/**
* Special case for converting clamped IEEE-754 floats to unsigned norms.
*
* The mathematical voodoo below may seem excessive but it is actually
* paramount we do it this way for several reasons. First, there is no single
* precision FP to unsigned integer conversion Intel SSE instruction. Second,
* secondly, even if there was, since the FP's mantissa takes only a fraction
* of register bits the typically scale and cast approach would require double
* precision for accurate results, and therefore half the throughput
*
* Although the result values can be scaled to an arbitrary bit width specified
* by dst_width, the actual result type will have the same width.
*/
LLVMValueRef
lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
struct lp_type src_type,
unsigned dst_width,
LLVMValueRef src)
{
LLVMTypeRef int_vec_type = lp_build_int_vec_type(src_type);
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(src_type.floating);
mantissa = lp_mantissa(src_type);
/* We cannot carry more bits than the mantissa */
n = MIN2(mantissa, dst_width);
/* This magic coefficients will make the desired result to appear in the
* lowest significant bits of the mantissa.
*/
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)mask/ubound;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = LLVMBuildMul(builder, src, lp_build_const_scalar(src_type, scale), "");
res = LLVMBuildAdd(builder, res, lp_build_const_scalar(src_type, bias), "");
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
if(dst_width > n) {
int shift = dst_width - n;
res = LLVMBuildShl(builder, res, lp_build_int_const_scalar(src_type, shift), "");
/* TODO: Fill in the empty lower bits for additional precision? */
#if 0
{
LLVMValueRef msb;
msb = LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, dst_width - 1), "");
msb = LLVMBuildShl(builder, msb, lp_build_int_const_scalar(src_type, shift), "");
msb = LLVMBuildSub(builder, msb, lp_build_int_const_scalar(src_type, 1), "");
res = LLVMBuildOr(builder, res, msb, "");
}
#elif 0
while(shift > 0) {
res = LLVMBuildOr(builder, res, LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, n), ""), "");
shift -= n;
n *= 2;
}
#endif
}
else
res = LLVMBuildAnd(builder, res, lp_build_int_const_scalar(src_type, mask), "");
return res;
}
/**
* Inverse of lp_build_clamped_float_to_unsigned_norm above.
* Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
* return {float, float, float, float} with values in range [0, 1].
*/
LLVMValueRef
lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
unsigned src_width,
struct lp_type dst_type,
LLVMValueRef src)
{
LLVMTypeRef vec_type = lp_build_vec_type(dst_type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(dst_type);
LLVMValueRef bias_;
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(dst_type.floating);
mantissa = lp_mantissa(dst_type);
n = MIN2(mantissa, src_width);
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)ubound/mask;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = src;
if(src_width > mantissa) {
int shift = src_width - mantissa;
res = LLVMBuildLShr(builder, res, lp_build_const_int_vec(dst_type, shift), "");
}
bias_ = lp_build_const_vec(dst_type, bias);
res = LLVMBuildOr(builder,
res,
LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
res = LLVMBuildBitCast(builder, res, vec_type, "");
res = LLVMBuildFSub(builder, res, bias_, "");
res = LLVMBuildFMul(builder, res, lp_build_const_vec(dst_type, scale), "");
return res;
}
/**
* Small vector x scale multiplication optimization.
*/
LLVMValueRef
lp_build_mul_imm(struct lp_build_context *bld,
LLVMValueRef a,
int b)
{
LLVMValueRef factor;
if(b == 0)
return bld->zero;
if(b == 1)
return a;
if(b == -1)
return LLVMBuildNeg(bld->builder, a, "");
if(b == 2 && bld->type.floating)
return lp_build_add(bld, a, a);
if(util_is_pot(b)) {
unsigned shift = ffs(b) - 1;
if(bld->type.floating) {
#if 0
/*
* Power of two multiplication by directly manipulating the mantissa.
*
* XXX: This might not be always faster, it will introduce a small error
* for multiplication by zero, and it will produce wrong results
* for Inf and NaN.
