWZ(swizzle, Y),
BRW_GET_SWZ(swizzle, Z),
BRW_GET_SWZ(swizzle, W) );
}
static struct brw_fp_src get_delta_xy( struct brw_wm_compile *c )
{
if (src_is_undef(c->fp_delta_xy)) {
struct brw_fp_dst delta_xy = get_temp(c);
struct brw_fp_src pixel_xy = get_pixel_xy(c);
struct brw_fp_src payload_r0_depth = src_reg(BRW_FILE_PAYLOAD, PAYLOAD_DEPTH);
/* deltas.xy = DELTAXY pixel_xy, payload[0]
*/
emit_op3(c,
WM_DELTAXY,
dst_mask(delta_xy, BRW_WRITEMASK_XY),
pixel_xy,
payload_r0_depth,
src_undef());
c->fp_delta_xy = src_reg_from_dst(delta_xy);
}
return c->fp_delta_xy;
}
static struct prog_src_register get_delta_xy( struct brw_wm_compile *c )
{
if (src_is_undef(c->delta_xy)) {
struct prog_dst_register delta_xy = get_temp(c);
struct prog_src_register pixel_xy = get_pixel_xy(c);
struct prog_src_register payload_r0_depth = src_reg(PROGRAM_PAYLOAD, PAYLOAD_DEPTH);
/* deltas.xy = DELTAXY pixel_xy, payload[0]
*/
emit_op(c,
WM_DELTAXY,
dst_mask(delta_xy, WRITEMASK_XY),
0,
pixel_xy,
payload_r0_depth,
src_undef());
c->delta_xy = src_reg_from_dst(delta_xy);
}
return c->delta_xy;
}
static struct prog_src_register get_pixel_w( struct brw_wm_compile *c )
{
if (src_is_undef(c->pixel_w)) {
struct prog_dst_register pixel_w = get_temp(c);
struct prog_src_register deltas = get_delta_xy(c);
struct prog_src_register interp_wpos = src_reg(PROGRAM_PAYLOAD, FRAG_ATTRIB_WPOS);
/* deltas.xyw = DELTAS2 deltas.xy, payload.interp_wpos.x
*/
emit_op(c,
WM_PIXELW,
dst_mask(pixel_w, WRITEMASK_W),
0,
interp_wpos,
deltas,
src_undef());
c->pixel_w = src_reg_from_dst(pixel_w);
}
return c->pixel_w;
}
static void emit_interp( struct brw_wm_compile *c,
GLuint idx )
{
struct prog_dst_register dst = dst_reg(PROGRAM_INPUT, idx);
struct prog_src_register interp = src_reg(PROGRAM_PAYLOAD, idx);
struct prog_src_register deltas = get_delta_xy(c);
/* Need to use PINTERP on attributes which have been
* multiplied by 1/W in the SF program, and LINTERP on those
* which have not:
*/
switch (idx) {
case FRAG_ATTRIB_WPOS:
/* Have to treat wpos.xy specially:
*/
emit_op(c,
WM_WPOSXY,
dst_mask(dst, WRITEMASK_XY),
0,
get_pixel_xy(c),
src_undef(),
src_undef());
dst = dst_mask(dst, WRITEMASK_ZW);
/* PROGRAM_INPUT.attr.xyzw = INTERP payload.interp[attr].x, deltas.xyw
*/
emit_op(c,
WM_LINTERP,
dst,
0,
interp,
deltas,
src_undef());
break;
case FRAG_ATTRIB_COL0:
case FRAG_ATTRIB_COL1:
if (c->key.flat_shade) {
emit_op(c,
WM_CINTERP,
dst,
0,
interp,
src_undef(),
src_undef());
}
else {
if (c->key.linear_color) {
emit_op(c,
WM_LINTERP,
dst,
0,
interp,
deltas,
src_undef());
}
else {
