Page MenuHome
Differential D14295 Diff 49287 intern/cycles/kernel/device/cpu/image.h

Changeset View

Standalone View

View Options

intern/cycles/kernel/device/cpu/image.h

Show All 9 Lines
#endif #endif
CCL_NAMESPACE_BEGIN CCL_NAMESPACE_BEGIN
/* Make template functions private so symbols don't conflict between kernels with different /* Make template functions private so symbols don't conflict between kernels with different
* instruction sets. */ * instruction sets. */
namespace { namespace {
#define ERROR_COLOR make_float4(0.0f, 0.0f, 0.0f, 0.0f)
#define SET_CUBIC_SPLINE_WEIGHTS(u, t) \ #define SET_CUBIC_SPLINE_WEIGHTS(u, t) \
{ \ { \
u[0] = (((-1.0f / 6.0f) * t + 0.5f) * t - 0.5f) * t + (1.0f / 6.0f); \ u[0] = (((-1.0f / 6.0f) * t + 0.5f) * t - 0.5f) * t + (1.0f / 6.0f); \
u[1] = ((0.5f * t - 1.0f) * t) * t + (2.0f / 3.0f); \ u[1] = ((0.5f * t - 1.0f) * t) * t + (2.0f / 3.0f); \
u[2] = ((-0.5f * t + 0.5f) * t + 0.5f) * t + (1.0f / 6.0f); \ u[2] = ((-0.5f * t + 0.5f) * t + 0.5f) * t + (1.0f / 6.0f); \
u[3] = (1.0f / 6.0f) * t * t * t; \ u[3] = (1.0f / 6.0f) * t * t * t; \
} \ } \
(void)0 (void)0
▲ Show 20 Lines • Show All 49 Lines • ▼ Show 20 Linestemplate<typename T> struct TextureInterpolator {
} }
static ccl_always_inline float4 read(ushort4 r) static ccl_always_inline float4 read(ushort4 r)
{ {
float f = 1.0f / 65535.0f; float f = 1.0f / 65535.0f;
return make_float4(r.x * f, r.y * f, r.z * f, r.w * f); return make_float4(r.x * f, r.y * f, r.z * f, r.w * f);
} }
/** Read 2D Texture Data
* -Does not check if data request is in bounds.
*/
static ccl_always_inline float4 read(const T *data, int x, int y, int width, int height) static ccl_always_inline float4 read(const T *data, int x, int y, int width, int height)
{ {
if (x < 0 || y < 0 || x >= width || y >= height) { return read(data[y * width + x]);
}
/** Read 2D Texture Data Clip
* -Returns transparent black if data request is out of bounds.
*/
static ccl_always_inline float4 read_clip(const T *data, int x, int y, int width, int height)
{
if (x < 0 || x >= width || y < 0 || y >= height) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
} }
return read(data[y * width + x]); return read(data[y * width + x]);
} }
/** Read 3D Texture Data
* -Does not check if data request is in bounds.
*/
static ccl_always_inline float4
read(const T *data, int x, int y, int z, int width, int height, int depth)
{
return read(data[x + y * width + z * width * height]);
}
/** Read 3D Texture Data Clip
* -Returns transparent black if data request is out of bounds.
*/
static ccl_always_inline float4
read_clip(const T *data, int x, int y, int z, int width, int height, int depth)
{
if (x < 0 || x >= width || y < 0 || y >= height || z < 0 || z >= depth) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
}
return read(data[x + y * width + z * width * height]);
}
typedef float4 (*ReadType3DTexture)(const T *, int, int, int, int, int, int);
/* Trilinear Interpolation */
static ccl_always_inline float4
trilenear_lookup(const T *data,
float tx,
float ty,
float tz,
int ix,
int iy,
int iz,
int nix,
int niy,
int niz,
int width,
int height,
int depth,
float4 read(const T *, int, int, int, int, int, int))
{
float4 r;
r = (1.0f - tz) * (1.0f - ty) * (1.0f - tx) * read(data, ix, iy, iz, width, height, depth);
r += (1.0f - tz) * (1.0f - ty) * tx * read(data, nix, iy, iz, width, height, depth);
r += (1.0f - tz) * ty * (1.0f - tx) * read(data, ix, niy, iz, width, height, depth);
r += (1.0f - tz) * ty * tx * read(data, nix, niy, iz, width, height, depth);
r += tz * (1.0f - ty) * (1.0f - tx) * read(data, ix, iy, niz, width, height, depth);
r += tz * (1.0f - ty) * tx * read(data, nix, iy, niz, width, height, depth);
r += tz * ty * (1.0f - tx) * read(data, ix, niy, niz, width, height, depth);
r += tz * ty * tx * read(data, nix, niy, niz, width, height, depth);
return r;
}
/** Tricubic Interpolation */
static ccl_always_inline float4
tricubic_lookup(const T *data,
float tx,
float ty,
float tz,
const int xc[4],
const int yc[4],
const int zc[4],
int width,
int height,
int depth,
float4 read(const T *, int, int, int, int, int, int))
{
float u[4], v[4], w[4];
/* Some helper macros to keep code size reasonable.
