/**
* Transform the trigonometric functions COS, SIN, and SCS
* so that the input to COS and SIN is always in the range [-PI, PI].
* SCS is replaced by one COS and one SIN instruction.
*/
int r300_transform_trig_scale_vertex(struct radeon_compiler *c,
struct rc_instruction *inst,
void *unused)
{
static const float cons[4] = {0.15915494309189535, 0.5, 6.28318530717959, -3.14159265358979};
unsigned int temp;
unsigned int constant;
if (inst->U.I.Opcode != RC_OPCODE_COS &&
inst->U.I.Opcode != RC_OPCODE_SIN &&
inst->U.I.Opcode != RC_OPCODE_SCS)
return 0;
/* Repeat x in the range [-PI, PI]:
*
* repeat(x) = frac(x / 2PI + 0.5) * 2PI - PI
*/
temp = rc_find_free_temporary(c);
constant = rc_constants_add_immediate_vec4(&c->Program.Constants, cons);
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(temp, RC_MASK_W),
swizzle_xxxx(inst->U.I.SrcReg[0]),
srcregswz(RC_FILE_CONSTANT, constant, RC_SWIZZLE_XXXX),
srcregswz(RC_FILE_CONSTANT, constant, RC_SWIZZLE_YYYY));
emit1(c, inst->Prev, RC_OPCODE_FRC, 0, dstregtmpmask(temp, RC_MASK_W),
srcreg(RC_FILE_TEMPORARY, temp));
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(temp, RC_MASK_W),
srcreg(RC_FILE_TEMPORARY, temp),
srcregswz(RC_FILE_CONSTANT, constant, RC_SWIZZLE_ZZZZ),
srcregswz(RC_FILE_CONSTANT, constant, RC_SWIZZLE_WWWW));
r300_transform_SIN_COS_SCS(c, inst, temp);
return 1;
}
/**
* Transform the trigonometric functions COS, SIN, and SCS
* to include pre-scaling by 1/(2*PI) and taking the fractional
* part, so that the input to COS and SIN is always in the range [0,1).
* SCS is replaced by one COS and one SIN instruction.
*
* @warning This transformation implicitly changes the semantics of SIN and COS!
*/
int radeonTransformTrigScale(struct radeon_compiler* c,
struct rc_instruction* inst,
void* unused)
{
static const float RCP_2PI = 0.15915494309189535;
unsigned int temp;
unsigned int constant;
unsigned int constant_swizzle;
if (inst->U.I.Opcode != RC_OPCODE_COS &&
inst->U.I.Opcode != RC_OPCODE_SIN &&
inst->U.I.Opcode != RC_OPCODE_SCS)
return 0;
temp = rc_find_free_temporary(c);
constant = rc_constants_add_immediate_scalar(&c->Program.Constants, RCP_2PI, &constant_swizzle);
emit2(c, inst->Prev, RC_OPCODE_MUL, 0, dstregtmpmask(temp, RC_MASK_W),
swizzle_xxxx(inst->U.I.SrcReg[0]),
srcregswz(RC_FILE_CONSTANT, constant, constant_swizzle));
emit1(c, inst->Prev, RC_OPCODE_FRC, 0, dstregtmpmask(temp, RC_MASK_W),
srcreg(RC_FILE_TEMPORARY, temp));
r300_transform_SIN_COS_SCS(c, inst, temp);
return 1;
}
/* dst = ROUND(src) :
* add = src + .5
* frac = FRC(add)
* dst = add - frac
*
* According to the GLSL spec, the implementor can decide which way to round
* when the fraction is .5. We round down for .5.
