doylet
/
DTRenderer
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Start separating inline SIMD to own functions

This commit is contained in:
Doyle Thai 2017-06-01 18:52:50 +10:00
parent 6bcdb6d1fb
commit 9b4072f5bc
2 changed files with 401 additions and 211 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
src/DTRendererDebug.h
View File

@ -44,12 +44,15 @@ enum DTRDebugCounter
enum DTRDebugCycleCount
{
DTRDebugCycleCount_RenderTexturedTriangle_Rasterise,
DTRDebugCycleCount_RenderTexturedTriangle_SampleTextureFunction,
DTRDebugCycleCount_RenderTexturedTriangle_SampleTexture,
DTRDebugCycleCount_RenderTriangle_Rasterise,
DTRDebugCycleCount_Count,
};
typedef struct DTRDebugCycles
{
char *name;
u64 totalCycles;
u64 numInvokes;

609
src/DTRendererRender.cpp
View File

@ -11,6 +11,18 @@
FILE_SCOPE const f32 COLOR_EPSILON = 0.9f;
inline void Make3PointsClockwise(DqnV3 *p1, DqnV3 *p2, DqnV3 *p3)
{
f32 area2Times = ((p2->x - p1->x) * (p2->y + p1->y)) +
((p3->x - p2->x) * (p3->y + p2->y)) +
((p1->x - p3->x) * (p1->y + p3->y));
if (area2Times > 0)
{
// Clockwise swap any point to make it clockwise
DQN_SWAP(DqnV3, *p2, *p3);
}
}
FILE_SCOPE inline DqnV4 PreMultiplyAlpha1(const DqnV4 color)
{
DQN_ASSERT(color.a >= 0.0f && color.a <= 1.0f);
@ -61,20 +73,9 @@ inline f32 DTRRender_SRGB1ToLinearSpacef(f32 val)
inline DqnV4 DTRRender_SRGB1ToLinearSpaceV4(DqnV4 color)
{
DqnV4 result;
if (globalDTRPlatformFlags.canUseSSE2)
{
__m128 simdColor = _mm_set_ps(color.r, color.g, color.b, 0);
__m128 simdResult = _mm_mul_ps(simdColor, simdColor);
result.r = ((f32 *)&simdResult)[3];
result.g = ((f32 *)&simdResult)[2];
result.b = ((f32 *)&simdResult)[1];
}
else
{
result.r = DTRRender_SRGB1ToLinearSpacef(color.r);
result.g = DTRRender_SRGB1ToLinearSpacef(color.g);
result.b = DTRRender_SRGB1ToLinearSpacef(color.b);
}
result.r = DTRRender_SRGB1ToLinearSpacef(color.r);
result.g = DTRRender_SRGB1ToLinearSpacef(color.g);
result.b = DTRRender_SRGB1ToLinearSpacef(color.b);
result.a = color.a;
return result;
@ -91,20 +92,9 @@ inline f32 DTRRender_LinearToSRGB1Spacef(f32 val)
inline DqnV4 DTRRender_LinearToSRGB1SpaceV4(DqnV4 color)
{
DqnV4 result;
if (globalDTRPlatformFlags.canUseSSE2)
{
__m128 simdColor = _mm_set_ps(color.r, color.g, color.b, 0);
__m128 simdResult = _mm_sqrt_ps(simdColor);
result.r = ((f32 *)&simdResult)[3];
result.g = ((f32 *)&simdResult)[2];
result.b = ((f32 *)&simdResult)[1];
}
else
{
result.r = DTRRender_LinearToSRGB1Spacef(color.r);
result.g = DTRRender_LinearToSRGB1Spacef(color.g);
result.b = DTRRender_LinearToSRGB1Spacef(color.b);
}
result.r = DTRRender_LinearToSRGB1Spacef(color.r);
result.g = DTRRender_LinearToSRGB1Spacef(color.g);
result.b = DTRRender_LinearToSRGB1Spacef(color.b);
result.a = color.a;
return result;
@ -139,7 +129,6 @@ FILE_SCOPE inline void SetPixel(DTRRenderBuffer *const renderBuffer, const i32 x
if (needGammaFix) color = DTRRender_SRGB1ToLinearSpaceV4(color);
u32 src = bitmapPtr[x + (y * pitchInU32)];
#if 0
f32 srcR = (f32)((src >> 16) & 0xFF) * DTRRENDER_INV_255;
f32 srcG = (f32)((src >> 8) & 0xFF) * DTRRENDER_INV_255;
f32 srcB = (f32)((src >> 0) & 0xFF) * DTRRENDER_INV_255;
@ -181,44 +170,6 @@ FILE_SCOPE inline void SetPixel(DTRRenderBuffer *const renderBuffer, const i32 x
destB = 255;
}
#else
__m128 simdSrc = _mm_set_ps(0.0f,
(f32)((src >> 16) & 0xFF),
(f32)((src >> 8) & 0xFF),
(f32)((src >> 0) & 0xFF));
__m128 inv255_4x = _mm_set_ps1(DTRRENDER_INV_255);
simdSrc = _mm_mul_ps(simdSrc, inv255_4x);
simdSrc = _mm_mul_ps(simdSrc, simdSrc); // to linear
f32 invANorm = 1 - color.a;