/* perspective-corrected color interpolation */
emit_op(c,
WM_PINTERP,
dst,
0,
interp,
deltas,
get_pixel_w(c));
}
}
break;
case FRAG_ATTRIB_FOGC:
/* Interpolate the fog coordinate */
emit_op(c,
WM_PINTERP,
dst_mask(dst, WRITEMASK_X),
0,
interp,
deltas,
get_pixel_w(c));
emit_op(c,
OPCODE_MOV,
dst_mask(dst, WRITEMASK_YZW),
0,
src_swizzle(interp,
SWIZZLE_ZERO,
SWIZZLE_ZERO,
SWIZZLE_ZERO,
SWIZZLE_ONE),
src_undef(),
src_undef());
break;
case FRAG_ATTRIB_FACE:
emit_op(c,
WM_FRONTFACING,
dst_mask(dst, WRITEMASK_X),
0,
src_undef(),
src_undef(),
src_undef());
break;
case FRAG_ATTRIB_PNTC:
/* XXX review/test this case */
emit_op(c,
WM_PINTERP,
dst_mask(dst, WRITEMASK_XY),
0,
interp,
deltas,
get_pixel_w(c));
emit_op(c,
OPCODE_MOV,
dst_mask(dst, WRITEMASK_ZW),
0,
src_swizzle(interp,
SWIZZLE_ZERO,
SWIZZLE_ZERO,
SWIZZLE_ZERO,
SWIZZLE_ONE),
src_undef(),
src_undef());
break;
default:
emit_op(c,
WM_PINTERP,
dst,
0,
interp,
deltas,
get_pixel_w(c));
break;
}
c->fp_interp_emitted |= 1<<idx;
}
static void precalc_tex( struct brw_wm_compile *c,
const struct prog_instruction *inst )
{
struct prog_src_register coord;
struct prog_dst_register tmpcoord;
if (inst->TexSrcTarget == TEXTURE_RECT_INDEX) {
struct prog_src_register scale =
search_or_add_param5( c,
STATE_INTERNAL,
STATE_TEXRECT_SCALE,
inst->TexSrcUnit,
0,0 );
tmpcoord = get_temp(c);
/* coord.xy = MUL inst->SrcReg[0], { 1/width, 1/height }
*/
emit_op(c,
OPCODE_MUL,
tmpcoord,
0, 0, 0,
inst->SrcReg[0],
scale,
src_undef());
coord = src_reg_from_dst(tmpcoord);
}
else {
coord = inst->SrcReg[0];
}
/* Need to emit YUV texture conversions by hand. Probably need to
* do this here - the alternative is in brw_wm_emit.c, but the
* conversion requires allocating a temporary variable which we
* don't have the facility to do that late in the compilation.
*/
if (!(c->key.yuvtex_mask & (1<<inst->TexSrcUnit))) {
emit_op(c,
OPCODE_TEX,
inst->DstReg,
inst->SaturateMode,
inst->TexSrcUnit,
inst->TexSrcTarget,
coord,
src_undef(),
src_undef());
}
else {
/*
CONST C0 = { -.5, -.0625, -.5, 1.164 }
CONST C1 = { 1.596, -0.813, 2.018, -.391 }
UYV = TEX ...
UYV.xyz = ADD UYV, C0
UYV.y = MUL UYV.y, C0.w
RGB.xyz = MAD UYV.xxz, C1, UYV.y
RGB.y = MAD UYV.z, C1.w, RGB.y
*/
struct prog_dst_register dst = inst->DstReg;
struct prog_src_register src0 = inst->SrcReg[0];
struct prog_dst_register tmp = get_temp(c);
struct prog_src_register tmpsrc = src_reg_from_dst(tmp);
struct prog_src_register C0 = search_or_add_const4f( c, -.5, -.0625, -.5, 1.164 );
struct prog_src_register C1 = search_or_add_const4f( c, 1.596, -0.813, 2.018, -.391 );
/* tmp = TEX ...