*
*/
static void transform_ROUND(struct radeon_compiler* c,
struct rc_instruction* inst)
{
unsigned int mask = inst->U.I.DstReg.WriteMask;
unsigned int frac_index, add_index;
struct rc_dst_register frac_dst, add_dst;
struct rc_src_register frac_src, add_src;
/* add = src + .5 */
add_index = rc_find_free_temporary(c);
add_dst = dstregtmpmask(add_index, mask);
emit2(c, inst->Prev, RC_OPCODE_ADD, 0, add_dst, inst->U.I.SrcReg[0],
builtin_half);
add_src = srcreg(RC_FILE_TEMPORARY, add_dst.Index);
/* frac = FRC(add) */
frac_index = rc_find_free_temporary(c);
frac_dst = dstregtmpmask(frac_index, mask);
emit1(c, inst->Prev, RC_OPCODE_FRC, 0, frac_dst, add_src);
frac_src = srcreg(RC_FILE_TEMPORARY, frac_dst.Index);
/* dst = add - frac */
emit2(c, inst->Prev, RC_OPCODE_ADD, 0, inst->U.I.DstReg,
add_src, negate(frac_src));
rc_remove_instruction(inst);
}
static void transform_r300_vertex_SSG(struct radeon_compiler* c,
struct rc_instruction* inst)
{
/* result = sign(x)
*
* SLT tmp0, 0, x;
* SLT tmp1, x, 0;
* ADD result, tmp0, -tmp1;
*/
struct rc_dst_register dst0 = try_to_reuse_dst(c, inst);
unsigned tmp1;
/* 0 < x */
dst0 = try_to_reuse_dst(c, inst);
emit2(c, inst->Prev, RC_OPCODE_SLT, 0,
dst0,
builtin_zero,
inst->U.I.SrcReg[0]);
/* x < 0 */
tmp1 = rc_find_free_temporary(c);
emit2(c, inst->Prev, RC_OPCODE_SLT, 0,
dstregtmpmask(tmp1, inst->U.I.DstReg.WriteMask),
inst->U.I.SrcReg[0],
builtin_zero);
/* Either both are zero, or one of them is one and the other is zero. */
/* result = tmp0 - tmp1 */
emit2(c, inst->Prev, RC_OPCODE_ADD, 0,
inst->U.I.DstReg,
srcreg(RC_FILE_TEMPORARY, dst0.Index),
negate(srcreg(RC_FILE_TEMPORARY, tmp1)));
rc_remove_instruction(inst);
}
static void transform_r300_vertex_SNE(struct radeon_compiler *c,
struct rc_instruction *inst)
{
/* x != y <==> x < y || y < x */
int tmp = rc_find_free_temporary(c);
/* x < y */
emit2(c, inst->Prev, RC_OPCODE_SLT, 0,
dstregtmpmask(tmp, inst->U.I.DstReg.WriteMask),
inst->U.I.SrcReg[0],
inst->U.I.SrcReg[1]);
/* y < x */
emit2(c, inst->Prev, RC_OPCODE_SLT, 0,
inst->U.I.DstReg,
inst->U.I.SrcReg[1],
inst->U.I.SrcReg[0]);
/* x || y = max(x, y) */
emit2(c, inst->Prev, RC_OPCODE_MAX, 0,
inst->U.I.DstReg,
srcreg(RC_FILE_TEMPORARY, tmp),
srcreg(inst->U.I.DstReg.File, inst->U.I.DstReg.Index));
rc_remove_instruction(inst);
}
static void transform_r300_vertex_SEQ(struct radeon_compiler *c,
struct rc_instruction *inst)
{
/* x = y <==> x >= y && y >= x */
int tmp = rc_find_free_temporary(c);
/* x <= y */
emit2(c, inst->Prev, RC_OPCODE_SGE, 0,
dstregtmpmask(tmp, inst->U.I.DstReg.WriteMask),
inst->U.I.SrcReg[0],
inst->U.I.SrcReg[1]);
/* y <= x */
emit2(c, inst->Prev, RC_OPCODE_SGE, 0,
inst->U.I.DstReg,
inst->U.I.SrcReg[1],
inst->U.I.SrcReg[0]);
/* x && y = x * y */
emit2(c, inst->Prev, RC_OPCODE_MUL, 0,
inst->U.I.DstReg,
srcreg(RC_FILE_TEMPORARY, tmp),
srcreg(inst->U.I.DstReg.File, inst->U.I.DstReg.Index));
rc_remove_instruction(inst);
}
static struct rc_dst_register try_to_reuse_dst(struct radeon_compiler *c,
struct rc_instruction *inst)
{
unsigned tmp;
if (is_dst_safe_to_reuse(inst))
tmp = inst->U.I.DstReg.Index;
else
tmp = rc_find_free_temporary(c);
return dstregtmpmask(tmp, inst->U.I.DstReg.WriteMask);
}
/**
* Approximate sin(x), where x is clamped to (-pi/2, pi/2).
