TeamHepta/SDL
1
0
Fork 0
mirror of https://github.com/libsdl-org/SDL.git synced 2026-04-19 23:05:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

[N-Gage] Add various micro-optimizations to rendering back-end

Browse Source 
- Add reusable line points buffer to eliminate per-call heap allocations in DrawLines.
- Cache last draw color to skip redundant SetPenColor/SetBrushColor calls.
- Pre-compute cardinal angle constants (0°, 90°, 180°, 270°) for CopyEx fast-path.
- Cache color modulation state to avoid redundant LUT rebuilds.
- Add missing break statement in HandleEvent.
- Initialize previously uninitialized lastTime variable in UpdateFPS.
This commit is contained in:
Michael Fitzmayer
2026-04-16 20:50:34 +02:00
parent 847fc72b1b
commit b53b31b74a
3 changed files with 125 additions and 74 deletions
Download Patch File Download Diff File Expand all files Collapse all files

151
src/render/ngage/SDL_render_ngage.cpp
View File

@@ -159,6 +159,15 @@ void NGAGE_SuspendScreenSaverInternal(bool suspend)
}
#endif
// Pre-calculated fixed-point angle constants for cardinal rotation checks.
// These avoid repeated Real2Fix conversions in CopyEx hot path.
static const TFixed kAngleZero = 0;
static const TFixed kAnglePi_2 = Real2Fix(M_PI / 2.0); // 90 degrees
static const TFixed kAnglePi = Real2Fix(M_PI); // 180 degrees
static const TFixed kAnglePi3_2 = Real2Fix(3.0 * M_PI / 2.0); // 270 degrees
static const TFixed kAnglePi2 = Real2Fix(2.0 * M_PI); // 360 degrees
static const TFixed kAngleTolerance = 100; // Tolerance for angle comparison
CRenderer *CRenderer::NewL()
{
CRenderer *self = new (ELeave) CRenderer();
@@ -168,7 +177,7 @@ CRenderer *CRenderer::NewL()
return self;
}
CRenderer::CRenderer() : iRenderer(0), iDirectScreen(0), iScreenGc(0), iWsSession(), iWsWindowGroup(), iWsWindowGroupID(0), iWsWindow(), iWsScreen(0), iWsEventStatus(), iWsEvent(), iShowFPS(EFalse), iFPS(0), iFont(0), iWorkBuffer1(0), iWorkBuffer2(0), iWorkBufferSize(0), iTempRenderBitmap(0), iTempRenderBitmapWidth(0), iTempRenderBitmapHeight(0), iLastColorR(-1), iLastColorG(-1), iLastColorB(-1) {}
CRenderer::CRenderer() : iRenderer(0), iDirectScreen(0), iScreenGc(0), iWsSession(), iWsWindowGroup(), iWsWindowGroupID(0), iWsWindow(), iWsScreen(0), iWsEventStatus(), iWsEvent(), iShowFPS(EFalse), iFPS(0), iFont(0), iWorkBuffer1(0), iWorkBuffer2(0), iWorkBufferSize(0), iTempRenderBitmap(0), iTempRenderBitmapWidth(0), iTempRenderBitmapHeight(0), iLastColorR(-1), iLastColorG(-1), iLastColorB(-1), iLinePointsBuffer(0), iLinePointsBufferCapacity(0), iLastDrawColor(0xFFFFFFFF) {}
CRenderer::~CRenderer()
{
@@ -187,6 +196,11 @@ CRenderer::~CRenderer()
iTempRenderBitmap = 0;
iTempRenderBitmapWidth = 0;
iTempRenderBitmapHeight = 0;
// Free line points buffer.
delete[] iLinePointsBuffer;
iLinePointsBuffer = 0;
iLinePointsBufferCapacity = 0;
}
void CRenderer::ConstructL()
@@ -335,10 +349,30 @@ bool CRenderer::EnsureWorkBufferCapacity(TInt aRequiredSize)
return true;
}
bool CRenderer::EnsureLinePointsCapacity(TInt aRequiredCount)
{
if (aRequiredCount <= iLinePointsBufferCapacity) {
return true;
}
// Free old buffer.
delete[] iLinePointsBuffer;
// Allocate new buffer.
