/* Special xdr_db_dict_desc that understands optional version number at end. */
bool_t
xdr_db_dict_desc(XDR *xdrs, db_dict_desc *objp)
{
if (!xdr_db_dict_version(xdrs, &objp->impl_vers))
return (FALSE);
if (!xdr_array(xdrs, (char **)&objp->tables.tables_val,
(uint_t *)&objp->tables.tables_len, ~0,
sizeof (db_table_desc_p), (xdrproc_t)xdr_db_table_desc_p))
return (FALSE);
if (!xdr_int(xdrs, &objp->count))
return (FALSE);
if (xdrs->x_op == XDR_DECODE) {
/* If no version was found, set version to 0. */
if (!xdr_vers(xdrs, (void**) &db_update_version))
make_zero(&db_update_version);
return (TRUE);
} else if (xdrs->x_op == XDR_ENCODE) {
/* Always write out version */
if (!xdr_vers(xdrs, (void**) &db_update_version))
return (FALSE);
} /* else XDR_FREE: do nothing */
return (TRUE);
}
Expr make_zero(Type t) {
if (t.is_handle()) {
return reinterpret(t, make_zero(UInt(64)));
} else {
return make_const(t, 0);
}
}
//==============================================================================
// calculates the exact result, as an expansion.
static void
expansion_orientation_time3d(const double* x0, int time0,
const double* x1, int time1,
const double* x2, int time2,
const double* x3, int time3,
const double* x4, int time4,
expansion& result)
{
make_zero(result);
expansion d;
if(time0){
expansion_orientation3d(x1, x2, x3, x4, d);
negative(d, result);
}
if(time1){
expansion_orientation3d(x0, x2, x3, x4, d);
add(result, d, result);
}
if(time2){
expansion_orientation3d(x0, x1, x3, x4, d);
subtract(result, d, result);
}
if(time3){
expansion_orientation3d(x0, x1, x2, x4, d);
add(result, d, result);
}
if(time3){
expansion_orientation3d(x0, x1, x2, x3, d);
subtract(result, d, result);
}
}
ieee_floatt &ieee_floatt::operator /= (const ieee_floatt &other)
{
assert(other.spec.f==spec.f);
// NaN/x = NaN
if(NaN) return *this;
// x/Nan = NaN
if(other.NaN) { make_NaN(); return *this; }
// 0/0 = NaN
if(is_zero() && other.is_zero()) { make_NaN(); return *this; }
// x/0 = +-inf
if(other.is_zero())
{
infinity=true;
if(other.sign) negate();
return *this;
}
// x/inf = NaN
if(other.infinity)
{
if(infinity) { make_NaN(); return *this; }
bool old_sign = sign;
make_zero();
sign = old_sign;
if(other.sign)
negate();
return *this;
} // inf/x = inf
else if(infinity)
{
if(other.sign) negate();
return *this;
}
exponent-=other.exponent;
fraction*=power(2, spec.f);
// to account for error
fraction*=power(2, spec.f);
exponent-=spec.f;
fraction/=other.fraction;
if(other.sign) negate();
align();
return *this;
}
void visit(const Variable *op) {
if (op->name == var) {
expr = make_one(op->type);
} else if (scope.contains(op->name)) {
expr = scope.get(op->name);
} else {
expr = make_zero(op->type);
}
}
int main()
{
char a[SIZE], b[SIZE], c[SIZE];
make_num_from_str("12",a);
make_num_from_str("12",b);
make_zero(c);
//mul_simple(a,2,c);
mul_simple(a,7,b);
print_big_num(b);
return 0;
}
int make_zero(long long int no, long long int* result){
if ((no == 0)||(no == 1) || (no % 2 != 0))
return 0;
if ( (no % 2 ) == 0){
result[no/2] = 0;
make_zero(no/2, result );
}
}
Expr halide_exp(Expr x_full) {
Type type = x_full.type();
internal_assert(type.element_of() == Float(32));
float ln2_part1 = 0.6931457519f;
float ln2_part2 = 1.4286067653e-6f;
float one_over_ln2 = 1.0f/logf(2.0f);
Expr scaled = x_full * one_over_ln2;
Expr k_real = floor(scaled);
Expr k = cast(Int(32, type.lanes()), k_real);
Expr x = x_full - k_real * ln2_part1;
x -= k_real * ln2_part2;
float coeff[] = {
0.00031965933071842413f,
0.00119156835564003744f,
0.00848988645943932717f,
0.04160188091348320655f,
0.16667983794100929562f,
0.49999899033463041098f,
1.0f,
1.0f
};
Expr result = evaluate_polynomial(x, coeff, sizeof(coeff)/sizeof(coeff[0]));
// Compute 2^k.
int fpbias = 127;
Expr biased = k + fpbias;
Expr inf = Call::make(type, "inf_f32", {}, Call::PureExtern);
// Shift the bits up into the exponent field and reinterpret this
// thing as float.