* Lets the compiler inline all the matrix multiplications.
*/
#define DATA(x, y, z) (read(data, xc[x], yc[y], zc[z], width, height, depth))
#define COL_TERM(col, row) \
(v[col] * (u[0] * DATA(0, col, row) + u[1] * DATA(1, col, row) + u[2] * DATA(2, col, row) + \
u[3] * DATA(3, col, row)))
#define ROW_TERM(row) \
(w[row] * (COL_TERM(0, row) + COL_TERM(1, row) + COL_TERM(2, row) + COL_TERM(3, row)))
SET_CUBIC_SPLINE_WEIGHTS(u, tx);
SET_CUBIC_SPLINE_WEIGHTS(v, ty);
SET_CUBIC_SPLINE_WEIGHTS(w, tz);
/* Actual interpolation. */
return ROW_TERM(0) + ROW_TERM(1) + ROW_TERM(2) + ROW_TERM(3);
#undef COL_TERM
#undef ROW_TERM
#undef DATA
}
static ccl_always_inline int wrap_periodic(int x, int width) static ccl_always_inline int wrap_periodic(int x, int width)
{ {
x %= width; x %= width;
if (x < 0) if (x < 0) {
x += width; x += width;
}
return x; return x;
} }
static ccl_always_inline int wrap_clamp(int x, int width) static ccl_always_inline int wrap_clamp(int x, int width)
{ {
return clamp(x, 0, width - 1); return clamp(x, 0, width - 1);
} }
/* ******** 2D interpolation ******** */ /* ******** 2D interpolation ******** */
static ccl_always_inline float4 interp_closest(const TextureInfo &info, float x, float y) static ccl_always_inline float4 interp_closest(const TextureInfo &info, float x, float y)
{ {
const T *data = (const T *)info.data;
const int width = info.width;
const int height = info.height;
int ix, iy; int ix, iy;
frac(x * (float)width, &ix);
frac(y * (float)height, &iy); frac(x * (float)info.width, &ix);
frac(y * (float)info.height, &iy);
switch (info.extension) { switch (info.extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, info.width);
iy = wrap_periodic(iy, height); iy = wrap_periodic(iy, info.height);
break; break;
case EXTENSION_CLIP: case EXTENSION_CLIP:
if (x < 0.0f || y < 0.0f || x > 1.0f || y > 1.0f) { if (x < 0.0f || x >= 1.0f || y < 0.0f || y >= 1.0f) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
} }
ATTR_FALLTHROUGH; break;
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
ix = wrap_clamp(ix, width); ix = wrap_clamp(ix, info.width);
iy = wrap_clamp(iy, height); iy = wrap_clamp(iy, info.height);
break; break;
default: default:
kernel_assert(0); kernel_assert(0);
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return ERROR_COLOR;
} }
return read(data[ix + iy * width]); const T *data = (const T *)info.data;
return read(data[ix + iy * info.width]);
} }
static ccl_always_inline float4 interp_linear(const TextureInfo &info, float x, float y) static ccl_always_inline float4 interp_linear(const TextureInfo &info, float x, float y)
{ {
const T *data = (const T *)info.data; int ix, iy;
const int width = info.width; int nix, niy;
const int height = info.height;
int ix, iy, nix, niy; /* A -0.5 offset is used to center the linear samples around the sample point. */
const float tx = frac(x * (float)width - 0.5f, &ix); const float tx = frac(x * (float)info.width - 0.5f, &ix);
const float ty = frac(y * (float)height - 0.5f, &iy); const float ty = frac(y * (float)info.height - 0.5f, &iy);
switch (info.extension) { switch (info.extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, info.width);
iy = wrap_periodic(iy, height); nix = wrap_periodic(ix + 1, info.width);
nix = wrap_periodic(ix + 1, width);
niy = wrap_periodic(iy + 1, height); iy = wrap_periodic(iy, info.height);
niy = wrap_periodic(iy + 1, info.height);
break; break;
case EXTENSION_CLIP: case EXTENSION_CLIP:
if (ix < -1 || ix >= info.width || iy < -1 || iy >= info.height) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
}
nix = ix + 1; nix = ix + 1;
niy = iy + 1; niy = iy + 1;
break; break;
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
nix = wrap_clamp(ix + 1, width); nix = ix + 1;
niy = wrap_clamp(iy + 1, height); if (ix < 0) {
ix = wrap_clamp(ix, width); ix = 0;