*/
emit_op(c,
OPCODE_TEX,
tmp,
inst->SaturateMode,
inst->TexSrcUnit,
inst->TexSrcTarget,
src0,
src_undef(),
src_undef());
/* tmp.xyz = ADD TMP, C0
*/
emit_op(c,
OPCODE_ADD,
dst_mask(tmp, WRITEMASK_XYZ),
0, 0, 0,
tmpsrc,
C0,
src_undef());
/* YUV.y = MUL YUV.y, C0.w
*/
emit_op(c,
OPCODE_MUL,
dst_mask(tmp, WRITEMASK_Y),
0, 0, 0,
tmpsrc,
src_swizzle1(C0, W),
src_undef());
/* RGB.xyz = MAD YUV.xxz, C1, YUV.y
*/
emit_op(c,
OPCODE_MAD,
dst_mask(dst, WRITEMASK_XYZ),
0, 0, 0,
src_swizzle(tmpsrc, X,X,Z,Z),
C1,
src_swizzle1(tmpsrc, Y));
/* RGB.y = MAD YUV.z, C1.w, RGB.y
*/
emit_op(c,
OPCODE_MAD,
dst_mask(dst, WRITEMASK_Y),
0, 0, 0,
src_swizzle1(tmpsrc, Z),
src_swizzle1(C1, W),
src_swizzle1(src_reg_from_dst(dst), Y));
release_temp(c, tmp);
}
if (inst->TexSrcTarget == GL_TEXTURE_RECTANGLE_NV)
release_temp(c, tmpcoord);
}
void brw_wm_pass_fp( struct brw_wm_compile *c )
{
struct brw_fragment_program *fp = c->fp;
GLuint insn;
if (INTEL_DEBUG & DEBUG_WM) {
_mesa_printf("\n\n\npre-fp:\n");
_mesa_print_program(&fp->program.Base);
_mesa_printf("\n");
}
c->pixel_xy = src_undef();
c->delta_xy = src_undef();
c->pixel_w = src_undef();
c->nr_fp_insns = 0;
/* Emit preamble instructions:
*/
for (insn = 0; insn < fp->program.Base.NumInstructions; insn++) {
const struct prog_instruction *inst = &fp->program.Base.Instructions[insn];
struct prog_instruction *out;
/* Check for INPUT values, emit INTERP instructions where
* necessary:
*/
validate_src_regs(c, inst);
switch (inst->Opcode) {
case OPCODE_SWZ:
out = emit_insn(c, inst);
out->Opcode = OPCODE_MOV;
break;
case OPCODE_ABS:
out = emit_insn(c, inst);
out->Opcode = OPCODE_MOV;
out->SrcReg[0].NegateBase = 0;
out->SrcReg[0].Abs = 1;
break;
case OPCODE_SUB:
out = emit_insn(c, inst);
out->Opcode = OPCODE_ADD;
out->SrcReg[1].NegateBase ^= 0xf;
break;
case OPCODE_SCS:
out = emit_insn(c, inst);
/* This should probably be done in the parser.
*/
out->DstReg.WriteMask &= WRITEMASK_XY;
break;
case OPCODE_DST:
precalc_dst(c, inst);
break;
case OPCODE_LIT:
precalc_lit(c, inst);
break;
case OPCODE_TXP:
precalc_txp(c, inst);
break;
case OPCODE_XPD:
out = emit_insn(c, inst);
/* This should probably be done in the parser.
*/
out->DstReg.WriteMask &= WRITEMASK_XYZ;
break;
case OPCODE_KIL:
out = emit_insn(c, inst);
/* This should probably be done in the parser.
*/
out->DstReg.WriteMask = 0;
break;
case OPCODE_END:
case OPCODE_PRINT:
break;
default:
emit_insn(c, inst);
break;
}
}
emit_fog(c);
emit_fb_write(c);
if (INTEL_DEBUG & DEBUG_WM) {
_mesa_printf("\n\n\npass_fp:\n");
print_insns( c->prog_instructions, c->nr_fp_insns );
_mesa_printf("\n");
}
}
/**
* Initial pass for fragment program code generation.
* This function is used by both the GLSL and non-GLSL paths.
