Skip to content

Commit

Permalink
struct converter double precision fix
Browse files Browse the repository at this point in the history
  • Loading branch information
@wjakob
wjakob committed Apr 22, 2020
1 parent 86ec189 commit 13c629c
Showing 1 changed file with 34 additions and 18 deletions.
52 changes: 34 additions & 18 deletions src/libcore/struct.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,12 @@ extern const float dither_matrix256[65536];

NAMESPACE_BEGIN(detail)

#if defined(DOUBLE_PRECISION)
# define Float double
#else
# define Float float
#endif

#if MTS_STRUCTCONVERTER_USE_JIT == 1

using namespace asmjit;
Expand Down Expand Up @@ -84,8 +90,7 @@ class StructCompiler {
}

/// Floating point move operation
template <typename X, typename Y>
void movs(const X &x, const Y &y) {
template <typename X, typename Y> void movs(const X &x, const Y &y) {
#if defined(ENOKI_X86_AVX)
#if !defined(DOUBLE_PRECISION)
cc.vmovss(x, y);
Expand Down Expand Up @@ -615,7 +620,7 @@ class StructCompiler {
}

std::pair<Key, Value> load_default(const Struct::Field &field) {
Key key { field.name, struct_type_v<float>, +Struct::Flags::None };
Key key { field.name, struct_type_v<Float>, +Struct::Flags::None };
Value value;
value.xmm = cc.newXmm();
movs(value.xmm, const_(field.default_));
Expand All @@ -632,13 +637,13 @@ class StructCompiler {
bool inv_gamma = false;

if (Struct::is_integer(kr.type)) {
kr.type = struct_type_v<float>;
kr.type = struct_type_v<Float>;
kr.flags &= ~Struct::Flags::Normalized;
int_to_float = true;
}

if (Struct::is_float(kr.type) && kr.type != struct_type_v<float>) {
kr.type = struct_type_v<float>;
if (Struct::is_float(kr.type) && kr.type != struct_type_v<Float>) {
kr.type = struct_type_v<Float>;
float_to_float = true;
}

Expand Down Expand Up @@ -694,7 +699,7 @@ class StructCompiler {
kr.type = Struct::Type::Float32;
}

if (kr.type == Struct::Type::Float32 && kr.type != struct_type_v<float>) {
if (kr.type == Struct::Type::Float32 && kr.type != struct_type_v<Float>) {
X86Xmm source = vr.xmm;
vr.xmm = cc.newXmm();
#if defined(ENOKI_X86_AVX)
Expand All @@ -705,7 +710,7 @@ class StructCompiler {
kr.type = Struct::Type::Float64;
}

if (kr.type == Struct::Type::Float64 && kr.type != struct_type_v<float>) {
if (kr.type == Struct::Type::Float64 && kr.type != struct_type_v<Float>) {
X86Xmm source = vr.xmm;
vr.xmm = cc.newXmm();
#if defined(ENOKI_X86_AVX)
Expand Down Expand Up @@ -741,8 +746,8 @@ class StructCompiler {

if (field.is_integer() && !Struct::is_integer(key.type)) {
auto range_dbl = field.range();
std::pair<float, float> range = range_dbl;
if (key.type != struct_type_v<float>)
std::pair<Float, Float> range = range_dbl;
if (key.type != struct_type_v<Float>)
Throw("Internal error!");

auto temp = cc.newXmm();
Expand All @@ -765,7 +770,18 @@ class StructCompiler {
X86Gp base = cc.newUInt64();
cc.mov(base.r64(), Imm((uintptr_t) dither_matrix256));
dither_value = cc.newXmm();
movs(dither_value, X86Mem(base, index, 2, 0, (uint32_t) sizeof(float)));
#if defined(ENOKI_X86_AVX)
cc.movss(dither_value, X86Mem(base, index, 2, 0, (uint32_t) sizeof(float)));
#else
cc.vmovss(dither_value, X86Mem(base, index, 2, 0, (uint32_t) sizeof(float)));
#endif
#if defined(DOUBLE_PRECISION)
#if defined(ENOKI_X86_AVX)
cc.vcvtss2sd(dither_value, dither_value, dither_value);
#else
cc.cvtss2sd(dither_value, dither_value);
#endif
#endif
dither_ready = true;
}
adds(value.xmm, dither_value);
Expand Down Expand Up @@ -1297,7 +1313,7 @@ StructConverter::StructConverter(const Struct *source, const Struct *target, boo

if (source_weight != nullptr && target_weight == nullptr) {
X86Xmm result = cc.newXmm();
if (kv.first.type != struct_type_v<float>)
if (kv.first.type != struct_type_v<Float>)
kv = sc.linearize(kv);
sc.muls(result, kv.second.xmm, scale_factor);
kv.second.xmm = result;
Expand Down Expand Up @@ -1466,7 +1482,7 @@ void StructConverter::linearize(Value &value) const {

if (has_flag(value.flags, Struct::Flags::Normalized))
value.f *= Float(1 / Struct::range(value.type).second);
} else if (Struct::is_float(value.type) && value.type != struct_type_v<float>) {
} else if (Struct::is_float(value.type) && value.type != struct_type_v<Float>) {
if (value.type == Struct::Type::Float32)
value.f = (Float) value.s;
else
Expand All @@ -1477,7 +1493,7 @@ void StructConverter::linearize(Value &value) const {
value.flags = value.flags & ~Struct::Flags::Gamma;
}

value.type = struct_type_v<float>;
value.type = struct_type_v<Float>;
}

void StructConverter::save(uint8_t *dst, const Struct::Field &f, Value value, size_t x, size_t y) const {
Expand All @@ -1491,7 +1507,7 @@ void StructConverter::save(uint8_t *dst, const Struct::Field &f, Value value, si
if (has_flag(f.flags, Struct::Flags::Gamma) && !has_flag(value.flags, Struct::Flags::Gamma))
value.f = enoki::linear_to_srgb(value.f);

if (f.is_integer() && value.type == struct_type_v<float>) {
if (f.is_integer() && value.type == struct_type_v<Float>) {
auto range = f.range();

if (has_flag(f.flags, Struct::Flags::Normalized))
Expand All @@ -1512,9 +1528,9 @@ void StructConverter::save(uint8_t *dst, const Struct::Field &f, Value value, si
value.u = (uint64_t) d;
}

if ((f.type == Struct::Type::Float16 || f.type == Struct::Type::Float32) && value.type == struct_type_v<float>)
if ((f.type == Struct::Type::Float16 || f.type == Struct::Type::Float32) && value.type == struct_type_v<Float>)
value.s = (float) value.f;
else if (f.type == Struct::Type::Float64 && value.type == struct_type_v<float>)
else if (f.type == Struct::Type::Float64 && value.type == struct_type_v<Float>)
value.d = (double) value.f;

switch (f.type) {
Expand Down Expand Up @@ -1658,7 +1674,7 @@ bool StructConverter::convert_2d(size_t width, size_t height, const void *src_,
return false;
}
} else {
value.type = struct_type_v<float>;
value.type = struct_type_v<Float>;
value.f = 0;
value.flags = +Struct::Flags::None;
for (auto kv : f.blend) {
Expand Down

0 comments on commit 13c629c

Please sign in to comment.