iy = wrap_clamp(iy, height); nix = 0;
}
else if (nix >= info.width) {
ix = info.width - 1;
nix = info.width - 1;
}
niy = iy + 1;
if (iy < 0) {
iy = 0;
niy = 0;
}
else if (niy >= info.height) {
iy = info.height - 1;
niy = info.height - 1;
}
break; break;
default: default:
kernel_assert(0); kernel_assert(0);
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return ERROR_COLOR;
} }
return (1.0f - ty) * (1.0f - tx) * read(data, ix, iy, width, height) + const T *data = (const T *)info.data;
(1.0f - ty) * tx * read(data, nix, iy, width, height) + return (1.0f - ty) * (1.0f - tx) * read_clip(data, ix, iy, info.width, info.height) +
ty * (1.0f - tx) * read(data, ix, niy, width, height) + (1.0f - ty) * tx * read_clip(data, nix, iy, info.width, info.height) +
ty * tx * read(data, nix, niy, width, height); ty * (1.0f - tx) * read_clip(data, ix, niy, info.width, info.height) +
ty * tx * read_clip(data, nix, niy, info.width, info.height);
} }
static ccl_always_inline float4 interp_cubic(const TextureInfo &info, float x, float y) static ccl_always_inline float4 interp_cubic(const TextureInfo &info, float x, float y)
{ {
const T *data = (const T *)info.data; int pix, piy;
const int width = info.width; int ix, iy;
const int height = info.height; int nix, niy;
int ix, iy, nix, niy; int nnix, nniy;
const float tx = frac(x * (float)width - 0.5f, &ix);
const float ty = frac(y * (float)height - 0.5f, &iy); /* A -0.5 offset is used to center the cubic samples around the sample point. */
int pix, piy, nnix, nniy; const float tx = frac(x * (float)info.width - 0.5f, &ix);
const float ty = frac(y * (float)info.height - 0.5f, &iy);
switch (info.extension) { switch (info.extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, info.width);
iy = wrap_periodic(iy, height); pix = wrap_periodic(ix - 1, info.width);
pix = wrap_periodic(ix - 1, width); nix = wrap_periodic(ix + 1, info.width);
piy = wrap_periodic(iy - 1, height); nnix = wrap_periodic(ix + 2, info.width);
nix = wrap_periodic(ix + 1, width);
niy = wrap_periodic(iy + 1, height); iy = wrap_periodic(iy, info.height);
nnix = wrap_periodic(ix + 2, width); piy = wrap_periodic(iy - 1, info.height);
nniy = wrap_periodic(iy + 2, height); niy = wrap_periodic(iy + 1, info.height);
nniy = wrap_periodic(iy + 2, info.height);
break; break;
case EXTENSION_CLIP: case EXTENSION_CLIP: {
if (ix < -2 || ix > info.width || iy < -2 || iy > info.height) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
}
pix = ix - 1; pix = ix - 1;
piy = iy - 1;
nix = ix + 1; nix = ix + 1;
niy = iy + 1;
nnix = ix + 2; nnix = ix + 2;
piy = iy - 1;
niy = iy + 1;
nniy = iy + 2; nniy = iy + 2;
break; }
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
pix = wrap_clamp(ix - 1, width); pix = wrap_clamp(ix - 1, info.width);
piy = wrap_clamp(iy - 1, height); nix = wrap_clamp(ix + 1, info.width);
nix = wrap_clamp(ix + 1, width); nnix = wrap_clamp(ix + 2, info.width);
niy = wrap_clamp(iy + 1, height); ix = wrap_clamp(ix, info.width);
nnix = wrap_clamp(ix + 2, width);
nniy = wrap_clamp(iy + 2, height); piy = wrap_clamp(iy - 1, info.height);
ix = wrap_clamp(ix, width); niy = wrap_clamp(iy + 1, info.height);
iy = wrap_clamp(iy, height); nniy = wrap_clamp(iy + 2, info.height);
iy = wrap_clamp(iy, info.height);
break; break;
default: default:
kernel_assert(0); kernel_assert(0);
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return ERROR_COLOR;
} }
const T *data = (const T *)info.data;
const int xc[4] = {pix, ix, nix, nnix}; const int xc[4] = {pix, ix, nix, nnix};
const int yc[4] = {piy, iy, niy, nniy}; const int yc[4] = {piy, iy, niy, nniy};
float u[4], v[4]; float u[4], v[4];
/* Some helper macro to keep code reasonable size,
* let compiler to inline all the matrix multiplications. /* Some helper macros to keep code size reasonable.