Expr two_to_the_n = reinterpret(type, biased << 23);
result *= two_to_the_n;
// Catch overflow and underflow
result = select(biased < 255, result, inf);
result = select(biased > 0, result, make_zero(type));
// This introduces lots of common subexpressions
result = common_subexpression_elimination(result);
return result;
}
void mul_big_num(char num1[SIZE], char num2[SIZE], char res[SIZE])
{
int i = 0, times;
char temp[SIZE];
make_zero(temp);
while(num2[i] == '0')
{
i += 1;
}
times = num2[i] - '0';
mul_simple(num1,times,temp);
while(i <= SIZE - 1)
{
add_big_num(temp, res, res);
i += 1;
times = num2[i] - '0';
mul_simple(num1,times,temp);
shift_left(temp);
}
add_big_num(temp,res,res);
}
Expr const_false(int w) {
return make_zero(UInt(1, w));
}
Expr BufferBuilder::build() const {
std::vector<Expr> args(10);
if (buffer_memory.defined()) {
args[0] = buffer_memory;
} else {
args[0] = Call::make(type_of<struct halide_buffer_t *>(), Call::alloca, {(int)sizeof(halide_buffer_t)}, Call::Intrinsic);
}
std::string shape_var_name = unique_name('t');
Expr shape_var = Variable::make(type_of<halide_dimension_t *>(), shape_var_name);
if (shape_memory.defined()) {
args[1] = shape_memory;
} else if (dimensions == 0) {
args[1] = make_zero(type_of<halide_dimension_t *>());
} else {
args[1] = shape_var;
}
if (host.defined()) {
args[2] = host;
} else {
args[2] = make_zero(type_of<void *>());
}
if (device.defined()) {
args[3] = device;
} else {
args[3] = make_zero(UInt(64));
}
if (device_interface.defined()) {
args[4] = device_interface;
} else {
args[4] = make_zero(type_of<struct halide_device_interface_t *>());
}
args[5] = (int)type.code();
args[6] = type.bits();
args[7] = dimensions;
std::vector<Expr> shape;
for (size_t i = 0; i < (size_t)dimensions; i++) {
if (i < mins.size()) {
shape.push_back(mins[i]);
} else {
shape.push_back(0);
}
if (i < extents.size()) {
shape.push_back(extents[i]);
} else {
shape.push_back(0);
}
if (i < strides.size()) {
shape.push_back(strides[i]);
} else {
shape.push_back(0);
}
// per-dimension flags, currently unused.
shape.push_back(0);
}
for (const Expr &e : shape) {
internal_assert(e.type() == Int(32))
<< "Buffer shape fields must be int32_t:" << e << "\n";
}
Expr shape_arg = Call::make(type_of<halide_dimension_t *>(), Call::make_struct, shape, Call::Intrinsic);
if (shape_memory.defined()) {
args[8] = shape_arg;
} else if (dimensions == 0) {
args[8] = make_zero(type_of<halide_dimension_t *>());
} else {
args[8] = shape_var;
}
Expr flags = make_zero(UInt(64));
if (host_dirty.defined()) {
flags = select(host_dirty,
make_const(UInt(64), halide_buffer_flag_host_dirty),
make_zero(UInt(64)));
}
if (device_dirty.defined()) {
flags = flags | select(device_dirty,
make_const(UInt(64), halide_buffer_flag_device_dirty),
make_zero(UInt(64)));
}
args[9] = flags;
Expr e = Call::make(type_of<struct halide_buffer_t *>(), Call::buffer_init, args, Call::Extern);
if (!shape_memory.defined() && dimensions != 0) {
e = Let::make(shape_var_name, shape_arg, e);
}
return e;
}
Expr lower_lerp(Expr zero_val, Expr one_val, Expr weight) {
Expr result;
internal_assert(zero_val.type() == one_val.type());
internal_assert(weight.type().is_uint() || weight.type().is_float());
Type result_type = zero_val.type();
Expr bias_value = make_zero(result_type);
Type computation_type = result_type;
if (zero_val.type().is_int()) {
computation_type = UInt(zero_val.type().bits(), zero_val.type().lanes());
bias_value = result_type.min();
}
// For signed integer types, just convert everything to unsigned
// and then back at the end to ensure proper rounding, etc.
