Skip to content

Commit

Permalink
fix(batched): Assume BatchedRendererServices texture derivatives are …
Browse files Browse the repository at this point in the history 
…in st space. (#1828)

The convention in the single-point RendererServices is that the texture call returns
derivatives in st space, and they are transformed to xy space before returning
from the wrapper to RenderServices.  This change makes BatchedRendererServices
follow the same convention.

---------

Signed-off-by: Stephen Friedman <[email protected]>
  • Loading branch information
@sfriedmapixar
sfriedmapixar authored Aug 15, 2024
1 parent f1aa009 commit 28b71e7
Showing 1 changed file with 189 additions and 34 deletions.
223 changes: 189 additions & 34 deletions src/liboslexec/wide/wide_optexture.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -117,14 +117,6 @@ default_texture(BatchedRendererServices* bsr, ustring filename,
has_derivs ? (float*)&dresultds_simd : NULL,
has_derivs ? (float*)&dresultdt_simd : NULL);

OIIO::simd::vfloat4 dresultdx_simd;
OIIO::simd::vfloat4 dresultdy_simd;
if (has_derivs) {
// Correct our st texture space gradients into xy-space gradients
dresultdx_simd = dresultds_simd * dsdx + dresultdt_simd * dtdx;
dresultdy_simd = dresultds_simd * dsdy + dresultdt_simd * dtdy;
}

// NOTE: regardless of the value of "retVal" we will always copy over the texture system's results.
// We are relying on the texture system properly filling in missing or fill colors

Expand All @@ -142,10 +134,10 @@ default_texture(BatchedRendererServices* bsr, ustring filename,
MaskedDx<Color3> resultDx(resultRef);
MaskedDy<Color3> resultDy(resultRef);

resultDx[lane] = Color3(dresultdx_simd[0], dresultdx_simd[1],
dresultdx_simd[2]);
resultDy[lane] = Color3(dresultdy_simd[0], dresultdy_simd[1],
dresultdy_simd[2]);
resultDx[lane] = Color3(dresultds_simd[0], dresultds_simd[1],
dresultds_simd[2]);
resultDy[lane] = Color3(dresultdt_simd[0], dresultdt_simd[1],
dresultdt_simd[2]);
}
} else if (Masked<float>::is(resultRef)) {
alphaChannelIndex = 1;
Expand All @@ -154,8 +146,8 @@ default_texture(BatchedRendererServices* bsr, ustring filename,
MaskedDy<float> resultDy(resultRef);
result[lane] = result_simd[0];
if (resultRef.has_derivs()) {
resultDx[lane] = dresultdx_simd[0];
resultDy[lane] = dresultdy_simd[0];
resultDx[lane] = dresultds_simd[0];
resultDy[lane] = dresultdt_simd[0];
}
}

Expand All @@ -165,8 +157,8 @@ default_texture(BatchedRendererServices* bsr, ustring filename,
if (alphaRef.has_derivs()) {
MaskedDx<float> alphaDx(alphaRef);
MaskedDy<float> alphaDy(alphaRef);
alphaDx[lane] = dresultdx_simd[alphaChannelIndex];
alphaDy[lane] = dresultdy_simd[alphaChannelIndex];
alphaDx[lane] = dresultds_simd[alphaChannelIndex];
alphaDy[lane] = dresultdt_simd[alphaChannelIndex];
}
}
//std::cout << "s: " << s.get(i) << " t: " << t.get(i) << " color: " << resultColor << " " << wideResult.get(i) << std::endl;
Expand Down Expand Up @@ -311,16 +303,6 @@ default_texture3d(BatchedRendererServices* bsr, ustring filename,
has_derivs ? (float*)&dresultdt_simd : nullptr,
has_derivs ? (float*)&dresultdr_simd : nullptr);

OIIO::simd::vfloat4 dresultdx_simd;
OIIO::simd::vfloat4 dresultdy_simd;
if (has_derivs) {
// Correct our str texture space gradients into xyz-space gradients
dresultdx_simd = dresultds_simd * dPdx.x + dresultdt_simd * dPdx.y
+ dresultdr_simd * dPdx.z;
dresultdy_simd = dresultds_simd * dPdy.x + dresultdt_simd * dPdy.y
+ dresultdr_simd * dPdy.z;
}

// NOTE: regardless of the value of "retVal" we will always copy over the texture system's results.
// We are relying on the texture system properly filling in missing or fill colors

