Skip to content

Commit

Permalink
Fix bilinear scale
Browse files Browse the repository at this point in the history
  • Loading branch information
@awxkee
awxkee committed Jan 7, 2024
1 parent c9190a4 commit ab15738
Show file tree
Hide file tree
Showing 2 changed files with 49 additions and 48 deletions.
2 changes: 1 addition & 1 deletion jxlcoder/build.gradle.kts
View file
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ afterEvaluate {
create<MavenPublication>("mavenJava") {
groupId = "com.github.awxkee"
artifactId = "jxl-coder"
version = "1.6.0"
version = "1.6.1"
from(components["release"])
// artifact(androidSourcesJar)
}
Expand Down
95 changes: 48 additions & 47 deletions jxlcoder/src/main/cpp/XScaler.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -56,7 +56,7 @@ namespace coder::HWY_NAMESPACE {
using hwy::HWY_NAMESPACE::BitCast;
using hwy::HWY_NAMESPACE::Min;
using hwy::HWY_NAMESPACE::ConvertTo;
using hwy::HWY_NAMESPACE::Ceil;
using hwy::HWY_NAMESPACE::Floor;
using hwy::HWY_NAMESPACE::Mul;
using hwy::HWY_NAMESPACE::ExtractLane;
using hwy::HWY_NAMESPACE::Zero;
Expand Down Expand Up @@ -296,10 +296,10 @@ namespace coder::HWY_NAMESPACE {
return T(0);
}
T s1 = x * (-p0 + 3 * p1 - 3 * p2 + p3);
T s2 = (T(2.0) * p0 - T(5.0) * p1 + 4 * p2 - p3);
T s3 = (s1 + s2) * x;
T s2 = (T(2.0)*p0 - T(5.0)*p1 + 4 * p2 - p3);
T s3 = (s1 + s2)*x;
T s4 = (-p0 + p2);
T s5 = (s3 + s4) * x * T(0.5);
T s5 = (s3 + s4)*x*T(0.5);

return s5 + p1;
// if (x < T(1.0)) {
Expand Down Expand Up @@ -330,6 +330,25 @@ namespace coder::HWY_NAMESPACE {
auto dst8 = reinterpret_cast<uint8_t *>(output) + y * dstStride;
auto dst16 = reinterpret_cast<uint16_t *>(dst8);

const FixedTag<float32_t, 4> dfx4;
const FixedTag<int32_t, 4> dix4;
const FixedTag<float16_t, 4> df16x4;
using VI4 = Vec<decltype(dix4)>;
using VF4 = Vec<decltype(dfx4)>;
using VF16x4 = Vec<decltype(df16x4)>;

const int shift[4] = {0, 1, 2, 3};
const VI4 shiftV = LoadU(dix4, shift);
const VF4 xScaleV = Set(dfx4, xScale);
const VF4 yScaleV = Set(dfx4, yScale);
const VI4 addOne = Set(dix4, 1);
const VF4 fOneV = Set(dfx4, 1.0f);
const VI4 maxWidth = Set(dix4, inputWidth - 1);
const VI4 maxHeight = Set(dix4, inputHeight - 1);
const VI4 iZeros = Zero(dix4);
const VF4 vfZeros = Zero(dfx4);
const VI4 srcStrideV = Set(dix4, srcStride);

for (int x = 0; x < outputWidth; ++x) {
float srcX = (float) x * xScale;
float srcY = (float) y * yScale;
Expand All @@ -339,31 +358,13 @@ namespace coder::HWY_NAMESPACE {

if (option == bilinear) {
if (components == 4 && x + 8 < outputWidth) {
FixedTag<float32_t, 4> dfx4;
FixedTag<int32_t, 4> dix4;
FixedTag<float16_t, 4> df16x4;
using VI4 = Vec<decltype(dix4)>;
using VF4 = Vec<decltype(dfx4)>;
using VF16x4 = Vec<decltype(df16x4)>;

int shift[4] = {0, 1, 2, 3};
VI4 shiftV = LoadU(dix4, &shift[0]);
VF4 xScaleV = Set(dfx4, xScale);
VF4 yScaleV = Set(dfx4, yScale);
VI4 addOne = Set(dix4, 1);
VF4 fOneV = Set(dfx4, 1.0f);
VI4 maxWidth = Set(dix4, inputWidth - 1);
VI4 maxHeight = Set(dix4, inputHeight - 1);
VI4 iZeros = Zero(dix4);
VF4 vfZeros = Zero(dfx4);
VI4 srcStrideV = Set(dix4, srcStride);
VI4 currentX = Set(dix4, x);
VI4 currentXV = Add(currentX, shiftV);
VF4 currentXVF = Mul(ConvertTo(dfx4, currentXV), xScaleV);
VF4 currentYVF = Mul(ConvertTo(dfx4, Set(dix4, y)), yScaleV);