*
* MUL tmp.xy, src, { 4/PI, -4/(PI^2) }
* MAD tmp.x, tmp.y, |src|, tmp.x
* MAD tmp.y, tmp.x, |tmp.x|, -tmp.x
* MAD dest, tmp.y, weight, tmp.x
*/
static void sin_approx(
struct radeon_compiler* c, struct rc_instruction * inst,
struct rc_dst_register dst, struct rc_src_register src, const unsigned int* constants)
{
unsigned int tempreg = rc_find_free_temporary(c);
emit2(c, inst->Prev, RC_OPCODE_MUL, 0, dstregtmpmask(tempreg, RC_MASK_XY),
swizzle_xxxx(src),
srcreg(RC_FILE_CONSTANT, constants[0]));
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(tempreg, RC_MASK_X),
swizzle_yyyy(srcreg(RC_FILE_TEMPORARY, tempreg)),
absolute(swizzle_xxxx(src)),
swizzle_xxxx(srcreg(RC_FILE_TEMPORARY, tempreg)));
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(tempreg, RC_MASK_Y),
swizzle_xxxx(srcreg(RC_FILE_TEMPORARY, tempreg)),
absolute(swizzle_xxxx(srcreg(RC_FILE_TEMPORARY, tempreg))),
negate(swizzle_xxxx(srcreg(RC_FILE_TEMPORARY, tempreg))));
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dst,
swizzle_yyyy(srcreg(RC_FILE_TEMPORARY, tempreg)),
swizzle_wwww(srcreg(RC_FILE_CONSTANT, constants[0])),
swizzle_xxxx(srcreg(RC_FILE_TEMPORARY, tempreg)));
}
static void transform_r300_vertex_fix_LIT(struct radeon_compiler* c,
struct rc_instruction* inst)
{
struct rc_dst_register dst = try_to_reuse_dst(c, inst);
unsigned constant_swizzle;
int constant = rc_constants_add_immediate_scalar(&c->Program.Constants,
0.0000000000000000001,
&constant_swizzle);
/* MOV dst, src */
dst.WriteMask = RC_MASK_XYZW;
emit1(c, inst->Prev, RC_OPCODE_MOV, 0,
dst,
inst->U.I.SrcReg[0]);
/* MAX dst.y, src, 0.00...001 */
emit2(c, inst->Prev, RC_OPCODE_MAX, 0,
dstregtmpmask(dst.Index, RC_MASK_Y),
srcreg(RC_FILE_TEMPORARY, dst.Index),
srcregswz(RC_FILE_CONSTANT, constant, constant_swizzle));
inst->U.I.SrcReg[0] = srcreg(RC_FILE_TEMPORARY, dst.Index);
}
/**
* Translate the trigonometric functions COS, SIN, and SCS
* using only the basic instructions
* MOV, ADD, MUL, MAD, FRC
*/
int r300_transform_trig_simple(struct radeon_compiler* c,
struct rc_instruction* inst,
void* unused)
{
unsigned int constants[2];
unsigned int tempreg;
if (inst->U.I.Opcode != RC_OPCODE_COS &&
inst->U.I.Opcode != RC_OPCODE_SIN &&
inst->U.I.Opcode != RC_OPCODE_SCS)
return 0;
tempreg = rc_find_free_temporary(c);
sincos_constants(c, constants);
if (inst->U.I.Opcode == RC_OPCODE_COS) {
/* MAD tmp.x, src, 1/(2*PI), 0.75 */
/* FRC tmp.x, tmp.x */
/* MAD tmp.z, tmp.x, 2*PI, -PI */
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(tempreg, RC_MASK_W),
swizzle_xxxx(inst->U.I.SrcReg[0]),
swizzle_zzzz(srcreg(RC_FILE_CONSTANT, constants[1])),
swizzle_xxxx(srcreg(RC_FILE_CONSTANT, constants[1])));
emit1(c, inst->Prev, RC_OPCODE_FRC, 0, dstregtmpmask(tempreg, RC_MASK_W),
