skia2/bench/MathBench.cpp
tomhudson@google.com 9dc2713fc4 Let SkBenchmark classes specify that they do no rendering.
Doing this gives us a 15-20% speedup in bench cycle time.
Here again I'm just picking the easy targets.

http://codereview.appspot.com/6500115/



git-svn-id: http://skia.googlecode.com/svn/trunk@5525 2bbb7eff-a529-9590-31e7-b0007b416f81
2012-09-13 15:50:24 +00:00

409 lines
11 KiB
C++

#include "SkBenchmark.h"
#include "SkColorPriv.h"
#include "SkMatrix.h"
#include "SkRandom.h"
#include "SkString.h"
#include "SkPaint.h"
static float sk_fsel(float pred, float result_ge, float result_lt) {
return pred >= 0 ? result_ge : result_lt;
}
static float fast_floor(float x) {
// float big = sk_fsel(x, 0x1.0p+23, -0x1.0p+23);
float big = sk_fsel(x, (float)(1 << 23), -(float)(1 << 23));
return (x + big) - big;
}
class MathBench : public SkBenchmark {
enum {
kBuffer = 100,
kLoop = 10000
};
SkString fName;
float fSrc[kBuffer], fDst[kBuffer];
public:
MathBench(void* param, const char name[]) : INHERITED(param) {
fName.printf("math_%s", name);
SkRandom rand;
for (int i = 0; i < kBuffer; ++i) {
fSrc[i] = rand.nextSScalar1();
}
fIsRendering = false;
}
virtual void performTest(float* SK_RESTRICT dst,
const float* SK_RESTRICT src,
int count) = 0;
protected:
virtual int mulLoopCount() const { return 1; }
virtual const char* onGetName() {
return fName.c_str();
}
virtual void onDraw(SkCanvas* canvas) {
int n = SkBENCHLOOP(kLoop * this->mulLoopCount());
for (int i = 0; i < n; i++) {
this->performTest(fDst, fSrc, kBuffer);
}
}
private:
typedef SkBenchmark INHERITED;
};
class MathBenchU32 : public MathBench {
public:
MathBenchU32(void* param, const char name[]) : INHERITED(param, name) {}
protected:
virtual void performITest(uint32_t* SK_RESTRICT dst,
const uint32_t* SK_RESTRICT src,
int count) = 0;
virtual void performTest(float* SK_RESTRICT dst,
const float* SK_RESTRICT src,
int count) SK_OVERRIDE {
uint32_t* d = SkTCast<uint32_t*>(dst);
const uint32_t* s = SkTCast<const uint32_t*>(src);
this->performITest(d, s, count);
}
private:
typedef MathBench INHERITED;
};
///////////////////////////////////////////////////////////////////////////////
class NoOpMathBench : public MathBench {
public:
NoOpMathBench(void* param) : INHERITED(param, "noOp") {}
protected:
virtual void performTest(float* SK_RESTRICT dst,
const float* SK_RESTRICT src,
int count) {
for (int i = 0; i < count; ++i) {
dst[i] = src[i] + 1;
}
}
private:
typedef MathBench INHERITED;
};
class SlowISqrtMathBench : public MathBench {
public:
SlowISqrtMathBench(void* param) : INHERITED(param, "slowIsqrt") {}
protected:
virtual void performTest(float* SK_RESTRICT dst,
const float* SK_RESTRICT src,
int count) {
for (int i = 0; i < count; ++i) {
dst[i] = 1.0f / sk_float_sqrt(src[i]);
}
}
private:
typedef MathBench INHERITED;
};
static inline float SkFastInvSqrt(float x) {
float xhalf = 0.5f*x;
int i = *(int*)&x;
i = 0x5f3759df - (i>>1);
x = *(float*)&i;
x = x*(1.5f-xhalf*x*x);
// x = x*(1.5f-xhalf*x*x); // this line takes err from 10^-3 to 10^-6
return x;
}
class FastISqrtMathBench : public MathBench {
public:
FastISqrtMathBench(void* param) : INHERITED(param, "fastIsqrt") {}
protected:
virtual void performTest(float* SK_RESTRICT dst,
const float* SK_RESTRICT src,
int count) {
for (int i = 0; i < count; ++i) {
dst[i] = SkFastInvSqrt(src[i]);
}
}
private:
typedef MathBench INHERITED;
};
static inline uint32_t QMul64(uint32_t value, U8CPU alpha) {
SkASSERT((uint8_t)alpha == alpha);
const uint32_t mask = 0xFF00FF;
uint64_t tmp = value;
tmp = (tmp & mask) | ((tmp & ~mask) << 24);
tmp *= alpha;
return (uint32_t) (((tmp >> 8) & mask) | ((tmp >> 32) & ~mask));
}
class QMul64Bench : public MathBenchU32 {
public:
QMul64Bench(void* param) : INHERITED(param, "qmul64") {}
protected:
virtual void performITest(uint32_t* SK_RESTRICT dst,
const uint32_t* SK_RESTRICT src,
int count) SK_OVERRIDE {
for (int i = 0; i < count; ++i) {
dst[i] = QMul64(src[i], (uint8_t)i);
}
}
private:
typedef MathBenchU32 INHERITED;
};
class QMul32Bench : public MathBenchU32 {
public:
QMul32Bench(void* param) : INHERITED(param, "qmul32") {}
protected:
virtual void performITest(uint32_t* SK_RESTRICT dst,
const uint32_t* SK_RESTRICT src,
int count) SK_OVERRIDE {
for (int i = 0; i < count; ++i) {
dst[i] = SkAlphaMulQ(src[i], (uint8_t)i);
}
}
private:
typedef MathBenchU32 INHERITED;
};
///////////////////////////////////////////////////////////////////////////////
static bool isFinite_int(float x) {
uint32_t bits = SkFloat2Bits(x); // need unsigned for our shifts
int exponent = bits << 1 >> 24;
return exponent != 0xFF;
}
static bool isFinite_float(float x) {
return SkToBool(sk_float_isfinite(x));
}
static bool isFinite_mulzero(float x) {
float y = x * 0;
return y == y;
}
static bool isfinite_and_int(const float data[4]) {
return isFinite_int(data[0]) && isFinite_int(data[1]) && isFinite_int(data[2]) && isFinite_int(data[3]);
}
static bool isfinite_and_float(const float data[4]) {
return isFinite_float(data[0]) && isFinite_float(data[1]) && isFinite_float(data[2]) && isFinite_float(data[3]);
}
static bool isfinite_and_mulzero(const float data[4]) {
return isFinite_mulzero(data[0]) && isFinite_mulzero(data[1]) && isFinite_mulzero(data[2]) && isFinite_mulzero(data[3]);
}
#define mulzeroadd(data) (data[0]*0 + data[1]*0 + data[2]*0 + data[3]*0)
static bool isfinite_plus_int(const float data[4]) {
return isFinite_int(mulzeroadd(data));
}
static bool isfinite_plus_float(const float data[4]) {
return !sk_float_isnan(mulzeroadd(data));
}
static bool isfinite_plus_mulzero(const float data[4]) {
float x = mulzeroadd(data);
return x == x;
}
typedef bool (*IsFiniteProc)(const float[]);
#define MAKEREC(name) { name, #name }
static const struct {
IsFiniteProc fProc;
const char* fName;
} gRec[] = {
MAKEREC(isfinite_and_int),
MAKEREC(isfinite_and_float),
MAKEREC(isfinite_and_mulzero),
MAKEREC(isfinite_plus_int),
MAKEREC(isfinite_plus_float),
MAKEREC(isfinite_plus_mulzero),
};
#undef MAKEREC
static bool isFinite(const SkRect& r) {
// x * 0 will be NaN iff x is infinity or NaN.
// a + b will be NaN iff either a or b is NaN.
float value = r.fLeft * 0 + r.fTop * 0 + r.fRight * 0 + r.fBottom * 0;
// value is either NaN or it is finite (zero).