iLinePointsBuffer = new TPoint[aRequiredCount];
if (!iLinePointsBuffer) {
iLinePointsBufferCapacity = 0;
return false;
}
iLinePointsBufferCapacity = aRequiredCount;
return true;
}
bool CRenderer::EnsureTempBitmapCapacity(TInt aWidth, TInt aHeight)
{
if (iTempRenderBitmap &&
iTempRenderBitmapWidth >= aWidth &&
if (iTempRenderBitmap &&
iTempRenderBitmapWidth >= aWidth &&
iTempRenderBitmapHeight >= aHeight) {
return true;
}
@@ -373,10 +407,10 @@ void CRenderer::BuildColorModLUT(TFixed rf, TFixed gf, TFixed bf)
{
// Build lookup tables for R, G, B channels.
for (int i = 0; i < 256; i++) {
TFixed val = i << 16; // Convert to fixed-point
iColorModLUT[i] = (TUint8)SDL_min(Fix2Int(FixMul(val, rf)), 255); // R
iColorModLUT[i + 256] = (TUint8)SDL_min(Fix2Int(FixMul(val, gf)), 255); // G
iColorModLUT[i + 512] = (TUint8)SDL_min(Fix2Int(FixMul(val, bf)), 255); // B
TFixed val = i << 16; // Convert to fixed-point
iColorModLUT[i] = (TUint8)SDL_min(Fix2Int(FixMul(val, rf)), 255); // R
iColorModLUT[i + 256] = (TUint8)SDL_min(Fix2Int(FixMul(val, gf)), 255); // G
iColorModLUT[i + 512] = (TUint8)SDL_min(Fix2Int(FixMul(val, bf)), 255); // B
}
// Remember the last color to avoid rebuilding unnecessarily.
@@ -385,7 +419,7 @@ void CRenderer::BuildColorModLUT(TFixed rf, TFixed gf, TFixed bf)
iLastColorB = bf;
}
CFbsBitmap* CRenderer::GetCardinalRotation(NGAGE_TextureData *aTextureData, TInt aAngleIndex)
CFbsBitmap *CRenderer::GetCardinalRotation(NGAGE_TextureData *aTextureData, TInt aAngleIndex)
{
// Check if already cached.
if (aTextureData->cardinalRotations[aAngleIndex]) {
@@ -426,27 +460,27 @@ CFbsBitmap* CRenderer::GetCardinalRotation(NGAGE_TextureData *aTextureData, TInt
int dstY = 0;
switch (aAngleIndex) {
case 0: // 0 degrees
dstX = x;
dstY = y;
break;
case 1: // 90 degrees
dstX = h - 1 - y;
dstY = x;
break;
case 2: // 180 degrees
dstX = w - 1 - x;
dstY = h - 1 - y;
break;
case 3: // 270 degrees
dstX = y;
dstY = w - 1 - x;
break;
default:
// Should never happen, but initialize to avoid warnings
dstX = x;
dstY = y;
break;
case 0: // 0 degrees
dstX = x;
dstY = y;
break;
case 1: // 90 degrees
dstX = h - 1 - y;
dstY = x;
break;
case 2: // 180 degrees
dstX = w - 1 - x;
dstY = h - 1 - y;
break;
case 3: // 270 degrees
dstX = y;
dstY = w - 1 - x;
break;
default:
// Should never happen, but initialize to avoid warnings
dstX = x;
dstY = y;
break;
}
dst[dstY * dstPitch + dstX] = pixel;
@@ -608,18 +642,18 @@ bool CRenderer::CopyEx(SDL_Renderer *renderer, SDL_Texture *texture, const NGAGE
TInt angleIndex = -1;
TFixed angle = copydata->angle;
// Convert angle to degrees and check if it's a cardinal angle.
TFixed zero = 0;
TFixed pi_2 = Real2Fix(M_PI / 2.0);
TFixed pi = Real2Fix(M_PI);
TFixed pi3_2 = Real2Fix(3.0 * M_PI / 2.0);
TFixed pi2 = Real2Fix(2.0 * M_PI);
if (angle == zero) angleIndex = 0;
else if (SDL_abs(angle - pi_2) < 100) angleIndex = 1; // 90°
else if (SDL_abs(angle - pi) < 100) angleIndex = 2; // 180°
else if (SDL_abs(angle - pi3_2) < 100) angleIndex = 3; // 270°
else if (SDL_abs(angle - pi2) < 100) angleIndex = 0; // 360° = 0°
// Use pre-calculated angle constants for comparison.