*/
int brw_wm_pass_fp( struct brw_wm_compile *c )
{
struct brw_fragment_shader *fs = c->fp;
struct tgsi_parse_context parse;
struct tgsi_full_instruction *inst;
struct tgsi_full_declaration *decl;
const float *imm;
GLuint size;
GLuint i;
if (BRW_DEBUG & DEBUG_WM) {
debug_printf("pre-fp:\n");
tgsi_dump(fs->tokens, 0);
}
c->fp_pixel_xy = src_undef();
c->fp_delta_xy = src_undef();
c->fp_pixel_w = src_undef();
c->nr_fp_insns = 0;
c->nr_immediates = 0;
/* Loop over all instructions doing assorted simplifications and
* transformations.
*/
tgsi_parse_init( &parse, fs->tokens );
while( !tgsi_parse_end_of_tokens( &parse ) ) {
tgsi_parse_token( &parse );
switch( parse.FullToken.Token.Type ) {
case TGSI_TOKEN_TYPE_DECLARATION:
/* Turn intput declarations into special WM_* instructions.
*
* XXX: For non-branching shaders, consider deferring variable
* initialization as late as possible to minimize register
* usage. This is how the original BRW driver worked.
*
* In a branching shader, must preamble instructions at decl
* time, as instruction order in the shader does not
* correspond to the order instructions are executed in the
* wild.
*
* This is where special instructions such as WM_CINTERP,
* WM_LINTERP, WM_PINTERP and WM_WPOSXY are emitted to
* compute shader inputs from the payload registers and pixel
* position.
*/
decl = &parse.FullToken.FullDeclaration;
if( decl->Declaration.File == TGSI_FILE_INPUT ) {
unsigned first, last, mask;
unsigned attrib;
first = decl->Range.First;
last = decl->Range.Last;
mask = decl->Declaration.UsageMask;
for (attrib = first; attrib <= last; attrib++) {
emit_interp(c,
attrib,
decl->Semantic.Name,
decl->Declaration.Interpolate );
}
}
break;
case TGSI_TOKEN_TYPE_IMMEDIATE:
/* Unlike VS programs we can probably manage fine encoding
* immediate values directly into the emitted EU
* instructions, as we probably only need to reference one
* float value per instruction. Just save the data for now
* and use directly later.
*/
i = c->nr_immediates++;
imm = &parse.FullToken.FullImmediate.u[i].Float;
size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
if (c->nr_immediates >= BRW_WM_MAX_CONST)
return PIPE_ERROR_OUT_OF_MEMORY;
for (i = 0; i < size; i++)
c->immediate[c->nr_immediates].v[i] = imm[i];
for (; i < 4; i++)
c->immediate[c->nr_immediates].v[i] = 0.0;
c->immediate[c->nr_immediates].nr = size;
c->nr_immediates++;
break;
case TGSI_TOKEN_TYPE_INSTRUCTION:
inst = &parse.FullToken.FullInstruction;
emit_insn(c, inst);
break;
}
}
if (BRW_DEBUG & DEBUG_WM) {
brw_wm_print_fp_program( c, "pass_fp" );
debug_printf("\n");
}
return c->error;
}
/**
* Initial pass for fragment program code generation.
* This function is used by both the GLSL and non-GLSL paths.
*/
void brw_wm_pass_fp( struct brw_wm_compile *c )
{
struct brw_fragment_program *fp = c->fp;
GLuint insn;
if (INTEL_DEBUG & DEBUG_WM) {
printf("pre-fp:\n");
_mesa_print_program(&fp->program.Base);
printf("\n");
}
c->pixel_xy = src_undef();
c->delta_xy = src_undef();
c->pixel_w = src_undef();
c->nr_fp_insns = 0;
c->fp->tex_units_used = 0x0;
/* Emit preamble instructions. This is where special instructions such as
* WM_CINTERP, WM_LINTERP, WM_PINTERP and WM_WPOSXY are emitted to
* compute shader inputs from varying vars.
*/
for (insn = 0; insn < fp->program.Base.NumInstructions; insn++) {
const struct prog_instruction *inst = &fp->program.Base.Instructions[insn];
validate_src_regs(c, inst);
validate_dst_regs(c, inst);
}
/* Loop over all instructions doing assorted simplifications and
* transformations.