Expand All @@ -337,10 +319,10 @@ default_texture3d(BatchedRendererServices* bsr, ustring filename,
if (resultRef.has_derivs()) {
MaskedDx<Color3> resultDx(resultRef);
MaskedDy<Color3> resultDy(resultRef);
resultDx[lane] = Color3(dresultdx_simd[0], dresultdx_simd[1],
dresultdx_simd[2]);
resultDy[lane] = Color3(dresultdy_simd[0], dresultdy_simd[1],
dresultdy_simd[2]);
resultDx[lane] = Color3(dresultds_simd[0], dresultds_simd[1],
dresultds_simd[2]);
resultDy[lane] = Color3(dresultdt_simd[0], dresultdt_simd[1],
dresultdt_simd[2]);
}
} else if (Masked<float>::is(resultRef)) {
alphaChannelIndex = 1;
Expand All @@ -349,8 +331,8 @@ default_texture3d(BatchedRendererServices* bsr, ustring filename,
if (resultRef.has_derivs()) {
MaskedDx<float> resultDx(resultRef);
MaskedDy<float> resultDy(resultRef);
resultDx[lane] = dresultdx_simd[0];
resultDy[lane] = dresultdy_simd[0];
resultDx[lane] = dresultds_simd[0];
resultDy[lane] = dresultdt_simd[0];
}
}

Expand All @@ -361,8 +343,8 @@ default_texture3d(BatchedRendererServices* bsr, ustring filename,
if (alphaRef.has_derivs()) {
MaskedDx<float> alphaDx(alphaRef);
MaskedDy<float> alphaDy(alphaRef);
alphaDx[lane] = dresultdx_simd[alphaChannelIndex];
alphaDy[lane] = dresultdy_simd[alphaChannelIndex];
alphaDx[lane] = dresultds_simd[alphaChannelIndex];
alphaDy[lane] = dresultdt_simd[alphaChannelIndex];
}
}

Expand Down Expand Up @@ -561,6 +543,164 @@ dispatch_environment(BatchedRendererServices* bsr, ustring filename,
} // namespace


static OSL_NOINLINE void
transformWideTextureGradients(BatchedTextureOutputs& outputs,
Wide<const float> dsdx, Wide<const float> dtdx,
Wide<const float> dsdy, Wide<const float> dtdy)
{
MaskedData resultRef = outputs.result();
if (resultRef.valid() && resultRef.has_derivs()) {
if (Masked<float>::is(resultRef)) {
OSL_FORCEINLINE_BLOCK
{
MaskedDx<float> drds(resultRef);
MaskedDy<float> drdt(resultRef);

OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
for (int i = 0; i < __OSL_WIDTH; ++i) {
float drdsVal = drds[i];
float drdtVal = drdt[i];
float drdx = drdsVal * dsdx[i] + drdtVal * dtdx[i];
float drdy = drdsVal * dsdy[i] + drdtVal * dtdy[i];
drds[i] = drdx;
drdt[i] = drdy;
}
}
} else {
// keep assert out of inlined code
OSL_DASSERT(Masked<Color3>::is(resultRef));
OSL_FORCEINLINE_BLOCK
{
//printf("doint color\n");
MaskedDx<Color3> widedrds(resultRef);
MaskedDy<Color3> widedrdt(resultRef);
OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
for (int i = 0; i < __OSL_WIDTH; ++i) {
Color3 drdsColor = widedrds[i];
Color3 drdtColor = widedrdt[i];

widedrds[i] = drdsColor * dsdx[i] + drdtColor * dtdx[i];
widedrdt[i] = drdsColor * dsdy[i] + drdtColor * dtdy[i];
}
}
}
}

MaskedData alphaRef = outputs.alpha();
OSL_FORCEINLINE_BLOCK
if (alphaRef.valid() && alphaRef.has_derivs()) {
MaskedDx<float> dads(alphaRef);
MaskedDy<float> dadt(alphaRef);
OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
for (int i = 0; i < __OSL_WIDTH; ++i) {
float dadsVal = dads[i];
float dadtVal = dadt[i];
float dadx = dadsVal * dsdx[i] + dadtVal * dtdx[i];
float dady = dadsVal * dsdy[i] + dadtVal * dtdy[i];
dads[i] = dadx;
dadt[i] = dady;
}
}
}

static OSL_NOINLINE void
transformWideTextureGradientsTexture3d(BatchedTextureOutputs& outputs,
Wide<const Vec3> Pdx,
Wide<const Vec3> Pdy,
Wide<const Vec3> Pdz)
{
MaskedData resultRef = outputs.result();
if (resultRef.valid() && resultRef.has_derivs()) {
if (Masked<float>::is(resultRef)) {
OSL_FORCEINLINE_BLOCK
{
MaskedDx<float> drds(resultRef);
MaskedDy<float> drdt(resultRef);
//MaskedDz<float> drdr(resultRef); // our duals don't actually have space for this

OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
for (int i = 0; i < __OSL_WIDTH; ++i) {
float dres_xVal = drds[i];
float dres_yVal = drdt[i];
//float dres_zVal = drdr[i];