if (angle == kAngleZero) {
angleIndex = 0;
} else if (SDL_abs(angle - kAnglePi_2) < kAngleTolerance) {
angleIndex = 1; // 90°
} else if (SDL_abs(angle - kAnglePi) < kAngleTolerance) {
angleIndex = 2; // 180°
} else if (SDL_abs(angle - kAnglePi3_2) < kAngleTolerance) {
angleIndex = 3; // 270°
} else if (SDL_abs(angle - kAnglePi2) < kAngleTolerance) {
angleIndex = 0; // 360° = 0°
}
if (angleIndex >= 0) {
CFbsBitmap *cached = GetCardinalRotation(phdata, angleIndex);
@@ -743,7 +777,7 @@ bool CRenderer::CreateTextureData(NGAGE_TextureData *aTextureData, const TInt aW
}
// Initialize dirty tracking.
aTextureData->isDirty = true; // New textures start dirty
aTextureData->isDirty = true; // New textures start dirty.
aTextureData->dirtyRect.x = 0;
aTextureData->dirtyRect.y = 0;
aTextureData->dirtyRect.w = aWidth;
@@ -755,10 +789,14 @@ bool CRenderer::CreateTextureData(NGAGE_TextureData *aTextureData, const TInt aW
void CRenderer::DrawLines(NGAGE_Vertex *aVerts, const TInt aCount)
{
if (iRenderer && iRenderer->Gc()) {
TPoint *aPoints = new TPoint[aCount];
// Ensure reusable buffer has sufficient capacity.
if (!EnsureLinePointsCapacity(aCount)) {
return;
}
// Fill points from vertex data.
for (TInt i = 0; i < aCount; i++) {
aPoints[i] = TPoint(aVerts[i].x, aVerts[i].y);
iLinePointsBuffer[i] = TPoint(aVerts[i].x, aVerts[i].y);
}
TUint32 aColor = (((TUint8)aVerts->color.a << 24) |
@@ -767,9 +805,7 @@ void CRenderer::DrawLines(NGAGE_Vertex *aVerts, const TInt aCount)
(TUint8)aVerts->color.r);
iRenderer->Gc()->SetPenColor(aColor);
iRenderer->Gc()->DrawPolyLineNoEndPoint(aPoints, aCount);
delete[] aPoints;
iRenderer->Gc()->DrawPolyLineNoEndPoint(iLinePointsBuffer, aCount);
}
}
@@ -777,7 +813,7 @@ void CRenderer::DrawPoints(NGAGE_Vertex *aVerts, const TInt aCount)
{
if (iRenderer && iRenderer->Gc()) {
// Batch points by color to minimize SetPenColor calls.
TUint32 currentColor = 0xFFFFFFFF; // Invalid initial color
TUint32 currentColor = 0xFFFFFFFF; // Invalid initial color
bool colorSet = false;
for (TInt i = 0; i < aCount; i++, aVerts++) {
@@ -802,7 +838,7 @@ void CRenderer::FillRects(NGAGE_Vertex *aVerts, const TInt aCount)
{
if (iRenderer && iRenderer->Gc()) {
// Batch rectangles by color to minimize SetPenColor/SetBrushColor calls.
TUint32 currentColor = 0xFFFFFFFF; // Invalid initial color
TUint32 currentColor = 0xFFFFFFFF; // Invalid initial color
bool colorSet = false;
// Process rectangles (each rect uses 2 vertices: position and size).
@@ -876,6 +912,11 @@ void CRenderer::Flip()
void CRenderer::SetDrawColor(TUint32 iColor)
{
if (iRenderer && iRenderer->Gc()) {
// Skip redundant calls if color hasn't changed.
if (iColor == iLastDrawColor) {
return;
}
iRenderer->Gc()->SetPenColor(iColor);
iRenderer->Gc()->SetBrushColor(iColor);
iRenderer->Gc()->SetBrushStyle(CGraphicsContext::ESolidBrush);
@@ -884,6 +925,8 @@ void CRenderer::SetDrawColor(TUint32 iColor)
if (err != KErrNone) {
return;
}
iLastDrawColor = iColor;
}
}
@@ -899,10 +942,17 @@ void CRenderer::UpdateFPS()
{
static TTime lastTime;
static TInt frameCount = 0;
static TBool initialized = EFalse;
TTime currentTime;
const TUint KOneSecond = 1000000; // 1s in ms.