* Lets the compiler inline all the matrix multiplications.
*/ */
#define DATA(x, y) (read(data, xc[x], yc[y], width, height)) #define DATA(x, y) (read_clip(data, xc[x], yc[y], info.width, info.height))
#define TERM(col) \ #define TERM(col) \
(v[col] * \ (v[col] * \
(u[0] * DATA(0, col) + u[1] * DATA(1, col) + u[2] * DATA(2, col) + u[3] * DATA(3, col))) (u[0] * DATA(0, col) + u[1] * DATA(1, col) + u[2] * DATA(2, col) + u[3] * DATA(3, col)))
SET_CUBIC_SPLINE_WEIGHTS(u, tx); SET_CUBIC_SPLINE_WEIGHTS(u, tx);
SET_CUBIC_SPLINE_WEIGHTS(v, ty); SET_CUBIC_SPLINE_WEIGHTS(v, ty);
/* Actual interpolation. */ /*Actual interpolation.*/
return TERM(0) + TERM(1) + TERM(2) + TERM(3); return TERM(0) + TERM(1) + TERM(2) + TERM(3);
#undef TERM #undef TERM
#undef DATA #undef DATA
} }
static ccl_always_inline float4 interp(const TextureInfo &info, float x, float y) static ccl_always_inline float4 interp(const TextureInfo &info, float x, float y)
{ {
if (UNLIKELY(!info.data)) { if (UNLIKELY(!info.data)) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return ERROR_COLOR;
} }
switch (info.interpolation) { switch (info.interpolation) {
case INTERPOLATION_CLOSEST: case INTERPOLATION_CLOSEST:
return interp_closest(info, x, y); return interp_closest(info, x, y);
case INTERPOLATION_LINEAR: case INTERPOLATION_LINEAR:
return interp_linear(info, x, y); return interp_linear(info, x, y);
default: default:
return interp_cubic(info, x, y); return interp_cubic(info, x, y);
} }
} }
/* ******** 3D interpolation ******** */ /* ******** 3D interpolation ******** */
static ccl_always_inline float4 interp_3d_closest(const TextureInfo &info, static ccl_always_inline float4 interp_3d_closest(const TextureInfo &info,
float x, float x,
float y, float y,
float z) float z)
{ {
int width = info.width;
int height = info.height;
int depth = info.depth;
int ix, iy, iz; int ix, iy, iz;
frac(x * (float)width, &ix); frac(x * (float)info.width, &ix);
frac(y * (float)height, &iy); frac(y * (float)info.height, &iy);
frac(z * (float)depth, &iz); frac(z * (float)info.depth, &iz);
switch (info.extension) { switch (info.extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, info.width);
iy = wrap_periodic(iy, height); iy = wrap_periodic(iy, info.height);
iz = wrap_periodic(iz, depth); iz = wrap_periodic(iz, info.depth);
break; break;
case EXTENSION_CLIP: case EXTENSION_CLIP:
if (x < 0.0f || y < 0.0f || z < 0.0f || x > 1.0f || y > 1.0f || z > 1.0f) { if (x < 0.0f || x >= 1.0f || y < 0.0f || y >= 1.0f || z < 0.0f || z >= 1.0f) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
} }
ATTR_FALLTHROUGH; break;
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
ix = wrap_clamp(ix, width); ix = wrap_clamp(ix, info.width);
iy = wrap_clamp(iy, height); iy = wrap_clamp(iy, info.height);
iz = wrap_clamp(iz, depth); iz = wrap_clamp(iz, info.depth);
break; break;
default: default:
kernel_assert(0); kernel_assert(0);
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return ERROR_COLOR;
} }
const T *data = (const T *)info.data; return read((const T *)info.data, ix, iy, iz, info.width, info.height, info.depth);
return read(data[ix + iy * width + iz * width * height]);
} }
static ccl_always_inline float4 interp_3d_linear(const TextureInfo &info, static ccl_always_inline float4 interp_3d_linear(const TextureInfo &info,
float x, float x,
float y, float y,
float z) float z)
{ {
int width = info.width;
int height = info.height;
int depth = info.depth;
int ix, iy, iz; int ix, iy, iz;
int nix, niy, niz; int nix, niy, niz;