Vec3 v3pdx = Pdx[i];
Vec3 v3pdy = Pdy[i];
//Vec3 v3pdz = Pdz[i];

float dres_x = dres_xVal * v3pdx.x
+ dres_yVal
* v3pdx.y; // + dres_zVal * v3pdx.z;
float dres_y = dres_xVal * v3pdy.x
+ dres_yVal
* v3pdy.y; // + dres_zVal * v3pdy.z;
//float dres_z = dres_xVal * v3pdz.x + dres_yVal * v3pdz.y + dres_zVal * v3pdz.z;

drds[i] = dres_x;
drdt[i] = dres_y;
//drdr[i] = dres_z;
}
}
} else {
// keep assert out of inlined code
OSL_DASSERT(Masked<Color3>::is(resultRef));
OSL_FORCEINLINE_BLOCK
{
MaskedDx<Color3> widedrp1(resultRef);
MaskedDy<Color3> widedrp2(resultRef);
//MaskedDz<Color3> widedrp3(resultRef);

OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
for (int i = 0; i < __OSL_WIDTH; ++i) {
Color3 drdp1Color = widedrp1[i];
Color3 drdp2Color = widedrp2[i];
//Color3 drdp3Color = widedrp3[i];

Vec3 v3pdx = Pdx[i];
Vec3 v3pdy = Pdy[i];
//Vec3 v3pdz = Pdz[i];

widedrp1[i] = drdp1Color * v3pdx.x
+ drdp2Color
* v3pdx.y; // + drdp3Color * v3pdx.z;
widedrp2[i] = drdp1Color * v3pdy.x
+ drdp2Color
* v3pdy.y; // + drdp3Color * v3pdy.z;
//widedrp3[i] = drdp1Color * v3pdz.x + drdp2Color * v3pdz.y + drdp3Color * v3pdz.z;
}
}
}
}

MaskedData alphaRef = outputs.alpha();
OSL_FORCEINLINE_BLOCK
if (alphaRef.valid() && alphaRef.has_derivs()) {
MaskedDx<float> dap1(alphaRef);
MaskedDy<float> dap2(alphaRef);
// MaskedDz<float> dap3(alphaRef);

OSL_OMP_PRAGMA(omp simd simdlen(__OSL_WIDTH))
for (int i = 0; i < __OSL_WIDTH; ++i) {
float dadp1Val = dap1[i];
float dadp2Val = dap2[i];
//float dadp3Val = dap3[i];

Vec3 v3pdx = Pdx[i];
Vec3 v3pdy = Pdy[i];
//Vec3 v3pdz = Pdz[i];

float dadpx = dadp1Val * v3pdx.x
+ dadp2Val * v3pdx.y; // + dadp3Val * v3pdx.z;
float dadpy = dadp1Val * v3pdy.x
+ dadp2Val * v3pdy.y; // + dadp3Val * v3pdy.z;
//float dadpz = dadp1Val * v3pdz.x + dadp2Val * v3pdz.y + dadp3Val * v3pdz.z;

dap1[i] = dadpx;
dap2[i] = dadpy;
//dap3[i] = dadpz;
}
}
}

OSL_BATCHOP int
__OSL_MASKED_OP(texture)(void* bsg_, ustring_pod name_, void* handle,
Expand Down Expand Up @@ -590,6 +730,14 @@ __OSL_MASKED_OP(texture)(void* bsg_, ustring_pod name_, void* handle,
Wide<const float>(dsdy), Wide<const float>(dtdy),
outputs);

// Correct our st texture space gradients into xy-space gradients
if (resultHasDerivs || alphaHasDerivs) {
transformWideTextureGradients(outputs, Wide<const float>(dsdx),
Wide<const float>(dtdx),
Wide<const float>(dsdy),
Wide<const float>(dtdy));
}

OSL_FORCEINLINE_BLOCK
if (outputs.errormessage().valid()) {
Masked<ustring> err(outputs.errormessage());
Expand Down Expand Up @@ -637,6 +785,13 @@ __OSL_MASKED_OP(texture3d)(void* bsg_, ustring_pod name_, void* handle,
Wide<const Vec3>(wPdy), Wide<const Vec3>(wPdz),
outputs);

// Correct our P (Vec3) space gradients into xyz-space gradients
if (resultHasDerivs || alphaHasDerivs) {
transformWideTextureGradientsTexture3d(outputs, Wide<const Vec3>(wPdx),
Wide<const Vec3>(wPdy),
Wide<const Vec3>(wPdz));
}

OSL_FORCEINLINE_BLOCK
if (outputs.errormessage().valid()) {
Masked<ustring> err(outputs.errormessage());
Expand Down

0 comments on commit 28b71e7

Please sign in to comment.