__m128 invANorm_4x = _mm_set_ps1(invANorm);
__m128 const255_4x = _mm_set_ps1(255.0f);
__m128 simdColor = _mm_set_ps(0, color.r, color.g, color.b);
__m128 dest = _mm_add_ps(simdColor, _mm_mul_ps(simdSrc, invANorm_4x)); // to 0->1 range
dest = _mm_sqrt_ps(dest); // to srgb
dest = _mm_mul_ps(dest, const255_4x); // to 0->255 range
DQN_ASSERT(((f32 *)&dest)[2] >= 0);
DQN_ASSERT(((f32 *)&dest)[1] >= 0);
DQN_ASSERT(((f32 *)&dest)[0] >= 0);
if (DTR_DEBUG)
{
DQN_ASSERT((((f32 *)&dest)[2] - 255.0f) < COLOR_EPSILON);
DQN_ASSERT((((f32 *)&dest)[1] - 255.0f) < COLOR_EPSILON);
DQN_ASSERT((((f32 *)&dest)[0] - 255.0f) < COLOR_EPSILON);
}
dest = _mm_min_ps(dest, const255_4x);
f32 destR = ((f32 *)&dest)[2];
f32 destG = ((f32 *)&dest)[1];
f32 destB = ((f32 *)&dest)[0];
#endif
u32 pixel = // ((u32)(destA) << 24 |
(u32)(destR) << 16 |
(u32)(destG) << 8 |
@ -582,29 +533,24 @@ typedef struct TriangleInclusionTest
typedef struct SIMDTriangleInclusionTest
{
DqnV2i boundsMin;
DqnV2i boundsMax;
__m128 vertexZValues;
__m128 signedAreaPixelDeltaX;
__m128 signedAreaPixelDeltaY;
__m128 invSignedAreaParallelogram_4x;
__m128 startPixel;
DqnV2i boundsMin;
DqnV2i boundsMax;
DqnV3 p1;
DqnV3 p2;
DqnV3 p3;
} SIMDTriangleInclusionTest;
FILE_SCOPE TriangleInclusionTest CreateTriangleInclusionTest(const i32 clipWidth,
const i32 clipHeight, DqnV3 p1,
DqnV3 p2, DqnV3 p3)
{
f32 area2Times = ((p2.x - p1.x) * (p2.y + p1.y)) + ((p3.x - p2.x) * (p3.y + p2.y)) +
((p1.x - p3.x) * (p1.y + p3.y));
if (area2Times > 0)
{
// Clockwise swap any point to make it clockwise
DQN_SWAP(DqnV3, p2, p3);
}
Make3PointsClockwise(&p1, &p2, &p3);
TriangleInclusionTest result = {};
result.boundsMin = DqnV2i_2f(DQN_MIN(DQN_MIN(p1.x, p2.x), p3.x), DQN_MIN(DQN_MIN(p1.y, p2.y), p3.y));
@ -757,6 +703,9 @@ FILE_SCOPE SIMDTriangleInclusionTest CreateSimdTriangleInclusionTest(
result.boundsMax = inclusionTest.boundsMax;
// NOTE: Order is important here!
result.p1 = p1;
result.p2 = p2;
result.p3 = p3;
result.vertexZValues = _mm_set_ps(0, p3.z, p2.z, p1.z);
result.signedAreaPixelDeltaX = _mm_set_ps(0,
inclusionTest.signedAreaP3DeltaX,
@ -774,7 +723,6 @@ FILE_SCOPE SIMDTriangleInclusionTest CreateSimdTriangleInclusionTest(
return result;
}
inline void RasteriseTexturedTriangle(DTRRenderBuffer *const renderBuffer, const DqnV3 p1,
const DqnV3 p2, const DqnV3 p3, const DqnV2 uv1,
const DqnV2 uv2, const DqnV2 uv3, DTRBitmap *const texture,
@ -910,14 +858,215 @@ inline void RasteriseTexturedTriangle(DTRRenderBuffer *const renderBuffer, const
}
}
inline void SIMDRasteriseTexturedTriangle(DTRRenderBuffer *const renderBuffer, const DqnV3 p1,
const DqnV3 p2, const DqnV3 p3, const DqnV2 uv1,
const DqnV2 uv2, const DqnV2 uv3,
DTRBitmap *const texture, const DqnV4 color)
FILE_SCOPE inline f32 Triangle2TimesSignedArea(const DqnV2 a, const DqnV2 b, const DqnV2 c)
{
f32 result = ((b.x - a.x) * (c.y - a.y)) - ((b.y - a.y) * (c.x - a.x));
return result;
}
////////////////////////////////////////////////////////////////////////////////
// SIMD
////////////////////////////////////////////////////////////////////////////////
// color: _mm_set_ps(a, b, g, r) ie. 0=r, 1=g, 2=b, 3=a
FILE_SCOPE inline void SIMDDebug_ColorInRange(__m128 color, f32 min, f32 max)
{
if (DTR_DEBUG)
{
f32 r = ((f32 *)&color)[0];
f32 g = ((f32 *)&color)[1];
f32 b = ((f32 *)&color)[2];
f32 a = ((f32 *)&color)[3];
DQN_ASSERT(r >= min && r <= max);
DQN_ASSERT(g >= min && g <= max);
DQN_ASSERT(b >= min && b <= max);
DQN_ASSERT(a >= min && a <= max);
}
}
FILE_SCOPE inline __m128 SIMD_SRGB1ToLinearSpace(__m128 color)
{
SIMDDebug_ColorInRange(color, 0.0f, 1.0f);
f32 preserveAlpha = ((f32 *)&color)[3];