*/
for (insn = 0; insn < fp->program.Base.NumInstructions; insn++) {
const struct prog_instruction *inst = &fp->program.Base.Instructions[insn];
struct prog_instruction *out;
/* Check for INPUT values, emit INTERP instructions where
* necessary:
*/
switch (inst->Opcode) {
case OPCODE_SWZ:
out = emit_insn(c, inst);
out->Opcode = OPCODE_MOV;
break;
case OPCODE_ABS:
out = emit_insn(c, inst);
out->Opcode = OPCODE_MOV;
out->SrcReg[0].Negate = NEGATE_NONE;
out->SrcReg[0].Abs = 1;
break;
case OPCODE_SUB:
out = emit_insn(c, inst);
out->Opcode = OPCODE_ADD;
out->SrcReg[1].Negate ^= NEGATE_XYZW;
break;
case OPCODE_SCS:
out = emit_insn(c, inst);
/* This should probably be done in the parser.
*/
out->DstReg.WriteMask &= WRITEMASK_XY;
break;
case OPCODE_DST:
precalc_dst(c, inst);
break;
case OPCODE_LIT:
precalc_lit(c, inst);
break;
case OPCODE_TEX:
precalc_tex(c, inst);
break;
case OPCODE_TXP:
precalc_txp(c, inst);
break;
case OPCODE_TXB:
out = emit_insn(c, inst);
out->TexSrcUnit = fp->program.Base.SamplerUnits[inst->TexSrcUnit];
assert(out->TexSrcUnit < BRW_MAX_TEX_UNIT);
break;
case OPCODE_XPD:
out = emit_insn(c, inst);
/* This should probably be done in the parser.
*/
out->DstReg.WriteMask &= WRITEMASK_XYZ;
break;
case OPCODE_KIL:
out = emit_insn(c, inst);
/* This should probably be done in the parser.
*/
out->DstReg.WriteMask = 0;
break;
case OPCODE_END:
emit_render_target_writes(c);
break;
case OPCODE_PRINT:
break;
default:
if (brw_wm_is_scalar_result(inst->Opcode))
emit_scalar_insn(c, inst);
else
emit_insn(c, inst);
break;
}
}
if (INTEL_DEBUG & DEBUG_WM) {
printf("pass_fp:\n");
print_insns( c->prog_instructions, c->nr_fp_insns );
printf("\n");
}
}
/**
* Some TEX instructions require extra code, cube map coordinate
* normalization, or coordinate scaling for RECT textures, etc.
* This function emits those extra instructions and the TEX
* instruction itself.
*/
static void precalc_tex( struct brw_wm_compile *c,
const struct prog_instruction *inst )
{
struct prog_src_register coord;
struct prog_dst_register tmpcoord;
const GLuint unit = c->fp->program.Base.SamplerUnits[inst->TexSrcUnit];
assert(unit < BRW_MAX_TEX_UNIT);
if (inst->TexSrcTarget == TEXTURE_CUBE_INDEX) {
struct prog_instruction *out;
struct prog_dst_register tmp0 = get_temp(c);
struct prog_src_register tmp0src = src_reg_from_dst(tmp0);
struct prog_dst_register tmp1 = get_temp(c);