const TUint KOneSecond = 1000000; // 1s in microseconds.
currentTime.HomeTime();
if (!initialized) {
lastTime = currentTime;
initialized = ETrue;
}
++frameCount;
TTimeIntervalMicroSeconds timeDiff = currentTime.MicroSecondsFrom(lastTime);
@@ -1022,6 +1072,7 @@ void CRenderer::HandleEvent(const TWsEvent &aWsEvent)
case EEventKeyUp: /* Key events */
timestamp = SDL_GetPerformanceCounter();
SDL_SendKeyboardKey(timestamp, 1, aWsEvent.Key()->iCode, ConvertScancode(aWsEvent.Key()->iScanCode), false);
break;
case EEventFocusGained:
DisableKeyBlocking();

8
src/render/ngage/SDL_render_ngage_c.hpp
View File

@@ -104,9 +104,17 @@ class CRenderer : public MDirectScreenAccess
TFixed iLastColorG;
TFixed iLastColorB;
// Reusable line points buffer to avoid per-frame allocations in DrawLines.
TPoint *iLinePointsBuffer;
TInt iLinePointsBufferCapacity;
// Cached draw color to avoid redundant SetPenColor/SetBrushColor calls.
TUint32 iLastDrawColor;
// Helper methods.
bool EnsureWorkBufferCapacity(TInt aRequiredSize);
bool EnsureTempBitmapCapacity(TInt aWidth, TInt aHeight);
bool EnsureLinePointsCapacity(TInt aRequiredCount);
void BuildColorModLUT(TFixed rf, TFixed gf, TFixed bf);
CFbsBitmap *GetCardinalRotation(NGAGE_TextureData *aTextureData, TInt aAngleIndex);
};

40
src/render/ngage/SDL_render_ops.cpp
View File

@@ -36,12 +36,12 @@ void ApplyColorMod(void *dest, void *source, int pitch, int width, int height, S
const TUint8 *lut_g = colorLUT + 256;
const TUint8 *lut_b = colorLUT + 512;
// Process 4 pixels at a time (loop unrolling).
// Process 4 pixels at a time.
for (int y = 0; y < height; ++y) {
const TInt rowOffset = y * pitchPixels;
int x = 0;
// Unrolled loop: process 4 pixels at once with optimized bit manipulation.
// Process 4 pixels at once with optimized bit manipulation.
for (; x < width - 3; x += 4) {
// Load 4 pixels at once.
TUint16 p0 = src_pixels[rowOffset + x];
@@ -51,9 +51,9 @@ void ApplyColorMod(void *dest, void *source, int pitch, int width, int height, S
// Pixel 0: Extract and modulate RGB4444 components.
// RGB4444 format: RRRR GGGG BBBB xxxx
TUint8 r0 = lut_r[(p0 >> 8) & 0xF0]; // Extract R (bits 12-15), shift to byte position
TUint8 g0 = lut_g[(p0 >> 3) & 0xF8]; // Extract G (bits 6-9), scale to 8-bit
TUint8 b0 = lut_b[(p0 << 3) & 0xF8]; // Extract B (bits 0-3), scale to 8-bit
TUint8 r0 = lut_r[(p0 >> 8) & 0xF0]; // Extract R (bits 12-15), shift to byte position
TUint8 g0 = lut_g[(p0 >> 3) & 0xF8]; // Extract G (bits 6-9), scale to 8-bit
TUint8 b0 = lut_b[(p0 << 3) & 0xF8]; // Extract B (bits 0-3), scale to 8-bit
dst_pixels[rowOffset + x] = ((r0 & 0xF0) << 8) | ((g0 & 0xF0) << 3) | ((b0 & 0xF0) >> 1);
// Pixel 1
@@ -112,7 +112,7 @@ void ApplyFlip(void *dest, void *source, int pitch, int width, int height, SDL_F
int x = 0;
// Unrolled loop: process 4 pixels at once.
// Process 4 pixels at once.
for (; x < width - 3; x += 4) {
if (flipHorizontal) {
dst_pixels[dstRowOffset + x] = src_pixels[srcRowOffset + (width_m1 - x)];
@@ -172,7 +172,7 @@ void ApplyRotation(void *dest, void *source, int pitch, int width, int height, T
int x = 0;
// Unrolled loop: process 4 pixels at once.