float tx = frac(x * (float)width - 0.5f, &ix); /* A -0.5 offset is used to center the linear samples around the sample point. */
float ty = frac(y * (float)height - 0.5f, &iy); float tx = frac(x * (float)info.width - 0.5f, &ix);
float tz = frac(z * (float)depth - 0.5f, &iz); float ty = frac(y * (float)info.height - 0.5f, &iy);
float tz = frac(z * (float)info.depth - 0.5f, &iz);
switch (info.extension) { switch (info.extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, info.width);
iy = wrap_periodic(iy, height); nix = wrap_periodic(ix + 1, info.width);
iz = wrap_periodic(iz, depth);
iy = wrap_periodic(iy, info.height);
nix = wrap_periodic(ix + 1, width); niy = wrap_periodic(iy + 1, info.height);
niy = wrap_periodic(iy + 1, height);
niz = wrap_periodic(iz + 1, depth); iz = wrap_periodic(iz, info.depth);
niz = wrap_periodic(iz + 1, info.depth);
break; break;
case EXTENSION_CLIP: case EXTENSION_CLIP:
if (x < 0.0f || y < 0.0f || z < 0.0f || x > 1.0f || y > 1.0f || z > 1.0f) { /* Case - No linear samples are inside the clip region. */
if (ix < -1 || ix >= info.width || iy < -1 || iy >= info.height || iz < -1 ||
iz >= info.depth) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
} }
ATTR_FALLTHROUGH;
nix = ix + 1;
niy = iy + 1;
niz = iz + 1;
/* Case - All linear samples are inside the clip region. */
if (ix >= 0 && nix < info.width && iy >= 0 && niy < info.height && iz >= 0 &&
niz < info.depth) {
break;
}
/* Case - The linear samples span the clip border.
* #read_clip is used to ensure proper interpolation across the clip border. */
return trilenear_lookup((const T *)info.data,
tx,
ty,
tz,
ix,
iy,
iz,
nix,
niy,
niz,
info.width,
info.height,
info.depth,
(ReadType3DTexture)&read_clip);
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
nix = wrap_clamp(ix + 1, width); nix = ix + 1;
niy = wrap_clamp(iy + 1, height); if (ix < 0) {
niz = wrap_clamp(iz + 1, depth); ix = 0;
nix = 0;
ix = wrap_clamp(ix, width); }
iy = wrap_clamp(iy, height); else if (nix >= info.width) {
iz = wrap_clamp(iz, depth); ix = info.width - 1;
nix = info.width - 1;
}
niy = iy + 1;
if (iy < 0) {
iy = 0;
niy = 0;
}
else if (niy >= info.height) {
iy = info.height - 1;
niy = info.height - 1;
}
niz = iz + 1;
if (iz < 0) {
iz = 0;
niz = iz;
}
else if (niz >= info.depth) {
iz = info.depth - 1;
niz = info.depth - 1;
}
break; break;
default: default:
kernel_assert(0); kernel_assert(0);
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return ERROR_COLOR;
} }
const T *data = (const T *)info.data; return trilenear_lookup((const T *)info.data,
float4 r; tx,
ty,
r = (1.0f - tz) * (1.0f - ty) * (1.0f - tx) * tz,
read(data[ix + iy * width + iz * width * height]); ix,
r += (1.0f - tz) * (1.0f - ty) * tx * read(data[nix + iy * width + iz * width * height]); iy,
r += (1.0f - tz) * ty * (1.0f - tx) * read(data[ix + niy * width + iz * width * height]); iz,
r += (1.0f - tz) * ty * tx * read(data[nix + niy * width + iz * width * height]); nix,
niy,
r += tz * (1.0f - ty) * (1.0f - tx) * read(data[ix + iy * width + niz * width * height]); niz,
r += tz * (1.0f - ty) * tx * read(data[nix + iy * width + niz * width * height]); info.width,
r += tz * ty * (1.0f - tx) * read(data[ix + niy * width + niz * width * height]); info.height,
r += tz * ty * tx * read(data[nix + niy * width + niz * width * height]); info.depth,
(ReadType3DTexture)&read);
return r;
} }
/* TODO(sergey): For some unspeakable reason both GCC-6 and Clang-3.9 are /* Tricubic b-spline interpolation.