__m128 result = _mm_mul_ps(color, color);
((f32 *)&result)[3] = preserveAlpha;
return result;
}
FILE_SCOPE inline __m128 SIMD_SRGB255ToLinearSpace1(__m128 color)
{
LOCAL_PERSIST const __m128 INV255_4X = _mm_set_ps1(DTRRENDER_INV_255);
color = _mm_mul_ps(color, INV255_4X);
f32 preserveAlpha = ((f32 *)&color)[3];
__m128 result = _mm_mul_ps(color, color);
((f32 *)&result)[3] = preserveAlpha;
return result;
}
FILE_SCOPE inline __m128 SIMD_LinearSpace1ToSRGB1(__m128 color)
{
SIMDDebug_ColorInRange(color, 0.0f, 1.0f);
f32 preserveAlpha = ((f32 *)&color)[3];
__m128 result = _mm_sqrt_ps(color);
((f32 *)&result)[3] = preserveAlpha;
return result;
}
// color: _mm_set_ps(a, b, g, r) ie. 0=r, 1=g, 2=b, 3=a
FILE_SCOPE inline __m128 SIMD_PreMultiplyAlpha1(__m128 color)
{
f32 alpha = ((f32 *)&color)[3];
__m128 simdAlpha = _mm_set_ps(1, alpha, alpha, alpha);
__m128 result = _mm_mul_ps(color, simdAlpha);
return result;
}
FILE_SCOPE inline DqnV2 Get2DOriginFromTransformAnchor(const DqnV2 p1, const DqnV2 p2,
const DqnV2 p3,
const DTRRenderTransform transform)
{
DqnV2 p1p2 = p2 - p1;
DqnV2 p1p3 = p3 - p1;
DqnV2 p1p2Anchored = p1p2 * transform.anchor;
DqnV2 p1p3Anchored = p1p3 * transform.anchor;
DqnV2 origin = p1 + p1p2Anchored + p1p3Anchored;
return origin;
}
// color: _mm_set_ps(a, b, g, r) ie. 0=r, 1=g, 2=b, 3=a
FILE_SCOPE inline void SIMD_SetPixel(DTRRenderBuffer *const renderBuffer, const i32 x, const i32 y,
__m128 color,
const enum ColorSpace colorSpace = ColorSpace_SRGB)
{
if (!renderBuffer) return;
if (x < 0 || x > (renderBuffer->width - 1)) return;
if (y < 0 || y > (renderBuffer->height - 1)) return;
DTR_DEBUG_EP_TIMED_FUNCTION();
SIMDDebug_ColorInRange(color, 0.0f, 1.0f);
u32 *const bitmapPtr = (u32 *)renderBuffer->memory;
const u32 pitchInU32 = (renderBuffer->width * renderBuffer->bytesPerPixel) / 4;
// If some alpha is involved, we need to apply gamma correction, but if the
// new pixel is totally opaque or invisible then we're just flat out
// overwriting/keeping the state of the pixel so we can save cycles by skipping.
f32 alpha = ((f32 *)&color)[3];
bool needGammaFix = (alpha > 0.0f || alpha < (1.0f + COLOR_EPSILON)) && (colorSpace == ColorSpace_SRGB);
if (needGammaFix) color = SIMD_SRGB1ToLinearSpace(color);
// Format: u32 == (XX, RR, GG, BB)
u32 srcPixel = bitmapPtr[x + (y * pitchInU32)];
__m128 src = _mm_set_ps(0,
(f32)((srcPixel >> 0) & 0xFF),
(f32)((srcPixel >> 8) & 0xFF),
(f32)((srcPixel >> 16) & 0xFF));
src = SIMD_SRGB255ToLinearSpace1(src);
f32 invA = 1 - alpha;
__m128 invA_4x = _mm_set_ps1(invA);
// PreAlphaMulColor + (1 - Alpha) * Src
__m128 oneMinusAlphaSrc = _mm_mul_ps(invA_4x, src);
__m128 dest = _mm_add_ps(color, oneMinusAlphaSrc);
dest = SIMD_LinearSpace1ToSRGB1(dest);
dest = _mm_mul_ps(dest, _mm_set_ps1(255.0f)); // to 0->255 range
SIMDDebug_ColorInRange(dest, 0.0f, 255.0f);
f32 destR = ((f32 *)&dest)[0];
f32 destG = ((f32 *)&dest)[1];
f32 destB = ((f32 *)&dest)[2];
u32 pixel = // ((u32)(destA) << 24 |
(u32)(destR) << 16 |
(u32)(destG) << 8 |
(u32)(destB) << 0;
bitmapPtr[x + (y * pitchInU32)] = pixel;
DTRDebug_CounterIncrement(DTRDebugCounter_SetPixels);
}
// colorModulate: _mm_set_ps(a, b, g, r) ie. 0=r, 1=g, 2=b, 3=a
// barycentric: _mm_set_ps(xx, p3, p2, p1) ie. 0=p1, 1=p2, 2=p3, 3=a
FILE_SCOPE __m128 SIMD_SampleTextureForTriangle(DTRBitmap *const texture, const DqnV2 uv1,
const DqnV2 uv2SubUv1, const DqnV2 uv3SubUv1,
const __m128 barycentric)
{
DTRDebug_BeginCycleCount(DTRDebugCycleCount_RenderTexturedTriangle_SampleTextureFunction);
LOCAL_PERSIST const __m128 INV255_4X = _mm_set_ps1(1.0f / 255.0f);