// Process 4 pixels at once.
for (; x < width - 3; x += 4) {
// Pixel 0
int final_x0 = Fix2Int(src_x);
@@ -199,14 +199,10 @@ void ApplyRotation(void *dest, void *source, int pitch, int width, int height, T
src_y += dx_sin;
// Write all 4 pixels with bounds checking.
dst_pixels[dstRowOffset + x] = (final_x0 >= 0 && final_x0 < width && final_y0 >= 0 && final_y0 < height) ?
src_pixels[final_y0 * pitchPixels + final_x0] : 0;
dst_pixels[dstRowOffset + x + 1] = (final_x1 >= 0 && final_x1 < width && final_y1 >= 0 && final_y1 < height) ?
src_pixels[final_y1 * pitchPixels + final_x1] : 0;
dst_pixels[dstRowOffset + x + 2] = (final_x2 >= 0 && final_x2 < width && final_y2 >= 0 && final_y2 < height) ?
src_pixels[final_y2 * pitchPixels + final_x2] : 0;
dst_pixels[dstRowOffset + x + 3] = (final_x3 >= 0 && final_x3 < width && final_y3 >= 0 && final_y3 < height) ?
src_pixels[final_y3 * pitchPixels + final_x3] : 0;
dst_pixels[dstRowOffset + x] = (final_x0 >= 0 && final_x0 < width && final_y0 >= 0 && final_y0 < height) ? src_pixels[final_y0 * pitchPixels + final_x0] : 0;
dst_pixels[dstRowOffset + x + 1] = (final_x1 >= 0 && final_x1 < width && final_y1 >= 0 && final_y1 < height) ? src_pixels[final_y1 * pitchPixels + final_x1] : 0;
dst_pixels[dstRowOffset + x + 2] = (final_x2 >= 0 && final_x2 < width && final_y2 >= 0 && final_y2 < height) ? src_pixels[final_y2 * pitchPixels + final_x2] : 0;
dst_pixels[dstRowOffset + x + 3] = (final_x3 >= 0 && final_x3 < width && final_y3 >= 0 && final_y3 < height) ? src_pixels[final_y3 * pitchPixels + final_x3] : 0;
}
// Handle remaining pixels.
@@ -265,7 +261,7 @@ void ApplyScale(void *dest, void *source, int pitch, int width, int height, TFix
int x = 0;
// Unrolled loop: process 4 pixels at once.
// Process 4 pixels at once.
for (; x < width - 3; x += 4) {
// Process 4 pixels using incremental approach.
int final_x0 = Fix2Int(src_x);
@@ -278,14 +274,10 @@ void ApplyScale(void *dest, void *source, int pitch, int width, int height, TFix
src_x += inv_scale_x;
// Write all 4 pixels with bounds checking.
dst_pixels[dstRowOffset + x] = (rowInBounds && final_x0 >= 0 && final_x0 < width) ?
src_pixels[srcRowOffset + final_x0] : 0;
dst_pixels[dstRowOffset + x + 1] = (rowInBounds && final_x1 >= 0 && final_x1 < width) ?
src_pixels[srcRowOffset + final_x1] : 0;
dst_pixels[dstRowOffset + x + 2] = (rowInBounds && final_x2 >= 0 && final_x2 < width) ?
src_pixels[srcRowOffset + final_x2] : 0;
dst_pixels[dstRowOffset + x + 3] = (rowInBounds && final_x3 >= 0 && final_x3 < width) ?
src_pixels[srcRowOffset + final_x3] : 0;
dst_pixels[dstRowOffset + x] = (rowInBounds && final_x0 >= 0 && final_x0 < width) ? src_pixels[srcRowOffset + final_x0] : 0;
dst_pixels[dstRowOffset + x + 1] = (rowInBounds && final_x1 >= 0 && final_x1 < width) ? src_pixels[srcRowOffset + final_x1] : 0;
dst_pixels[dstRowOffset + x + 2] = (rowInBounds && final_x2 >= 0 && final_x2 < width) ? src_pixels[srcRowOffset + final_x2] : 0;
dst_pixels[dstRowOffset + x + 3] = (rowInBounds && final_x3 >= 0 && final_x3 < width) ? src_pixels[srcRowOffset + final_x3] : 0;
}
// Handle remaining pixels.