swizzle_wwww(srcreg(RC_FILE_TEMPORARY, tempreg)));
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(tempreg, RC_MASK_W),
swizzle_wwww(srcreg(RC_FILE_TEMPORARY, tempreg)),
swizzle_wwww(srcreg(RC_FILE_CONSTANT, constants[1])),
negate(swizzle_zzzz(srcreg(RC_FILE_CONSTANT, constants[0]))));
sin_approx(c, inst, inst->U.I.DstReg,
swizzle_wwww(srcreg(RC_FILE_TEMPORARY, tempreg)),
constants);
} else if (inst->U.I.Opcode == RC_OPCODE_SIN) {
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(tempreg, RC_MASK_W),
swizzle_xxxx(inst->U.I.SrcReg[0]),
swizzle_zzzz(srcreg(RC_FILE_CONSTANT, constants[1])),
swizzle_yyyy(srcreg(RC_FILE_CONSTANT, constants[1])));
emit1(c, inst->Prev, RC_OPCODE_FRC, 0, dstregtmpmask(tempreg, RC_MASK_W),
swizzle_wwww(srcreg(RC_FILE_TEMPORARY, tempreg)));
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(tempreg, RC_MASK_W),
swizzle_wwww(srcreg(RC_FILE_TEMPORARY, tempreg)),
swizzle_wwww(srcreg(RC_FILE_CONSTANT, constants[1])),
negate(swizzle_zzzz(srcreg(RC_FILE_CONSTANT, constants[0]))));
sin_approx(c, inst, inst->U.I.DstReg,
swizzle_wwww(srcreg(RC_FILE_TEMPORARY, tempreg)),
constants);
} else {
struct rc_dst_register dst;
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(tempreg, RC_MASK_XY),
swizzle_xxxx(inst->U.I.SrcReg[0]),
swizzle_zzzz(srcreg(RC_FILE_CONSTANT, constants[1])),
swizzle(srcreg(RC_FILE_CONSTANT, constants[1]), RC_SWIZZLE_X, RC_SWIZZLE_Y, RC_SWIZZLE_Z, RC_SWIZZLE_W));
emit1(c, inst->Prev, RC_OPCODE_FRC, 0, dstregtmpmask(tempreg, RC_MASK_XY),
srcreg(RC_FILE_TEMPORARY, tempreg));
emit3(c, inst->Prev, RC_OPCODE_MAD, 0, dstregtmpmask(tempreg, RC_MASK_XY),
srcreg(RC_FILE_TEMPORARY, tempreg),
swizzle_wwww(srcreg(RC_FILE_CONSTANT, constants[1])),
negate(swizzle_zzzz(srcreg(RC_FILE_CONSTANT, constants[0]))));
dst = inst->U.I.DstReg;
dst.WriteMask = inst->U.I.DstReg.WriteMask & RC_MASK_X;
sin_approx(c, inst, dst,
swizzle_xxxx(srcreg(RC_FILE_TEMPORARY, tempreg)),
constants);
dst.WriteMask = inst->U.I.DstReg.WriteMask & RC_MASK_Y;
sin_approx(c, inst, dst,
swizzle_yyyy(srcreg(RC_FILE_TEMPORARY, tempreg)),
constants);
}
rc_remove_instruction(inst);
return 1;
}
/**
* Definition of LIT (from ARB_fragment_program):
*
* tmp = VectorLoad(op0);
* if (tmp.x < 0) tmp.x = 0;
* if (tmp.y < 0) tmp.y = 0;
* if (tmp.w < -(128.0-epsilon)) tmp.w = -(128.0-epsilon);
* else if (tmp.w > 128-epsilon) tmp.w = 128-epsilon;
* result.x = 1.0;
* result.y = tmp.x;
* result.z = (tmp.x > 0) ? RoughApproxPower(tmp.y, tmp.w) : 0.0;
* result.w = 1.0;
*
* The longest path of computation is the one leading to result.z,
* consisting of 5 operations. This implementation of LIT takes
* 5 slots, if the subsequent optimization passes are clever enough
* to pair instructions correctly.