struct prog_src_register tmp1src = src_reg_from_dst(tmp1);
struct prog_src_register src0 = inst->SrcReg[0];
/* find longest component of coord vector and normalize it */
tmpcoord = get_temp(c);
coord = src_reg_from_dst(tmpcoord);
/* tmpcoord = src0 (i.e.: coord = src0) */
out = emit_op(c, OPCODE_MOV,
tmpcoord,
0,
src0,
src_undef(),
src_undef());
out->SrcReg[0].Negate = NEGATE_NONE;
out->SrcReg[0].Abs = 1;
/* tmp0 = MAX(coord.X, coord.Y) */
emit_op(c, OPCODE_MAX,
tmp0,
0,
src_swizzle1(coord, X),
src_swizzle1(coord, Y),
src_undef());
/* tmp1 = MAX(tmp0, coord.Z) */
emit_op(c, OPCODE_MAX,
tmp1,
0,
tmp0src,
src_swizzle1(coord, Z),
src_undef());
/* tmp0 = 1 / tmp1 */
emit_op(c, OPCODE_RCP,
dst_mask(tmp0, WRITEMASK_X),
0,
tmp1src,
src_undef(),
src_undef());
/* tmpCoord = src0 * tmp0 */
emit_op(c, OPCODE_MUL,
tmpcoord,
0,
src0,
src_swizzle1(tmp0src, SWIZZLE_X),
src_undef());
release_temp(c, tmp0);
release_temp(c, tmp1);
}
else if (inst->TexSrcTarget == TEXTURE_RECT_INDEX) {
struct prog_src_register scale =
search_or_add_param5( c,
STATE_INTERNAL,
STATE_TEXRECT_SCALE,
unit,
0,0 );
tmpcoord = get_temp(c);
/* coord.xy = MUL inst->SrcReg[0], { 1/width, 1/height }
*/
emit_op(c,
OPCODE_MUL,
tmpcoord,
0,
inst->SrcReg[0],
src_swizzle(scale,
SWIZZLE_X,
SWIZZLE_Y,
SWIZZLE_ONE,
SWIZZLE_ONE),
src_undef());
coord = src_reg_from_dst(tmpcoord);
}
else {
coord = inst->SrcReg[0];
}
/* Need to emit YUV texture conversions by hand. Probably need to
* do this here - the alternative is in brw_wm_emit.c, but the
* conversion requires allocating a temporary variable which we
* don't have the facility to do that late in the compilation.
*/
if (c->key.yuvtex_mask & (1 << unit)) {
/* convert ycbcr to RGBA */
GLboolean swap_uv = c->key.yuvtex_swap_mask & (1<<unit);
/*
CONST C0 = { -.5, -.0625, -.5, 1.164 }
CONST C1 = { 1.596, -0.813, 2.018, -.391 }
UYV = TEX ...
UYV.xyz = ADD UYV, C0
UYV.y = MUL UYV.y, C0.w
if (UV swaped)
RGB.xyz = MAD UYV.zzx, C1, UYV.y
else
RGB.xyz = MAD UYV.xxz, C1, UYV.y
RGB.y = MAD UYV.z, C1.w, RGB.y
*/
struct prog_dst_register dst = inst->DstReg;
struct prog_dst_register tmp = get_temp(c);
struct prog_src_register tmpsrc = src_reg_from_dst(tmp);
struct prog_src_register C0 = search_or_add_const4f( c, -.5, -.0625, -.5, 1.164 );
struct prog_src_register C1 = search_or_add_const4f( c, 1.596, -0.813, 2.018, -.391 );
/* tmp = TEX ...