// value==value will be true iff value is not NaN
return value == value;
}
class IsFiniteBench : public SkBenchmark {
enum {
N = SkBENCHLOOP(1000),
NN = SkBENCHLOOP(1000),
};
float fData[N];
public:
IsFiniteBench(void* param, int index) : INHERITED(param) {
SkRandom rand;
for (int i = 0; i < N; ++i) {
fData[i] = rand.nextSScalar1();
}
if (index < 0) {
fProc = NULL;
fName = "isfinite_rect";
} else {
fProc = gRec[index].fProc;
fName = gRec[index].fName;
}
fIsRendering = false;
}
protected:
virtual void onDraw(SkCanvas* canvas) {
IsFiniteProc proc = fProc;
const float* data = fData;
// do this so the compiler won't throw away the function call
int counter = 0;
if (proc) {
for (int j = 0; j < NN; ++j) {
for (int i = 0; i < N - 4; ++i) {
counter += proc(&data[i]);
}
}
} else {
for (int j = 0; j < NN; ++j) {
for (int i = 0; i < N - 4; ++i) {
const SkRect* r = reinterpret_cast<const SkRect*>(&data[i]);
if (false) { // avoid bit rot, suppress warning
isFinite(*r);
}
counter += r->isFinite();
}
}
}
SkPaint paint;
if (paint.getAlpha() == 0) {
SkDebugf("%d\n", counter);
}
}
virtual const char* onGetName() {
return fName;
}
private:
IsFiniteProc fProc;
const char* fName;
typedef SkBenchmark INHERITED;
};
class FloorBench : public SkBenchmark {
enum {
ARRAY = SkBENCHLOOP(1000),
LOOP = SkBENCHLOOP(1000),
};
float fData[ARRAY];
bool fFast;
public:
FloorBench(void* param, bool fast) : INHERITED(param), fFast(fast) {
SkRandom rand;
for (int i = 0; i < ARRAY; ++i) {
fData[i] = rand.nextSScalar1();
}
if (fast) {
fName = "floor_fast";
} else {
fName = "floor_std";
}
fIsRendering = false;
}
virtual void process(float) {}
protected:
virtual void onDraw(SkCanvas* canvas) {
SkRandom rand;
float accum = 0;
const float* data = fData;
if (fFast) {
for (int j = 0; j < LOOP; ++j) {
for (int i = 0; i < ARRAY; ++i) {
accum += fast_floor(data[i]);
}
this->process(accum);
}
} else {
for (int j = 0; j < LOOP; ++j) {
for (int i = 0; i < ARRAY; ++i) {
accum += sk_float_floor(data[i]);
}
this->process(accum);
}
}
}
virtual const char* onGetName() {
return fName;
}
private:
const char* fName;
typedef SkBenchmark INHERITED;
};
///////////////////////////////////////////////////////////////////////////////
static SkBenchmark* M0(void* p) { return new NoOpMathBench(p); }
static SkBenchmark* M1(void* p) { return new SlowISqrtMathBench(p); }
static SkBenchmark* M2(void* p) { return new FastISqrtMathBench(p); }
static SkBenchmark* M3(void* p) { return new QMul64Bench(p); }
static SkBenchmark* M4(void* p) { return new QMul32Bench(p); }
static SkBenchmark* M5neg1(void* p) { return new IsFiniteBench(p, -1); }
static SkBenchmark* M50(void* p) { return new IsFiniteBench(p, 0); }
static SkBenchmark* M51(void* p) { return new IsFiniteBench(p, 1); }
static SkBenchmark* M52(void* p) { return new IsFiniteBench(p, 2); }
static SkBenchmark* M53(void* p) { return new IsFiniteBench(p, 3); }
static SkBenchmark* M54(void* p) { return new IsFiniteBench(p, 4); }
static SkBenchmark* M55(void* p) { return new IsFiniteBench(p, 5); }
static SkBenchmark* F0(void* p) { return new FloorBench(p, false); }
static SkBenchmark* F1(void* p) { return new FloorBench(p, true); }
static BenchRegistry gReg0(M0);
static BenchRegistry gReg1(M1);
static BenchRegistry gReg2(M2);
static BenchRegistry gReg3(M3);
static BenchRegistry gReg4(M4);
static BenchRegistry gReg5neg1(M5neg1);
static BenchRegistry gReg50(M50);
static BenchRegistry gReg51(M51);
static BenchRegistry gReg52(M52);
static BenchRegistry gReg53(M53);
static BenchRegistry gReg54(M54);
static BenchRegistry gReg55(M55);
static BenchRegistry gRF0(F0);
static BenchRegistry gRF1(F1);