const f32 barycentricP2 = ((f32 *)&barycentric)[1];
const f32 barycentricP3 = ((f32 *)&barycentric)[2];
DqnV2 uv = uv1 + (uv2SubUv1 * barycentricP2) + (uv3SubUv1 * barycentricP3);
const f32 EPSILON = 0.1f;
DQN_ASSERT(uv.x >= 0 && uv.x < 1.0f + EPSILON);
DQN_ASSERT(uv.y >= 0 && uv.y < 1.0f + EPSILON);
uv.x = DqnMath_Clampf(uv.x, 0.0f, 1.0f);
uv.y = DqnMath_Clampf(uv.y, 0.0f, 1.0f);
f32 texelXf = uv.x * texture->dim.w;
f32 texelYf = uv.y * texture->dim.h;
DQN_ASSERT(texelXf >= 0 && texelXf < texture->dim.w);
DQN_ASSERT(texelYf >= 0 && texelYf < texture->dim.h);
i32 texelX = (i32)texelXf;
i32 texelY = (i32)texelYf;
const u32 texturePitch = texture->bytesPerPixel * texture->dim.w;
const u8 *const texturePtr = texture->memory;
u32 texel1 = *(u32 *)(texturePtr + (texelX * texture->bytesPerPixel) + (texelY * texturePitch));
__m128 color = _mm_set_ps((f32)(texel1 >> 24),
(f32)((texel1 >> 16) & 0xFF),
(f32)((texel1 >> 8) & 0xFF),
(f32)((texel1 >> 0) & 0xFF));
color = SIMD_SRGB255ToLinearSpace1(color);
DTRDebug_EndCycleCount(DTRDebugCycleCount_RenderTexturedTriangle_SampleTextureFunction);
return color;
}
FILE_SCOPE void SIMD_TexturedTriangle(DTRRenderBuffer *const renderBuffer, DqnV3 p1, DqnV3 p2,
DqnV3 p3, DqnV2 uv1, DqnV2 uv2, DqnV2 uv3,
DTRBitmap *const texture, DqnV4 color,
const DTRRenderTransform transform)
{
DTR_DEBUG_EP_TIMED_FUNCTION();
////////////////////////////////////////////////////////////////////////////
// Calculate Bounding Box
// Convert color
////////////////////////////////////////////////////////////////////////////
__m128 simdColor = _mm_set_ps(color.a, color.b, color.g, color.r);
simdColor = SIMD_SRGB1ToLinearSpace(simdColor);
simdColor = SIMD_PreMultiplyAlpha1(simdColor);
////////////////////////////////////////////////////////////////////////////
// Transform vertexes p1, p2, p3 inplace
////////////////////////////////////////////////////////////////////////////
{
Make3PointsClockwise(&p1, &p2, &p3);
// TODO(doyle): Transform is only in 2d right now
DqnV2 origin = Get2DOriginFromTransformAnchor(p1.xy, p2.xy, p3.xy, transform);
DqnV2 pList[3] = {p1.xy - origin, p2.xy - origin, p3.xy - origin};
TransformPoints(origin, pList, DQN_ARRAY_COUNT(pList), transform.scale, transform.rotation);
p1.xy = pList[0];
p2.xy = pList[1];
p3.xy = pList[2];
}
DTRDebug_BeginCycleCount(DTRDebugCycleCount_RenderTexturedTriangle_Rasterise);
DqnV2i max = DqnV2i_2f(DQN_MAX(DQN_MAX(p1.x, p2.x), p3.x),
DQN_MAX(DQN_MAX(p1.y, p2.y), p3.y));
DqnV2i min = DqnV2i_2f(DQN_MIN(DQN_MIN(p1.x, p2.x), p3.x),
@ -927,56 +1076,73 @@ inline void SIMDRasteriseTexturedTriangle(DTRRenderBuffer *const renderBuffer, c
max.x = DQN_MIN(max.x, renderBuffer->width - 1);
max.y = DQN_MIN(max.y, renderBuffer->height - 1);
const u32 zBufferPitch = renderBuffer->width;
const DqnV3 a = p1;
const DqnV3 b = p2;
const DqnV3 c = p3;
DqnV2i startP = min;
f32 signedAreaC = ((b.x - a.x) * (startP.y - a.y)) - ((b.y - a.y) * (startP.x - a.x));
f32 signedAreaCDeltaX = a.y - b.y;
f32 signedAreaCDeltaY = b.x - a.x;
f32 signedAreaA = ((c.x - b.x) * (startP.y - b.y)) - ((c.y - b.y) * (startP.x - b.x));
f32 signedAreaADeltaX = b.y - c.y;
f32 signedAreaADeltaY = c.x - b.x;
f32 signedAreaB = ((a.x - c.x) * (startP.y - c.y)) - ((a.y - c.y) * (startP.x - c.x));
f32 signedAreaBDeltaX = c.y - a.y;
f32 signedAreaBDeltaY = a.x - c.x;
f32 signedAreaParallelogram = signedAreaC + signedAreaA + signedAreaB;
if (signedAreaParallelogram == 0) return;
f32 invSignedAreaParallelogram = 1.0f / signedAreaParallelogram;
__m128 invSignedAreaParallelogram_4x = _mm_set_ps1(invSignedAreaParallelogram);
// NOTE: Order is important here!