*/
emit_tex_op(c,
OPCODE_TEX,
tmp,
inst->SaturateMode,
unit,
inst->TexSrcTarget,
inst->TexShadow,
coord,
src_undef(),
src_undef());
/* tmp.xyz = ADD TMP, C0
*/
emit_op(c,
OPCODE_ADD,
dst_mask(tmp, WRITEMASK_XYZ),
0,
tmpsrc,
C0,
src_undef());
/* YUV.y = MUL YUV.y, C0.w
*/
emit_op(c,
OPCODE_MUL,
dst_mask(tmp, WRITEMASK_Y),
0,
tmpsrc,
src_swizzle1(C0, W),
src_undef());
/*
* if (UV swaped)
* RGB.xyz = MAD YUV.zzx, C1, YUV.y
* else
* RGB.xyz = MAD YUV.xxz, C1, YUV.y
*/
emit_op(c,
OPCODE_MAD,
dst_mask(dst, WRITEMASK_XYZ),
0,
swap_uv?src_swizzle(tmpsrc, Z,Z,X,X):src_swizzle(tmpsrc, X,X,Z,Z),
C1,
src_swizzle1(tmpsrc, Y));
/* RGB.y = MAD YUV.z, C1.w, RGB.y
*/
emit_op(c,
OPCODE_MAD,
dst_mask(dst, WRITEMASK_Y),
0,
src_swizzle1(tmpsrc, Z),
src_swizzle1(C1, W),
src_swizzle1(src_reg_from_dst(dst), Y));
release_temp(c, tmp);
}
else {
/* ordinary RGBA tex instruction */
emit_tex_op(c,
OPCODE_TEX,
inst->DstReg,
inst->SaturateMode,
unit,
inst->TexSrcTarget,
inst->TexShadow,
coord,
src_undef(),
src_undef());
}
/* For GL_EXT_texture_swizzle: */
if (c->key.tex_swizzles[unit] != SWIZZLE_NOOP) {
/* swizzle the result of the TEX instruction */
struct prog_src_register tmpsrc = src_reg_from_dst(inst->DstReg);
emit_op(c, OPCODE_SWZ,
inst->DstReg,
SATURATE_OFF, /* saturate already done above */
src_swizzle4(tmpsrc, c->key.tex_swizzles[unit]),
src_undef(),
src_undef());
}
if ((inst->TexSrcTarget == TEXTURE_RECT_INDEX) ||
(inst->TexSrcTarget == TEXTURE_CUBE_INDEX))
release_temp(c, tmpcoord);
}
/**
* Some TEX instructions require extra code, cube map coordinate
* normalization, or coordinate scaling for RECT textures, etc.
* This function emits those extra instructions and the TEX
* instruction itself.
*/
static void precalc_tex( struct brw_wm_compile *c,
struct brw_fp_dst dst,
unsigned target,
unsigned unit,
struct brw_fp_src src0,
struct brw_fp_src sampler )
{
struct brw_fp_src coord;
struct brw_fp_dst tmp = dst_undef();
assert(unit < BRW_MAX_TEX_UNIT);
/* Cubemap: find longest component of coord vector and normalize
* it.
*/
if (target == TGSI_TEXTURE_CUBE) {
struct brw_fp_src tmpsrc;
tmp = get_temp(c);
tmpsrc = src_reg_from_dst(tmp);
/* tmp = abs(src0) */
emit_op1(c,
TGSI_OPCODE_MOV,
tmp,
src_abs(src0));
/* tmp.X = MAX(tmp.X, tmp.Y) */
emit_op2(c, TGSI_OPCODE_MAX,
dst_mask(tmp, BRW_WRITEMASK_X),
src_scalar(tmpsrc, X),
src_scalar(tmpsrc, Y));
/* tmp.X = MAX(tmp.X, tmp.Z) */
emit_op2(c, TGSI_OPCODE_MAX,
dst_mask(tmp, BRW_WRITEMASK_X),
tmpsrc,
src_scalar(tmpsrc, Z));
/* tmp.X = 1 / tmp.X */
emit_op1(c, TGSI_OPCODE_RCP,
dst_mask(tmp, BRW_WRITEMASK_X),
tmpsrc);
/* tmp = src0 * tmp.xxxx */
emit_op2(c, TGSI_OPCODE_MUL,
tmp,
src0,
src_scalar(tmpsrc, X));
coord = tmpsrc;
}
else if (target == TGSI_TEXTURE_RECT ||
target == TGSI_TEXTURE_SHADOWRECT) {
/* XXX: need a mechanism for internally generated constants.
*/
coord = src0;
}
else {
coord = src0;
}
/* Need to emit YUV texture conversions by hand. Probably need to
* do this here - the alternative is in brw_wm_emit.c, but the
* conversion requires allocating a temporary variable which we
* don't have the facility to do that late in the compilation.
*/
if (c->key.yuvtex_mask & (1 << unit)) {
/* convert ycbcr to RGBA */
GLboolean swap_uv = c->key.yuvtex_swap_mask & (1<<unit);
struct brw_fp_dst tmp = get_temp(c);
struct brw_fp_src tmpsrc = src_reg_from_dst(tmp);
struct brw_fp_src C0 = src_imm4f( c, -.5, -.0625, -.5, 1.164 );
struct brw_fp_src C1 = src_imm4f( c, 1.596, -0.813, 2.018, -.391 );
/* tmp = TEX ...