*
* TODO(sergey): For some unspeakable reason both GCC-6 and Clang-3.9 are
* causing stack overflow issue in this function unless it is inlined. * causing stack overflow issue in this function unless it is inlined.
* *
* Only happens for AVX2 kernel and global __KERNEL_SSE__ vectorization * Only happens for AVX2 kernel and global __KERNEL_SSE__ vectorization
* enabled. * enabled.
*/ */
#if defined(__GNUC__) || defined(__clang__) #if defined(__GNUC__) || defined(__clang__)
static ccl_always_inline static ccl_always_inline
#else #else
static ccl_never_inline static ccl_never_inline
#endif #endif
float4 float4
interp_3d_cubic(const TextureInfo &info, float x, float y, float z) interp_3d_cubic(const TextureInfo &info, float x, float y, float z)
{ {
int width = info.width; int pix, piy, piz;
int height = info.height;
int depth = info.depth;
int ix, iy, iz; int ix, iy, iz;
int nix, niy, niz; int nix, niy, niz;
/* Tricubic b-spline interpolation. */ int nnix, nniy, nniz;
const float tx = frac(x * (float)width - 0.5f, &ix);
const float ty = frac(y * (float)height - 0.5f, &iy); /* A -0.5 offset is used to center the cubic samples around the sample point. */
const float tz = frac(z * (float)depth - 0.5f, &iz); const float tx = frac(x * (float)info.width - 0.5f, &ix);
int pix, piy, piz, nnix, nniy, nniz; const float ty = frac(y * (float)info.height - 0.5f, &iy);
const float tz = frac(z * (float)info.depth - 0.5f, &iz);
switch (info.extension) { switch (info.extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, info.width);
iy = wrap_periodic(iy, height); pix = wrap_periodic(ix - 1, info.width);
iz = wrap_periodic(iz, depth); nix = wrap_periodic(ix + 1, info.width);
nnix = wrap_periodic(ix + 2, info.width);
pix = wrap_periodic(ix - 1, width);
piy = wrap_periodic(iy - 1, height); iy = wrap_periodic(iy, info.height);
piz = wrap_periodic(iz - 1, depth); niy = wrap_periodic(iy + 1, info.height);
piy = wrap_periodic(iy - 1, info.height);
nix = wrap_periodic(ix + 1, width); nniz = wrap_periodic(iz + 2, info.depth);
niy = wrap_periodic(iy + 1, height);
niz = wrap_periodic(iz + 1, depth); iz = wrap_periodic(iz, info.depth);
piz = wrap_periodic(iz - 1, info.depth);
nnix = wrap_periodic(ix + 2, width); niz = wrap_periodic(iz + 1, info.depth);
nniy = wrap_periodic(iy + 2, height); nniy = wrap_periodic(iy + 2, info.height);
nniz = wrap_periodic(iz + 2, depth); break;
break; case EXTENSION_CLIP: {
case EXTENSION_CLIP: /* Case - No cubic samples are inside the clip region. */
if (x < 0.0f || y < 0.0f || z < 0.0f || x > 1.0f || y > 1.0f || z > 1.0f) { if (ix < -2 || ix > info.width || iy < -2 || iy > info.height || iz < -2 ||
iz > info.depth) {
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
} }
ATTR_FALLTHROUGH;
pix = ix - 1;
nnix = ix + 2;
nix = ix + 1;
piy = iy - 1;
niy = iy + 1;
nniy = iy + 2;
piz = iz - 1;
niz = iz + 1;
nniz = iz + 2;
/* Case - All linear samples are inside the clip region. */
if (pix >= 0 && nnix < info.width && piy >= 0 && nniy < info.height && piz >= 0 &&
nniz < info.depth) {
break;
}
/* Case - The Cubic samples span the clip border.