__m128 triangleZ = _mm_set_ps(0, b.z, a.z, c.z);
__m128 signedAreaPixelDeltaX = _mm_set_ps(0, signedAreaBDeltaX, signedAreaADeltaX, signedAreaCDeltaX);
__m128 signedAreaPixelDeltaY = _mm_set_ps(0, signedAreaBDeltaY, signedAreaADeltaY, signedAreaCDeltaY);
__m128 signedAreaPixel1 = _mm_set_ps(0, signedAreaB, signedAreaA, signedAreaC);
__m128 signedAreaPixel2 = _mm_add_ps(signedAreaPixel1, signedAreaPixelDeltaX);
////////////////////////////////////////////////////////////////////////////
// Setup SIMD data
////////////////////////////////////////////////////////////////////////////
const u32 NUM_X_PIXELS_TO_SIMD = 2;
const u32 NUM_Y_PIXELS_TO_SIMD = 1;
const __m128 INV255_4X = _mm_set_ps1(1.0f / 255.0f);
const __m128 ZERO_4X = _mm_set_ps1(0.0f);
const u32 IS_GREATER_MASK = 0xF;
const u32 NUM_X_PIXELS_TO_SIMD = 2;
const u32 NUM_Y_PIXELS_TO_SIMD = 1;
const __m128 STEP_X_4X = _mm_set_ps1((f32)NUM_X_PIXELS_TO_SIMD);
const __m128 STEP_Y_4X = _mm_set_ps1((f32)NUM_Y_PIXELS_TO_SIMD);
// SignedArea: _mm_set_ps(unused, p3, p2, p1) ie 0=p1, 1=p1, 2=p3, 3=unused
__m128 signedAreaPixel1;
__m128 signedAreaPixel2;
// NOTE: Increase step size to the number of pixels rasterised with SIMD
signedAreaPixelDeltaX = _mm_mul_ps(signedAreaPixelDeltaX, STEP_X_4X);
signedAreaPixelDeltaY = _mm_mul_ps(signedAreaPixelDeltaY, STEP_Y_4X);
__m128 signedAreaPixelDeltaX;
__m128 signedAreaPixelDeltaY;
__m128 invSignedAreaParallelogram_4x;
__m128 triangleZ = _mm_set_ps(0, p3.z, p2.z, p1.z);
{
DqnV2i startP = min;
f32 signedArea1Start = Triangle2TimesSignedArea(p2.xy, p3.xy, DqnV2_V2i(startP));
f32 signedArea1DeltaX = p2.y - p3.y;
f32 signedArea1DeltaY = p3.x - p2.x;
f32 signedArea2Start = Triangle2TimesSignedArea(p3.xy, p1.xy, DqnV2_V2i(startP));
f32 signedArea2DeltaX = p3.y - p1.y;
f32 signedArea2DeltaY = p1.x - p3.x;
f32 signedArea3Start = Triangle2TimesSignedArea(p1.xy, p2.xy, DqnV2_V2i(startP));
f32 signedArea3DeltaX = p1.y - p2.y;
f32 signedArea3DeltaY = p2.x - p1.x;
f32 signedAreaParallelogram = signedArea1Start + signedArea2Start + signedArea3Start;
if (signedAreaParallelogram == 0) return;
f32 invSignedAreaParallelogram = 1.0f / signedAreaParallelogram;
invSignedAreaParallelogram_4x = _mm_set_ps1(invSignedAreaParallelogram);
// NOTE: Order is important here!
signedAreaPixelDeltaX = _mm_set_ps(0, signedArea3DeltaX, signedArea2DeltaX, signedArea1DeltaX);
signedAreaPixelDeltaY = _mm_set_ps(0, signedArea3DeltaY, signedArea2DeltaY, signedArea1DeltaY);
signedAreaPixel1 = _mm_set_ps(0, signedArea3Start, signedArea2Start, signedArea1Start);
signedAreaPixel2 = _mm_add_ps(signedAreaPixel1, signedAreaPixelDeltaX);
// NOTE: Increase step size to the number of pixels rasterised with SIMD
{
const __m128 STEP_X_4X = _mm_set_ps1((f32)NUM_X_PIXELS_TO_SIMD);
const __m128 STEP_Y_4X = _mm_set_ps1((f32)NUM_Y_PIXELS_TO_SIMD);
signedAreaPixelDeltaX = _mm_mul_ps(signedAreaPixelDeltaX, STEP_X_4X);
signedAreaPixelDeltaY = _mm_mul_ps(signedAreaPixelDeltaY, STEP_Y_4X);
}
}
const DqnV2 uv2SubUv1 = uv2 - uv1;
const DqnV2 uv3SubUv1 = uv3 - uv1;
const __m128 colorModulate = _mm_set_ps(color.a, color.b, color.g, color.r);
const u32 texturePitch = texture->bytesPerPixel * texture->dim.w;
const u8 *const texturePtr = texture->memory;
const u32 zBufferPitch = renderBuffer->width;
////////////////////////////////////////////////////////////////////////////
// Scan and Render
////////////////////////////////////////////////////////////////////////////
for (i32 bufferY = min.y; bufferY < max.y; bufferY += NUM_Y_PIXELS_TO_SIMD)
{
__m128 signedArea1 = signedAreaPixel1;
@ -984,6 +1150,7 @@ inline void SIMDRasteriseTexturedTriangle(DTRRenderBuffer *const renderBuffer, c
for (i32 bufferX = min.x; bufferX < max.x; bufferX += NUM_X_PIXELS_TO_SIMD)
{
// Rasterise buffer(X, Y) pixel
{