*/
unsigned mantissa = lp_mantissa(bld->type);
factor = lp_build_int_const_scalar(bld->type, (unsigned long long)shift << mantissa);
a = LLVMBuildBitCast(bld->builder, a, lp_build_int_vec_type(bld->type), "");
a = LLVMBuildAdd(bld->builder, a, factor, "");
a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(bld->type), "");
return a;
#endif
}
else {
factor = lp_build_const_scalar(bld->type, shift);
return LLVMBuildShl(bld->builder, a, factor, "");
}
}
factor = lp_build_const_scalar(bld->type, (double)b);
return lp_build_mul(bld, a, factor);
}
/**
* Convert float[] to int[] with floor().
*/
LLVMValueRef
lp_build_ifloor(struct lp_build_context *bld,
LLVMValueRef a)
{
const struct lp_type type = bld->type;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef res;
assert(type.floating);
assert(lp_check_value(type, a));
if(util_cpu_caps.has_sse4_1) {
res = lp_build_round_sse41(bld, a, LP_BUILD_ROUND_SSE41_FLOOR);
}
else {
/* Take the sign bit and add it to 1 constant */
LLVMTypeRef vec_type = lp_build_vec_type(type);
unsigned mantissa = lp_mantissa(type);
LLVMValueRef mask = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
LLVMValueRef sign;
LLVMValueRef offset;
/* sign = a < 0 ? ~0 : 0 */
sign = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
sign = LLVMBuildAnd(bld->builder, sign, mask, "");
sign = LLVMBuildAShr(bld->builder, sign, lp_build_int_const_scalar(type, type.width - 1), "");
lp_build_name(sign, "floor.sign");
/* offset = -0.99999(9)f */
offset = lp_build_const_scalar(type, -(double)(((unsigned long long)1 << mantissa) - 1)/((unsigned long long)1 << mantissa));
offset = LLVMConstBitCast(offset, int_vec_type);
/* offset = a < 0 ? -0.99999(9)f : 0.0f */
offset = LLVMBuildAnd(bld->builder, offset, sign, "");
offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "");
lp_build_name(offset, "floor.offset");
res = LLVMBuildAdd(bld->builder, a, offset, "");
lp_build_name(res, "floor.res");
}
res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
lp_build_name(res, "floor");
return res;
}
/**
* Inverse of lp_build_clamped_float_to_unsigned_norm above.
* Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
* return {float, float, float, float} with values in range [0, 1].
*/
LLVMValueRef
lp_build_unsigned_norm_to_float(struct gallivm_state *gallivm,
unsigned src_width,
struct lp_type dst_type,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, dst_type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, dst_type);
LLVMValueRef bias_;
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(dst_type.floating);
mantissa = lp_mantissa(dst_type);
if (src_width <= (mantissa + 1)) {
/*
* The source width matches fits what can be represented in floating
* point (i.e., mantissa + 1 bits). So do a straight multiplication
* followed by casting. No further rounding is necessary.
*/
scale = 1.0/(double)((1ULL << src_width) - 1);
res = LLVMBuildSIToFP(builder, src, vec_type, "");
res = LLVMBuildFMul(builder, res,
lp_build_const_vec(gallivm, dst_type, scale), "");
return res;
}
else {
/*
* The source width exceeds what can be represented in floating
* point. So truncate the incoming values.
*/
n = MIN2(mantissa, src_width);
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)ubound/mask;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = src;
if (src_width > mantissa) {
int shift = src_width - mantissa;
res = LLVMBuildLShr(builder, res,
lp_build_const_int_vec(gallivm, dst_type, shift), "");
}
bias_ = lp_build_const_vec(gallivm, dst_type, bias);
res = LLVMBuildOr(builder,
res,
LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
res = LLVMBuildBitCast(builder, res, vec_type, "");
res = LLVMBuildFSub(builder, res, bias_, "");
res = LLVMBuildFMul(builder, res, lp_build_const_vec(gallivm, dst_type, scale), "");
}
return res;
}
/**
* Special case for converting clamped IEEE-754 floats to unsigned norms.