*/
static void transform_LIT(struct radeon_compiler* c,
struct rc_instruction* inst)
{
unsigned int constant;
unsigned int constant_swizzle;
unsigned int temp;
struct rc_src_register srctemp;
constant = rc_constants_add_immediate_scalar(&c->Program.Constants, -127.999999, &constant_swizzle);
if (inst->U.I.DstReg.WriteMask != RC_MASK_XYZW || inst->U.I.DstReg.File != RC_FILE_TEMPORARY) {
struct rc_instruction * inst_mov;
inst_mov = emit1(c, inst,
RC_OPCODE_MOV, 0, inst->U.I.DstReg,
srcreg(RC_FILE_TEMPORARY, rc_find_free_temporary(c)));
inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
inst->U.I.DstReg.Index = inst_mov->U.I.SrcReg[0].Index;
inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
}
temp = inst->U.I.DstReg.Index;
srctemp = srcreg(RC_FILE_TEMPORARY, temp);
/* tmp.x = max(0.0, Src.x); */
/* tmp.y = max(0.0, Src.y); */
/* tmp.w = clamp(Src.z, -128+eps, 128-eps); */
emit2(c, inst->Prev, RC_OPCODE_MAX, 0,
dstregtmpmask(temp, RC_MASK_XYW),
inst->U.I.SrcReg[0],
swizzle(srcreg(RC_FILE_CONSTANT, constant),
RC_SWIZZLE_ZERO, RC_SWIZZLE_ZERO, RC_SWIZZLE_ZERO, constant_swizzle&3));
emit2(c, inst->Prev, RC_OPCODE_MIN, 0,
dstregtmpmask(temp, RC_MASK_Z),
swizzle_wwww(srctemp),
negate(srcregswz(RC_FILE_CONSTANT, constant, constant_swizzle)));
/* tmp.w = Pow(tmp.y, tmp.w) */
emit1(c, inst->Prev, RC_OPCODE_LG2, 0,
dstregtmpmask(temp, RC_MASK_W),
swizzle_yyyy(srctemp));
emit2(c, inst->Prev, RC_OPCODE_MUL, 0,
dstregtmpmask(temp, RC_MASK_W),
swizzle_wwww(srctemp),
swizzle_zzzz(srctemp));
emit1(c, inst->Prev, RC_OPCODE_EX2, 0,
dstregtmpmask(temp, RC_MASK_W),
swizzle_wwww(srctemp));
/* tmp.z = (tmp.x > 0) ? tmp.w : 0.0 */
emit3(c, inst->Prev, RC_OPCODE_CMP, inst->U.I.SaturateMode,
dstregtmpmask(temp, RC_MASK_Z),
negate(swizzle_xxxx(srctemp)),
swizzle_wwww(srctemp),
builtin_zero);
/* tmp.x, tmp.y, tmp.w = 1.0, tmp.x, 1.0 */
emit1(c, inst->Prev, RC_OPCODE_MOV, inst->U.I.SaturateMode,
dstregtmpmask(temp, RC_MASK_XYW),
swizzle(srctemp, RC_SWIZZLE_ONE, RC_SWIZZLE_X, RC_SWIZZLE_ONE, RC_SWIZZLE_ONE));
rc_remove_instruction(inst);
}