* #read_clip is used to ensure proper interpolation across the clip border. */
const int xc[4] = {pix, ix, nix, nnix};
const int yc[4] = {piy, iy, niy, nniy};
const int zc[4] = {piz, iz, niz, nniz};
return tricubic_lookup((const T *)info.data,
tx,
ty,
tz,
xc,
yc,
zc,
info.width,
info.height,
info.depth,
(ReadType3DTexture)&read_clip);
}
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
pix = wrap_clamp(ix - 1, width); pix = wrap_clamp(ix - 1, info.width);
piy = wrap_clamp(iy - 1, height); nix = wrap_clamp(ix + 1, info.width);
piz = wrap_clamp(iz - 1, depth); nnix = wrap_clamp(ix + 2, info.width);
ix = wrap_clamp(ix, info.width);
nix = wrap_clamp(ix + 1, width);
niy = wrap_clamp(iy + 1, height); piy = wrap_clamp(iy - 1, info.height);
niz = wrap_clamp(iz + 1, depth); niy = wrap_clamp(iy + 1, info.height);
nniy = wrap_clamp(iy + 2, info.height);
nnix = wrap_clamp(ix + 2, width); iy = wrap_clamp(iy, info.height);
nniy = wrap_clamp(iy + 2, height);
nniz = wrap_clamp(iz + 2, depth); piz = wrap_clamp(iz - 1, info.depth);
niz = wrap_clamp(iz + 1, info.depth);
ix = wrap_clamp(ix, width); nniz = wrap_clamp(iz + 2, info.depth);
iy = wrap_clamp(iy, height); iz = wrap_clamp(iz, info.depth);
iz = wrap_clamp(iz, depth);
break; break;
default: default:
kernel_assert(0); kernel_assert(0);
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return ERROR_COLOR;
} }
const int xc[4] = {pix, ix, nix, nnix}; const int xc[4] = {pix, ix, nix, nnix};
const int yc[4] = {width * piy, width * iy, width * niy, width * nniy}; const int yc[4] = {piy, iy, niy, nniy};
const int zc[4] = { const int zc[4] = {piz, iz, niz, nniz};
width * height * piz, width * height * iz, width * height * niz, width * height * nniz}; return tricubic_lookup((const T *)info.data,
float u[4], v[4], w[4]; tx,
ty,
/* Some helper macro to keep code reasonable size, tz,
* let compiler to inline all the matrix multiplications. xc,
*/ yc,
#define DATA(x, y, z) (read(data[xc[x] + yc[y] + zc[z]])) zc,
#define COL_TERM(col, row) \ info.width,
(v[col] * (u[0] * DATA(0, col, row) + u[1] * DATA(1, col, row) + u[2] * DATA(2, col, row) + \ info.height,
u[3] * DATA(3, col, row))) info.depth,
#define ROW_TERM(row) \ (ReadType3DTexture)&read);
(w[row] * (COL_TERM(0, row) + COL_TERM(1, row) + COL_TERM(2, row) + COL_TERM(3, row)))
SET_CUBIC_SPLINE_WEIGHTS(u, tx);
SET_CUBIC_SPLINE_WEIGHTS(v, ty);
SET_CUBIC_SPLINE_WEIGHTS(w, tz);
/* Actual interpolation. */
const T *data = (const T *)info.data;
return ROW_TERM(0) + ROW_TERM(1) + ROW_TERM(2) + ROW_TERM(3);
#undef COL_TERM
#undef ROW_TERM
#undef DATA
} }
static ccl_always_inline float4 static ccl_always_inline float4
interp_3d(const TextureInfo &info, float x, float y, float z, InterpolationType interp) interp_3d(const TextureInfo &info, float x, float y, float z, InterpolationType interp)
{ {
if (UNLIKELY(!info.data)) if (UNLIKELY(!info.data))
return make_float4(0.0f, 0.0f, 0.0f, 0.0f); return ERROR_COLOR;
switch ((interp == INTERPOLATION_NONE) ? info.interpolation : interp) { switch ((interp == INTERPOLATION_NONE) ? info.interpolation : interp) {
case INTERPOLATION_CLOSEST: case INTERPOLATION_CLOSEST:
return interp_3d_closest(info, x, y, z); return interp_3d_closest(info, x, y, z);
case INTERPOLATION_LINEAR: case INTERPOLATION_LINEAR:
return interp_3d_linear(info, x, y, z); return interp_3d_linear(info, x, y, z);
default: default:
return interp_3d_cubic(info, x, y, z); return interp_3d_cubic(info, x, y, z);
Show All 29 Lines static ccl_always_inline float4 interp_3d_linear(const AccessorType &acc,
float x, float x,
float y, float y,
float z) float z)
{ {
const nanovdb::Vec3f xyz(x - 0.5f, y - 0.5f, z - 0.5f); const nanovdb::Vec3f xyz(x - 0.5f, y - 0.5f, z - 0.5f);
return read(nanovdb::SampleFromVoxels<AccessorType, 1, false>(acc)(xyz)); return read(nanovdb::SampleFromVoxels<AccessorType, 1, false>(acc)(xyz));
} }
/* Tricubic b-spline interpolation. */
# if defined(__GNUC__) || defined(__clang__) # if defined(__GNUC__) || defined(__clang__)
static ccl_always_inline static ccl_always_inline
# else # else
static ccl_never_inline static ccl_never_inline
# endif # endif
float4 float4
interp_3d_cubic(const AccessorType &acc, float x, float y, float z) interp_3d_cubic(const AccessorType &acc, float x, float y, float z)
{ {
int ix, iy, iz; int ix, iy, iz;
int nix, niy, niz; int nix, niy, niz;
int pix, piy, piz; int pix, piy, piz;
int nnix, nniy, nniz; int nnix, nniy, nniz;
/* Tricubic b-spline interpolation. */
/* A -0.5 offset is used to center the cubic samples around the sample point. */
const float tx = frac(x - 0.5f, &ix); const float tx = frac(x - 0.5f, &ix);
const float ty = frac(y - 0.5f, &iy); const float ty = frac(y - 0.5f, &iy);
const float tz = frac(z - 0.5f, &iz); const float tz = frac(z - 0.5f, &iz);
pix = ix - 1; pix = ix - 1;
piy = iy - 1; piy = iy - 1;
piz = iz - 1; piz = iz - 1;
nix = ix + 1; nix = ix + 1;
niy = iy + 1; niy = iy + 1;
niz = iz + 1; niz = iz + 1;
nnix = ix + 2; nnix = ix + 2;
nniy = iy + 2; nniy = iy + 2;
nniz = iz + 2; nniz = iz + 2;
const int xc[4] = {pix, ix, nix, nnix}; const int xc[4] = {pix, ix, nix, nnix};
const int yc[4] = {piy, iy, niy, nniy}; const int yc[4] = {piy, iy, niy, nniy};
const int zc[4] = {piz, iz, niz, nniz}; const int zc[4] = {piz, iz, niz, nniz};
float u[4], v[4], w[4]; float u[4], v[4], w[4];
/* Some helper macro to keep code reasonable size, /* Some helper macros to keep code size reasonable.
* let compiler to inline all the matrix multiplications. * Lets the compiler inline all the matrix multiplications.
*/ */
# define DATA(x, y, z) (read(acc.getValue(nanovdb::Coord(xc[x], yc[y], zc[z])))) # define DATA(x, y, z) (read(acc.getValue(nanovdb::Coord(xc[x], yc[y], zc[z]))))
# define COL_TERM(col, row) \ # define COL_TERM(col, row) \
(v[col] * (u[0] * DATA(0, col, row) + u[1] * DATA(1, col, row) + u[2] * DATA(2, col, row) + \ (v[col] * (u[0] * DATA(0, col, row) + u[1] * DATA(1, col, row) + u[2] * DATA(2, col, row) + \
u[3] * DATA(3, col, row))) u[3] * DATA(3, col, row)))
# define ROW_TERM(row) \ # define ROW_TERM(row) \
(w[row] * (COL_TERM(0, row) + COL_TERM(1, row) + COL_TERM(2, row) + COL_TERM(3, row))) (w[row] * (COL_TERM(0, row) + COL_TERM(1, row) + COL_TERM(2, row) + COL_TERM(3, row)))
▲ Show 20 Lines • Show All 106 Lines • Show Last 20 Lines