*
* The mathematical voodoo below may seem excessive but it is actually
* paramount we do it this way for several reasons. First, there is no single
* precision FP to unsigned integer conversion Intel SSE instruction. Second,
* secondly, even if there was, since the FP's mantissa takes only a fraction
* of register bits the typically scale and cast approach would require double
* precision for accurate results, and therefore half the throughput
*
* Although the result values can be scaled to an arbitrary bit width specified
* by dst_width, the actual result type will have the same width.
*
* Ex: src = { float, float, float, float }
* return { i32, i32, i32, i32 } where each value is in [0, 2^dst_width-1].
*/
LLVMValueRef
lp_build_clamped_float_to_unsigned_norm(struct gallivm_state *gallivm,
struct lp_type src_type,
unsigned dst_width,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, src_type);
LLVMValueRef res;
unsigned mantissa;
assert(src_type.floating);
assert(dst_width <= src_type.width);
src_type.sign = FALSE;
mantissa = lp_mantissa(src_type);
if (dst_width <= mantissa) {
/*
* Apply magic coefficients that will make the desired result to appear
* in the lowest significant bits of the mantissa, with correct rounding.
*
* This only works if the destination width fits in the mantissa.
*/
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
ubound = (1ULL << dst_width);
mask = ubound - 1;
scale = (double)mask/ubound;
bias = (double)(1ULL << (mantissa - dst_width));
res = LLVMBuildFMul(builder, src, lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFAdd(builder, res, lp_build_const_vec(gallivm, src_type, bias), "");
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
res = LLVMBuildAnd(builder, res,
lp_build_const_int_vec(gallivm, src_type, mask), "");
}
else if (dst_width == (mantissa + 1)) {
/*
* The destination width matches exactly what can be represented in
* floating point (i.e., mantissa + 1 bits). So do a straight
* multiplication followed by casting. No further rounding is necessary.
*/
double scale;
scale = (double)((1ULL << dst_width) - 1);
res = LLVMBuildFMul(builder, src,
lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
}
else {
/*
* The destination exceeds what can be represented in the floating point.
* So multiply by the largest power two we get away with, and when
* subtract the most significant bit to rescale to normalized values.
*
* The largest power of two factor we can get away is
* (1 << (src_type.width - 1)), because we need to use signed . In theory it
* should be (1 << (src_type.width - 2)), but IEEE 754 rules states
* INT_MIN should be returned in FPToSI, which is the correct result for
* values near 1.0!
*
* This means we get (src_type.width - 1) correct bits for values near 0.0,
* and (mantissa + 1) correct bits for values near 1.0. Equally or more
* important, we also get exact results for 0.0 and 1.0.
*/
unsigned n = MIN2(src_type.width - 1, dst_width);
double scale = (double)(1ULL << n);
unsigned lshift = dst_width - n;
unsigned rshift = n;
LLVMValueRef lshifted;
LLVMValueRef rshifted;
res = LLVMBuildFMul(builder, src,
lp_build_const_vec(gallivm, src_type, scale), "");
res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
/*
* Align the most significant bit to its final place.
*
* This will cause 1.0 to overflow to 0, but the later adjustment will
* get it right.
*/
if (lshift) {
lshifted = LLVMBuildShl(builder, res,
lp_build_const_int_vec(gallivm, src_type,
lshift), "");
} else {
lshifted = res;
}
/*
* Align the most significant bit to the right.
*/
rshifted = LLVMBuildLShr(builder, res,
lp_build_const_int_vec(gallivm, src_type, rshift),
"");
/*
* Subtract the MSB to the LSB, therefore re-scaling from
* (1 << dst_width) to ((1 << dst_width) - 1).
*/
res = LLVMBuildSub(builder, lshifted, rshifted, "");
}
return res;
}