__m128 checkArea = signedArea1;
@ -992,58 +1159,25 @@ inline void SIMDRasteriseTexturedTriangle(DTRRenderBuffer *const renderBuffer, c
i32 posX = bufferX;
i32 posY = bufferY;
if ((isGreaterResult & IS_GREATER_MASK) == IS_GREATER_MASK && posX < max.x)
if ((isGreaterResult & IS_GREATER_MASK) == IS_GREATER_MASK)
{
__m128 barycentric = _mm_mul_ps(checkArea, invSignedAreaParallelogram_4x);
__m128 barycentricZ = _mm_mul_ps(triangleZ, barycentric);
i32 zBufferIndex = posX + (posY * zBufferPitch);
f32 pixelZValue = ((f32 *)&barycentricZ)[0] +
((f32 *)&barycentricZ)[1] +
((f32 *)&barycentricZ)[2];
((f32 *)&barycentricZ)[1] +
((f32 *)&barycentricZ)[2];
f32 currZValue = renderBuffer->zBuffer[zBufferIndex];
if (pixelZValue > currZValue)
{
renderBuffer->zBuffer[zBufferIndex] = pixelZValue;
f32 barycentricB = ((f32 *)&barycentric)[2];
f32 barycentricC = ((f32 *)&barycentric)[0];
DqnV2 uv = uv1 + (uv2SubUv1 * barycentricB) + (uv3SubUv1 * barycentricC);
const f32 EPSILON = 0.1f;
DQN_ASSERT(uv.x >= 0 && uv.x < 1.0f + EPSILON);
DQN_ASSERT(uv.y >= 0 && uv.y < 1.0f + EPSILON);
uv.x = DqnMath_Clampf(uv.x, 0.0f, 1.0f);
uv.y = DqnMath_Clampf(uv.y, 0.0f, 1.0f);
f32 texelXf = uv.x * texture->dim.w;
f32 texelYf = uv.y * texture->dim.h;
DQN_ASSERT(texelXf >= 0 && texelXf < texture->dim.w);
DQN_ASSERT(texelYf >= 0 && texelYf < texture->dim.h);
i32 texelX = (i32)texelXf;
i32 texelY = (i32)texelYf;
u32 texel1 = *(u32 *)(texturePtr + (texelX * texture->bytesPerPixel) + (texelY * texturePitch));
__m128 color1 = _mm_set_ps((f32)(texel1 >> 24),
(f32)((texel1 >> 16) & 0xFF),
(f32)((texel1 >> 8) & 0xFF),
(f32)((texel1 >> 0) & 0xFF));
color1 = _mm_mul_ps(color1, INV255_4X);
f32 preserveAlpha = ((f32 *)&color1)[3];
color1 = _mm_mul_ps(color1, color1); // to linear space
((f32 *)&color1)[3] = preserveAlpha;
color1 = _mm_mul_ps(color1, colorModulate);
DqnV4 blend = {};
blend.r = ((f32 *)&color1)[0];
blend.g = ((f32 *)&color1)[1];
blend.b = ((f32 *)&color1)[2];
blend.a = ((f32 *)&color1)[3];
SetPixel(renderBuffer, posX, posY, blend, ColorSpace_Linear);
__m128 texSampledColor = SIMD_SampleTextureForTriangle(texture, uv1, uv2SubUv1, uv3SubUv1, barycentric);
__m128 finalColor = _mm_mul_ps(texSampledColor, simdColor);
SIMD_SetPixel(renderBuffer, posX, posY, finalColor, ColorSpace_Linear);
}
}
signedArea1 = _mm_add_ps(signedArea1, signedAreaPixelDeltaX);
}
// Rasterise buffer(X + 1, Y) pixel
@ -1066,60 +1200,73 @@ inline void SIMDRasteriseTexturedTriangle(DTRRenderBuffer *const renderBuffer, c
if (pixelZValue > currZValue)
{
renderBuffer->zBuffer[zBufferIndex] = pixelZValue;
f32 barycentricB = ((f32 *)&barycentric)[2];
f32 barycentricC = ((f32 *)&barycentric)[0];
DqnV2 uv = uv1 + (uv2SubUv1 * barycentricB) + (uv3SubUv1 * barycentricC);
const f32 EPSILON = 0.1f;
DQN_ASSERT(uv.x >= 0 && uv.x < 1.0f + EPSILON);
DQN_ASSERT(uv.y >= 0 && uv.y < 1.0f + EPSILON);
uv.x = DqnMath_Clampf(uv.x, 0.0f, 1.0f);
uv.y = DqnMath_Clampf(uv.y, 0.0f, 1.0f);
f32 texelXf = uv.x * texture->dim.w;
f32 texelYf = uv.y * texture->dim.h;
DQN_ASSERT(texelXf >= 0 && texelXf < texture->dim.w);
DQN_ASSERT(texelYf >= 0 && texelYf < texture->dim.h);
i32 texelX = (i32)texelXf;
i32 texelY = (i32)texelYf;
u32 texel1 = *(u32 *)(texturePtr + (texelX * texture->bytesPerPixel) + (texelY * texturePitch));
__m128 color1 = _mm_set_ps((f32)(texel1 >> 24),
(f32)((texel1 >> 16) & 0xFF),
(f32)((texel1 >> 8) & 0xFF),
(f32)((texel1 >> 0) & 0xFF));
color1 = _mm_mul_ps(color1, INV255_4X);
f32 preserveAlpha = ((f32 *)&color1)[3];
color1 = _mm_mul_ps(color1, color1); // to linear space
((f32 *)&color1)[3] = preserveAlpha;
color1 = _mm_mul_ps(color1, colorModulate);
DqnV4 blend = {};
blend.r = ((f32 *)&color1)[0];
blend.g = ((f32 *)&color1)[1];
blend.b = ((f32 *)&color1)[2];
blend.a = ((f32 *)&color1)[3];
SetPixel(renderBuffer, posX, posY, blend, ColorSpace_Linear);
