added sve

CMakeLists.txt

@@ -5,7 +5,9 @@ project(yuv)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_STANDARD 20)
 
-add_definitions(-DHWY_DISABLED_TARGETS=\(HWY_SVE|HWY_SVE2|HWY_SVE_256|HWY_SVE2_128|HWY_SCALAR\))
+add_definitions(-DHWY_DISABLED_TARGETS=\(HWY_SCALAR\))
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=armv8-a+simd+crypto+sve")
+#set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv8-a+simd+crypto+sve")
 
 include(CheckCXXCompilerFlag)
 check_cxx_compiler_flag("-pthread" COMPILER_SUPPORTS_THREADS)

Dockerfile

@@ -1,4 +1,4 @@
-FROM --platform=linux/ppc64le ubuntu:latest
+FROM ubuntu:latest
 
 RUN apt-get -y update && apt-get -y install build-essential ninja-build cmake
 

src/BlendAlpha-inl.h

@@ -161,11 +161,12 @@ UnpremultiplyAlpha8HWY(const uint8_t *SPARKYUV_RESTRICT src, const uint32_t srcS
       auto Ah = PromoteUpperTo(du16, A);
       auto Al = PromoteLowerTo(du16, A);
 
+#if HWY_TARGET != HWY_SVE && HWY_TARGET != HWY_SVE2 && HWY_TARGET != HWY_SVE_256 && HWY_TARGET != HWY_SVE2_128
       const auto ahMask = Ah == zeros;
       const auto alMask = Ah == zeros;
-
       Ah = IfThenElse(ahMask, mScale, Ah);
       Al = IfThenElse(alMask, mScale, Al);
+#endif
 
       const auto Rh = Div(WidenMulHigh(du8, R, scale), Ah);
       const auto Rl = Div(WidenMul(dhu8, LowerHalf(R), lScale), Al);

src/ITUR/eotf-inl.h

@@ -110,7 +110,7 @@ static float HLGOetf(float linear) {
   } else if (linear <= (1.0f / 12.0f)) {
     return ::sqrtf(3.0f * linear);
   } else {
-    return 0.17883277f * logf(12.0f * linear - 0.28466892f) + 0.55991073f;
+    return 0.17883277f * ::logf(12.0f * linear - 0.28466892f) + 0.55991073f;
   }
 }
 

src/YcCbcCrc-inl.h

@@ -26,6 +26,12 @@
 #include "sparkyuv-internal.h"
 #include "Eotf-inl.h"
 
+#if HWY_TARGET != HWY_SVE && HWY_TARGET != HWY_SVE2 && HWY_TARGET != HWY_SVE_256 && HWY_TARGET != HWY_SVE2_128
+#define YCCBCCRC_HWY_ENABLED 1
+#else
+#define YCCBCCRC_HWY_ENABLED 0
+#endif
+
 HWY_BEFORE_NAMESPACE();
 namespace sparkyuv::HWY_NAMESPACE {
 
@@ -42,13 +48,15 @@ void computeYcCbcCrcCutoffs(const SparkYuvTransferFunction transferFunction,
   Pr = 1.f - Oetf(kr, transferFunction);
 }
 
+#if YCCBCCRC_HWY_ENABLED
 template<typename D, typename V = Vec<D>, HWY_IF_FLOAT_D(D)>
 EOTF_INLINE V computeYcCbcCrcEquation(D d, V dx, V low, V high, V scaleLow, V scaleHigh) {
   const auto zeros = Zero(d);
   auto v = IfThenElse(And(dx >= low, dx <= zeros), Mul(dx, scaleLow), zeros);
   v = IfThenElse(And(dx > zeros, dx < high), Mul(dx, scaleHigh), v);
   return v;
 }
+#endif
 
 EOTF_INLINE float computeYcCbcCrcEquation(float dx, float low, float high, float scaleLow, float scaleHigh) {
   float Eb;
@@ -113,17 +121,17 @@ PixelToYcCbcCrcHWY(const T *SPARKYUV_RESTRICT src, const uint32_t srcStride,
   const float divisor2 = 1.f / 2.f;
   const float divisor4 = 1.f / 4.f;
 