VI4 xi1 = ConvertTo(dix4, Ceil(currentXVF));
VI4 yi1 = ConvertTo(dix4, Ceil(currentYVF));
VI4 xi1 = ConvertTo(dix4, Floor(currentXVF));
VI4 yi1 = Min(ConvertTo(dix4, Floor(currentYVF)), maxHeight);

VI4 xi2 = Min(Add(xi1, addOne), maxWidth);
VI4 yi2 = Min(Add(yi1, addOne), maxHeight);
Expand Down Expand Up @@ -601,6 +602,27 @@ namespace coder::HWY_NAMESPACE {
auto dst8 = reinterpret_cast<uint8_t *>(output + y * dstStride);
auto dst = reinterpret_cast<uint8_t *>(dst8);

const FixedTag<float32_t, 4> dfx4;
const FixedTag<int32_t, 4> dix4;
const FixedTag<uint32_t, 4> dux4;
const FixedTag<uint8_t, 4> du8x4;
using VI4 = Vec<decltype(dix4)>;
using VF4 = Vec<decltype(dfx4)>;
using VU8x4 = Vec<decltype(du8x4)>;

const int shift[4] = {0, 1, 2, 3};
const VI4 shiftV = LoadU(dix4, shift);
const VF4 xScaleV = Set(dfx4, xScale);
const VF4 yScaleV = Set(dfx4, yScale);
const VI4 addOne = Set(dix4, 1);
const VF4 fOneV = Set(dfx4, 1.0f);
const VI4 maxWidth = Set(dix4, inputWidth - 1);
const VI4 maxHeight = Set(dix4, inputHeight - 1);
const VI4 iZeros = Zero(dix4);
const VF4 vfZeros = Zero(dfx4);
const VI4 srcStrideV = Set(dix4, srcStride);
const VF4 maxColorsV = Set(dfx4, maxColors);

for (int x = 0; x < outputWidth; ++x) {
float srcX = (float) x * xScale;
float srcY = (float) y * yScale;
Expand All @@ -610,34 +632,13 @@ namespace coder::HWY_NAMESPACE {

if (option == bilinear) {
if (components == 4 && x + 8 < outputWidth) {
FixedTag<float32_t, 4> dfx4;
FixedTag<int32_t, 4> dix4;
FixedTag<uint32_t, 4> dux4;
FixedTag<uint8_t, 4> du8x4;
using VI4 = Vec<decltype(dix4)>;
using VF4 = Vec<decltype(dfx4)>;
using VU8x4 = Vec<decltype(du8x4)>;

int shift[4] = {0, 1, 2, 3};
VI4 shiftV = LoadU(dix4, &shift[0]);
VF4 xScaleV = Set(dfx4, xScale);
VF4 yScaleV = Set(dfx4, yScale);
VI4 addOne = Set(dix4, 1);
VF4 fOneV = Set(dfx4, 1.0f);
VI4 maxWidth = Set(dix4, inputWidth - 1);
VI4 maxHeight = Set(dix4, inputHeight - 1);
VI4 iZeros = Zero(dix4);
VF4 vfZeros = Zero(dfx4);
VI4 srcStrideV = Set(dix4, srcStride);
VF4 maxColorsV = Set(dfx4, maxColors);

VI4 currentX = Set(dix4, x);
VI4 currentXV = Add(currentX, shiftV);
VF4 currentXVF = Mul(ConvertTo(dfx4, currentXV), xScaleV);
VF4 currentYVF = Mul(ConvertTo(dfx4, Set(dix4, y)), yScaleV);

VI4 xi1 = ConvertTo(dix4, Ceil(currentXVF));
VI4 yi1 = ConvertTo(dix4, Ceil(currentYVF));
VI4 xi1 = ConvertTo(dix4, Floor(currentXVF));
VI4 yi1 = Min(ConvertTo(dix4, Floor(currentYVF)), maxHeight);

VI4 xi2 = Min(Add(xi1, addOne), maxWidth);
VI4 yi2 = Min(Add(yi1, addOne), maxHeight);
Expand Down

0 comments on commit ab15738

Please sign in to comment.