__m128 texSampledColor = SIMD_SampleTextureForTriangle(texture, uv1, uv2SubUv1, uv3SubUv1, barycentric);
__m128 finalColor = _mm_mul_ps(texSampledColor, simdColor);
SIMD_SetPixel(renderBuffer, posX, posY, finalColor, ColorSpace_Linear);
}
}
signedArea2 = _mm_add_ps(signedArea2, signedAreaPixelDeltaX);
}
signedArea1 = _mm_add_ps(signedArea1, signedAreaPixelDeltaX);
signedArea2 = _mm_add_ps(signedArea2, signedAreaPixelDeltaX);
}
signedAreaPixel1 = _mm_add_ps(signedAreaPixel1, signedAreaPixelDeltaY);
signedAreaPixel2 = _mm_add_ps(signedAreaPixel2, signedAreaPixelDeltaY);
}
DTRDebug_EndCycleCount(DTRDebugCycleCount_RenderTexturedTriangle_Rasterise);
////////////////////////////////////////////////////////////////////////////
// Debug
////////////////////////////////////////////////////////////////////////////
DTRDebug_CounterIncrement(DTRDebugCounter_RenderTriangle);
if (DTR_DEBUG_RENDER)
{
DqnV2 origin = Get2DOriginFromTransformAnchor(p1.xy, p2.xy, p3.xy, transform);
// Draw Bounding box
if (0)
{
DTRRender_Line(renderBuffer, DqnV2i_2i(min.x, min.y), DqnV2i_2i(min.x, max.y), color);
DTRRender_Line(renderBuffer, DqnV2i_2i(min.x, max.y), DqnV2i_2i(max.x, max.y), color);
DTRRender_Line(renderBuffer, DqnV2i_2i(max.x, max.y), DqnV2i_2i(max.x, min.y), color);
DTRRender_Line(renderBuffer, DqnV2i_2i(max.x, min.y), DqnV2i_2i(min.x, min.y), color);
}
// Draw Triangle Coordinate Basis
if (0)
{
DqnV2 xAxis = DqnV2_2f(cosf(transform.rotation), sinf(transform.rotation)) * transform.scale.x;
DqnV2 yAxis = DqnV2_2f(-xAxis.y, xAxis.x) * transform.scale.y;
DqnV4 coordSysColor = DqnV4_4f(0, 1, 1, 1);
i32 axisLen = 50;
DTRRender_Line(renderBuffer, DqnV2i_V2(origin), DqnV2i_V2(origin) + DqnV2i_V2(xAxis * axisLen), coordSysColor);
DTRRender_Line(renderBuffer, DqnV2i_V2(origin), DqnV2i_V2(origin) + DqnV2i_V2(yAxis * axisLen), coordSysColor);
}
// Draw axis point
if (0)
{
DqnV4 green = DqnV4_4f(0, 1, 0, 1);
DqnV4 blue = DqnV4_4f(0, 0, 1, 1);
DqnV4 purple = DqnV4_4f(1, 0, 1, 1);
DTRRender_Rectangle(renderBuffer, p1.xy - DqnV2_1f(5), p1.xy + DqnV2_1f(5), green);
DTRRender_Rectangle(renderBuffer, p2.xy - DqnV2_1f(5), p2.xy + DqnV2_1f(5), blue);
DTRRender_Rectangle(renderBuffer, p3.xy - DqnV2_1f(5), p3.xy + DqnV2_1f(5), purple);
}
}
}
void DTRRender_TexturedTriangle(DTRRenderBuffer *const renderBuffer, DqnV3 p1, DqnV3 p2, DqnV3 p3,
DqnV2 uv1, DqnV2 uv2, DqnV2 uv3, DTRBitmap *const texture,
DqnV4 color, const DTRRenderTransform transform)
{
if (globalDTRPlatformFlags.canUseSSE2)
{
SIMD_TexturedTriangle(renderBuffer, p1, p2, p3, uv1, uv2, uv3, texture, color, transform);
return;
}
DTR_DEBUG_EP_TIMED_FUNCTION();
////////////////////////////////////////////////////////////////////////////
// Transform vertexes
@ -1149,18 +1296,11 @@ void DTRRender_TexturedTriangle(DTRRenderBuffer *const renderBuffer, DqnV3 p1, D
DQN_SWAP(DqnV3, p2, p3);
}
DTRDebug_BeginCycleCount(DTRDebugCycleCount_RenderTexturedTriangle_Rasterise);
////////////////////////////////////////////////////////////////////////////
// Scan and Render
////////////////////////////////////////////////////////////////////////////
if (globalDTRPlatformFlags.canUseSSE2)
{
SIMDRasteriseTexturedTriangle(renderBuffer, p1, p2, p3, uv1, uv2, uv3, texture, color);
}
else
{
RasteriseTexturedTriangle(renderBuffer, p1, p2, p3, uv1, uv2, uv3, texture, color);
}
DTRDebug_BeginCycleCount(DTRDebugCycleCount_RenderTexturedTriangle_Rasterise);
RasteriseTexturedTriangle(renderBuffer, p1, p2, p3, uv1, uv2, uv3, texture, color);
DTRDebug_EndCycleCount(DTRDebugCycleCount_RenderTexturedTriangle_Rasterise);
////////////////////////////////////////////////////////////////////////////
@ -1250,12 +1390,14 @@ void DTRRender_Mesh(DTRRenderBuffer *const renderBuffer, DTRMesh *const mesh, co
// TODO(doyle): Why do we need rounding here? Maybe it's because
// I don't do any interpolation in the triangle routine for jagged
// edges.