*/
emit_tex_op(c,
TGSI_OPCODE_TEX,
dst_saturate(tmp, dst.saturate),
unit,
target,
sampler.index,
coord,
src_undef(),
src_undef());
/* tmp.xyz = ADD TMP, C0
*/
emit_op2(c, TGSI_OPCODE_ADD,
dst_mask(tmp, BRW_WRITEMASK_XYZ),
tmpsrc,
C0);
/* YUV.y = MUL YUV.y, C0.w
*/
emit_op2(c, TGSI_OPCODE_MUL,
dst_mask(tmp, BRW_WRITEMASK_Y),
tmpsrc,
src_scalar(C0, W));
/*
* if (UV swaped)
* RGB.xyz = MAD YUV.zzx, C1, YUV.y
* else
* RGB.xyz = MAD YUV.xxz, C1, YUV.y
*/
emit_op3(c, TGSI_OPCODE_MAD,
dst_mask(dst, BRW_WRITEMASK_XYZ),
( swap_uv ?
src_swizzle(tmpsrc, Z,Z,X,X) :
src_swizzle(tmpsrc, X,X,Z,Z)),
C1,
src_scalar(tmpsrc, Y));
/* RGB.y = MAD YUV.z, C1.w, RGB.y
*/
emit_op3(c,
TGSI_OPCODE_MAD,
dst_mask(dst, BRW_WRITEMASK_Y),
src_scalar(tmpsrc, Z),
src_scalar(C1, W),
src_scalar(src_reg_from_dst(dst), Y));
release_temp(c, tmp);
}
else {
/* ordinary RGBA tex instruction */
emit_tex_op(c,
TGSI_OPCODE_TEX,
dst,
unit,
target,
sampler.index,
coord,
src_undef(),
src_undef());
}
/* XXX: add GL_EXT_texture_swizzle support to gallium -- by
* generating shader variants in mesa state tracker.
*/
/* Release this temp if we ended up allocating it:
*/
if (!dst_is_undef(tmp))
release_temp(c, tmp);
}
static void emit_interp( struct brw_wm_compile *c,
GLuint idx )
{
struct prog_dst_register dst = dst_reg(PROGRAM_INPUT, idx);
struct prog_src_register interp = src_reg(PROGRAM_PAYLOAD, idx);
struct prog_src_register deltas = get_delta_xy(c);
struct prog_src_register arg2;
GLuint opcode;
/* Need to use PINTERP on attributes which have been
* multiplied by 1/W in the SF program, and LINTERP on those
* which have not:
*/
switch (idx) {
case FRAG_ATTRIB_WPOS:
opcode = WM_LINTERP;
arg2 = src_undef();
/* Have to treat wpos.xy specially:
*/
emit_op(c,
WM_WPOSXY,
dst_mask(dst, WRITEMASK_XY),
0, 0, 0,
get_pixel_xy(c),
src_undef(),
src_undef());
dst = dst_mask(dst, WRITEMASK_ZW);
/* PROGRAM_INPUT.attr.xyzw = INTERP payload.interp[attr].x, deltas.xyw
*/
emit_op(c,
WM_LINTERP,
dst,
0, 0, 0,
interp,
deltas,
arg2);
break;
case FRAG_ATTRIB_COL0:
case FRAG_ATTRIB_COL1:
if (c->key.flat_shade) {
emit_op(c,
WM_CINTERP,
dst,
0, 0, 0,
interp,
src_undef(),
src_undef());
}
else {
emit_op(c,
WM_LINTERP,
dst,
0, 0, 0,
interp,
deltas,
src_undef());
}
break;
default:
emit_op(c,
WM_PINTERP,
dst,
0, 0, 0,
interp,
deltas,
get_pixel_w(c));
break;
}
c->fp_interp_emitted |= 1<<idx;
}