+#if YCCBCCRC_HWY_ENABLED
   const ScalableTag<uint16_t> du16;
   const Half<decltype(du16)> dhu16;
   const Rebind<T, decltype(du16)> d;
   const Half<decltype(d)> dh;
-  const int lanes = Lanes(d);
-  const int uvLanes = (chromaSubsample == YUV_SAMPLE_444) ? lanes : Lanes(dh);
   const Rebind<uint32_t, decltype(dhu16)> du32;
   const Rebind<float, decltype(du32)> df;
   using VF = Vec<decltype(df)>;
   using VU = Vec<decltype(du16)>;
-
+  const int lanes = Lanes(d);
+  const int uvLanes = (chromaSubsample == YUV_SAMPLE_444) ? lanes : Lanes(dh);
   const auto vScaleRangeY = Set(df, rangeY);
   const auto vScaleRangeUV = Set(df, rangeUV);
   const auto vBiasY = Set(df, biasY);
@@ -142,6 +150,7 @@ PixelToYcCbcCrcHWY(const T *SPARKYUV_RESTRICT src, const uint32_t srcStride,
   const auto vKg = Set(df, kg);
   const auto alpha = Set(d, maxColors);
   const auto v420Scale = Set(df, 0.5f);
+#endif
 
   for (uint32_t y = 0; y < height; ++y) {
     uint32_t x = 0;
@@ -151,6 +160,7 @@ PixelToYcCbcCrcHWY(const T *SPARKYUV_RESTRICT src, const uint32_t srcStride,
     auto vDst = reinterpret_cast<T *>(vStore);
     auto mSrc = reinterpret_cast<const T *>(mSource);
 
+#if YCCBCCRC_HWY_ENABLED
     for (; x + lanes < width; x += lanes) {
       VF Rh, Rl, Gh, Gl, Bh, Bl;
       if (std::is_same<T, uint16_t>::value) {
@@ -317,6 +327,7 @@ PixelToYcCbcCrcHWY(const T *SPARKYUV_RESTRICT src, const uint32_t srcStride,
       vDst += uvLanes;
       mSrc += lanes*components;
     }
+#endif
 
     for (; x < width; x += lanesForward) {
       float r;
@@ -570,6 +581,7 @@ EOTF_INLINE T YcCbcCrcInverse(T Y, T v, T scaleLow, T scaleHigh, T low, T high)
   return x;
 }
 
+#if YCCBCCRC_HWY_ENABLED
 template<class D, class V = Vec<D>>
 EOTF_INLINE V YcCbcCrcInverse(D d, V Y, V v, V scaleLow, V scaleHigh, V low, V high) {
   const V l = Mul(v, scaleLow);
@@ -581,6 +593,7 @@ EOTF_INLINE V YcCbcCrcInverse(D d, V Y, V v, V scaleLow, V scaleHigh, V low, V h
   cv = IfThenElse(And(diffh > zeros, diffh <= high), Add(h, Y), cv);
   return cv;
 }
+#endif
 
 /**
 * @brief It will be good to declare a type of transfer function
@@ -632,6 +645,7 @@ void YcCbcCrcToXRGB(T *SPARKYUV_RESTRICT rgbaData, const uint32_t dstStride,
 
   const float ekg = 1.f / kg;
 
+#if YCCBCCRC_HWY_ENABLED
   const ScalableTag<uint16_t> du16;
   const Half<decltype(du16)> dhu16;
   const Rebind<T, decltype(du16)> d;
@@ -690,6 +704,7 @@ void YcCbcCrcToXRGB(T *SPARKYUV_RESTRICT rgbaData, const uint32_t dstStride,
 #endif
 
   const auto A = Set(du16, maxColors);
+#endif
 
   for (int y = 0; y < height; ++y) {
     auto CbSource = reinterpret_cast<const T *>(mUSrc);
@@ -699,6 +714,7 @@ void YcCbcCrcToXRGB(T *SPARKYUV_RESTRICT rgbaData, const uint32_t dstStride,
 
     uint32_t x = 0;
 
+#if YCCBCCRC_HWY_ENABLED
     for (; x + lanes < width; x += lanes) {
 #if SPARKYUV_ALLOW_FLOAT16
       VF Y, Cb, Cr;
@@ -835,6 +851,7 @@ void YcCbcCrcToXRGB(T *SPARKYUV_RESTRICT rgbaData, const uint32_t dstStride,
       CrSource += uvLanes;
       store += lanes * components;
     }
+#endif
 
     for (; x < width; x += lanesForward) {
       const T uValue = reinterpret_cast<const T *>(CbSource)[0];

src/sampler/BilinearRowSampler-inl.hpp

@@ -23,13 +23,22 @@
 
 #include <hwy/highway.h>
 #include "ScaleRowSampler.hpp"
-#include "sampler.h"
-#include "sampler-inl.h"
 #include "../yuv-inl.h"
+#include "sampler.h"
 #include <cstdint>
 #include <algorithm>
 #include <cmath>
 