// There is likely a better way to handle this.
if (result_type != computation_type) {
zero_val = Cast::make(computation_type, zero_val - bias_value);
one_val = Cast::make(computation_type, one_val - bias_value);
}
if (result_type.is_bool()) {
Expr half_weight;
if (weight.type().is_float())
half_weight = 0.5f;
else {
half_weight = weight.type().max() / 2;
}
result = select(weight > half_weight, one_val, zero_val);
} else {
Expr typed_weight;
Expr inverse_typed_weight;
if (weight.type().is_float()) {
typed_weight = weight;
if (computation_type.is_uint()) {
// TODO: Verify this reduces to efficient code or
// figure out a better way to express a multiply
// of unsigned 2^32-1 by a double promoted weight
if (computation_type.bits() == 32) {
typed_weight =
Cast::make(computation_type,
cast<double>(Expr(65535.0f)) * cast<double>(Expr(65537.0f)) *
Cast::make(Float(64, typed_weight.type().lanes()), typed_weight));
} else {
typed_weight =
Cast::make(computation_type,
computation_type.max() * typed_weight);
}
inverse_typed_weight = computation_type.max() - typed_weight;
} else {
inverse_typed_weight = 1.0f - typed_weight;
}
} else {
if (computation_type.is_float()) {
int weight_bits = weight.type().bits();
if (weight_bits == 32) {
// Should use ldexp, but can't make Expr from result
// that is double
typed_weight =
Cast::make(computation_type,
cast<double>(weight) / (pow(cast<double>(2), 32) - 1));
} else {
typed_weight =
Cast::make(computation_type,
weight / ((float)ldexp(1.0f, weight_bits) - 1));
}
inverse_typed_weight = 1.0f - typed_weight;
} else {
// This code rescales integer weights to the right number of bits.
// It takes advantage of (2^n - 1) == (2^(n/2) - 1)(2^(n/2) + 1)
// e.g. 65535 = 255 * 257. (Ditto for the 32-bit equivalent.)
// To recale a weight of m bits to be n bits, we need to do:
// scaled_weight = (weight / (2^m - 1)) * (2^n - 1)
// which power of two values for m and n, results in a series like
// so:
// (2^(m/2) + 1) * (2^(m/4) + 1) ... (2^(n*2) + 1)
// The loop below computes a scaling constant and either multiples
// or divides by the constant and relies on lowering and llvm to
// generate efficient code for the operation.
int bit_size_difference = weight.type().bits() - computation_type.bits();
if (bit_size_difference == 0) {
typed_weight = weight;
} else {
typed_weight = Cast::make(computation_type, weight);
int bits_left = ::abs(bit_size_difference);
int shift_amount = std::min(computation_type.bits(), weight.type().bits());
uint64_t scaling_factor = 1;
while (bits_left != 0) {
internal_assert(bits_left > 0);
scaling_factor = scaling_factor + (scaling_factor << shift_amount);
bits_left -= shift_amount;
shift_amount *= 2;
}
if (bit_size_difference < 0) {
typed_weight =
Cast::make(computation_type, weight) *
cast(computation_type, (int32_t)scaling_factor);
} else {
typed_weight =
Cast::make(computation_type,
weight / cast(weight.type(), (int32_t)scaling_factor));
}
}
inverse_typed_weight =
Cast::make(computation_type,
computation_type.max() - typed_weight);
}
}
if (computation_type.is_float()) {
result = zero_val * inverse_typed_weight +
one_val * typed_weight;
} else {
int32_t bits = computation_type.bits();
switch (bits) {
case 1:
result = select(typed_weight, one_val, zero_val);
break;
case 8:
case 16:
case 32: {
Expr zero_expand = Cast::make(UInt(2 * bits, computation_type.lanes()),
zero_val);
Expr one_expand = Cast::make(UInt(2 * bits, one_val.type().lanes()),
one_val);
Expr rounding = Cast::make(UInt(2 * bits), 1) << Cast::make(UInt(2 * bits), (bits - 1));
Expr divisor = Cast::make(UInt(2 * bits), 1) << Cast::make(UInt(2 * bits), bits);
Expr prod_sum = zero_expand * inverse_typed_weight +
one_expand * typed_weight + rounding;
Expr divided = ((prod_sum / divisor) + prod_sum) / divisor;
result = Cast::make(UInt(bits, computation_type.lanes()), divided);
break;
}
case 64:
// TODO: 64-bit lerp is not supported as current approach
// requires double-width multiply.
// There is an informative error message in IROperator.h.
internal_error << "Can't do a 64-bit lerp.\n";
break;
default:
break;
}
}
if (!is_zero(bias_value)) {
result = Cast::make(result_type, result) + bias_value;
}
}
return simplify(result);
}