#if 1
screenVA.x = (f32)(i32)(screenVA.x + 0.5f);
screenVA.y = (f32)(i32)(screenVA.y + 0.5f);
screenVB.x = (f32)(i32)(screenVB.x + 0.5f);
screenVB.y = (f32)(i32)(screenVB.y + 0.5f);
screenVC.x = (f32)(i32)(screenVC.x + 0.5f);
screenVC.y = (f32)(i32)(screenVC.y + 0.5f);
#endif
i32 textureAIndex = face.texIndex[0];
i32 textureBIndex = face.texIndex[1];
@ -1295,7 +1437,7 @@ void DTRRender_Mesh(DTRRenderBuffer *const renderBuffer, DTRMesh *const mesh, co
FILE_SCOPE inline void SIMDRasteriseTriangle(DTRRenderBuffer *const renderBuffer,
const SIMDTriangleInclusionTest simdTri,
const i32 posX, const i32 posY, const DqnV4 color,
const i32 posX, const i32 posY, DqnV4 color,
__m128 *const signedArea)
{
__m128 ZERO_4X = _mm_set_ps1(0.0f);
@ -1316,7 +1458,52 @@ FILE_SCOPE inline void SIMDRasteriseTriangle(DTRRenderBuffer *const renderBuffer
if (pixelZValue > currZValue)
{
renderBuffer->zBuffer[zBufferIndex] = pixelZValue;
#if 1
// NOTE: Supersampling
const i32 NUM_SUB_PIXEL = 8;
const f32 STEP_SIZE = 1.0f / NUM_SUB_PIXEL;
const f32 COVERAGE_GRANULARITY = 1.0f / (f32)DQN_SQUARED(NUM_SUB_PIXEL);
f32 coverage = 0;
const __m128 STEP_SIZE_4X = _mm_set_ps1(STEP_SIZE);
const __m128 SUB_PIXEL_DELTA_X = _mm_mul_ps(simdTri.signedAreaPixelDeltaX, STEP_SIZE_4X);
const __m128 SUB_PIXEL_DELTA_Y = _mm_mul_ps(simdTri.signedAreaPixelDeltaY, STEP_SIZE_4X);
DqnV2 sample = DqnV2_2f(posX + (0.5f / NUM_SUB_PIXEL),
posY + (0.5f / NUM_SUB_PIXEL));
f32 resultP1 = Triangle2TimesSignedArea(simdTri.p2.xy, simdTri.p3.xy, sample);
f32 resultP2 = Triangle2TimesSignedArea(simdTri.p3.xy, simdTri.p1.xy, sample);
f32 resultP3 = Triangle2TimesSignedArea(simdTri.p1.xy, simdTri.p2.xy, sample);
__m128 startSubPixel = _mm_set_ps(0, resultP3, resultP2, resultP1);
__m128 checkPixel = startSubPixel;
for (i32 subY = 0; subY < NUM_SUB_PIXEL; subY++)
{
checkPixel = startSubPixel;
for (i32 subX = 0; subX < NUM_SUB_PIXEL; subX++)
{
resultP1 = ((f32 *)&checkPixel)[0];
resultP2 = ((f32 *)&checkPixel)[1];
resultP3 = ((f32 *)&checkPixel)[2];
if ((resultP1 >= 0 && resultP2 >= 0 && resultP3 >= 0) ||
(resultP1 < 0 && resultP2 < 0 && resultP3 < 0))
{
coverage += COVERAGE_GRANULARITY;
}
checkPixel = _mm_add_ps(checkPixel, SUB_PIXEL_DELTA_X);
}
startSubPixel = _mm_add_ps(startSubPixel, SUB_PIXEL_DELTA_Y);
}
if (coverage > 0)
{
color *= coverage;
SetPixel(renderBuffer, posX, posY, color, ColorSpace_Linear);
}
#else
SetPixel(renderBuffer, posX, posY, color, ColorSpace_Linear);
#endif
}
}
*signedArea = _mm_add_ps(*signedArea, simdTri.signedAreaPixelDeltaX);