+#if HWY_TARGET != HWY_SVE && HWY_TARGET != HWY_SVE2 && HWY_TARGET != HWY_SVE_256 && HWY_TARGET != HWY_SVE2_128
+#define BILINEAR_ENABLE_HWY 1
+#else
+#define BILINEAR_ENABLE_HWY 0
+#endif
+
+#if BILINEAR_ENABLE_HWY
+#include "sampler-inl.h"
+#endif
+
 HWY_BEFORE_NAMESPACE();
 namespace sparkyuv::HWY_NAMESPACE {
 using namespace sparkyuv;
@@ -61,6 +70,7 @@ class BilinearRowSampler4Chan8Bit : public ScaleRowSampler<uint8_t> {
   ~BilinearRowSampler4Chan8Bit() override = default;
 
   void sample(const int row) override {
+#if BILINEAR_ENABLE_HWY
     const FixedTag<float32_t, 4> dfx4;
     const FixedTag<uint32_t, 4> dix4;
     const FixedTag<uint8_t, 4> du8x4;
@@ -82,7 +92,7 @@ class BilinearRowSampler4Chan8Bit : public ScaleRowSampler<uint8_t> {
     const VF4 vfZeros = Zero(dfx4);
     const VI4 srcStrideV = Set(dix4, this->srcStride);
     const VF4 maxColorsV = Set(dfx4, maxColors);
-
+#endif
     auto dst8 = reinterpret_cast<uint8_t *>(reinterpret_cast<uint8_t *>(this->mDestination) + row * this->dstStride);
     auto dst = reinterpret_cast<uint8_t *>(dst8);
 
@@ -91,6 +101,7 @@ class BilinearRowSampler4Chan8Bit : public ScaleRowSampler<uint8_t> {
 
     uint32_t x = 0;
 
+#if BILINEAR_ENABLE_HWY
 #if !NOACCELERATED_SAMPLER
     for (; x + 8 < this->outputWidth && components == 4; ++x) {
       VI4 currentX = Set(dix4, x);
@@ -135,6 +146,7 @@ class BilinearRowSampler4Chan8Bit : public ScaleRowSampler<uint8_t> {
 
       x += components - 1;
     }
+#endif
 #endif
 
     for (; x < this->outputWidth; ++x) {
@@ -196,6 +208,7 @@ class BilinearRowSamplerF16Bit : public ScaleRowSampler<uint16_t> {
   ~BilinearRowSamplerF16Bit() override = default;
 
   void sample(const int y) override {
+#if BILINEAR_ENABLE_HWY
     const FixedTag<float32_t, 4> dfx4;
     const FixedTag<int32_t, 4> dix4;
     const FixedTag<hwy::float16_t, 4> df16x4;
@@ -213,6 +226,7 @@ class BilinearRowSamplerF16Bit : public ScaleRowSampler<uint16_t> {
     const VI4 maxHeight = Set(dix4, this->inputHeight - 1);
     const VF4 vfZeros = Zero(dfx4);
     const VI4 srcStrideV = Set(dix4, this->srcStride);
+#endif
 
     const auto src8 = reinterpret_cast<const uint8_t *>(this->mSource);
     auto dst16 = reinterpret_cast<uint16_t *>(reinterpret_cast<uint8_t *>(this->mDestination) + y * this->dstStride);
@@ -221,6 +235,7 @@ class BilinearRowSamplerF16Bit : public ScaleRowSampler<uint16_t> {
 
     uint32_t x = 0;
 
+#if BILINEAR_ENABLE_HWY
 #if !NOACCELERATED_SAMPLER
     for (; x + 8 < this->outputWidth && components == 4; ++x) {
       VI4 currentX = Set(dix4, x);
@@ -263,6 +278,7 @@ class BilinearRowSamplerF16Bit : public ScaleRowSampler<uint16_t> {
 
       x += components - 1;
     }
+#endif
 #endif
 
     for (; x < this->outputWidth; ++x) {
@@ -460,4 +476,6 @@ class BilinearRowSampler10Bit : public ScaleRowSampler<uint32_t> {
 } // sparkyuv
 HWY_AFTER_NAMESPACE();
 
+#undef BILINEAR_ENABLE_HWY
+
 #endif //SPARKYUV_BILINEAR_ROW_SAMPLER

src/sampler/NearestRowSampler-inl.hpp

@@ -23,7 +23,6 @@
 
 #include "hwy/highway.h"
 #include "ScaleRowSampler.hpp"
-#include "sampler-inl.h"
 #include "sampler.h"
 #include <cstdint>
 #include <algorithm>