Page MenuHome
Differential D2595 Diff 8567 intern/cycles/kernel/kernel_compat_cpu.h

Changeset View

Standalone View

View Options

intern/cycles/kernel/kernel_compat_cpu.h

Show First 20 Lines • Show All 310 Lines • ▼ Show 20 Lines#undef DATA
} }
} }
ccl_always_inline float4 interp_3d(float x, float y, float z) ccl_always_inline float4 interp_3d(float x, float y, float z)
{ {
return interp_3d_ex(x, y, z, interpolation); return interp_3d_ex(x, y, z, interpolation);
} }
ccl_always_inline float4 interp_3d_ex(float x, float y, float z, ccl_always_inline float4 interp_3d_ex_closest(float x, float y, float z)
int interpolation = INTERPOLATION_LINEAR)
{ {
if(UNLIKELY(!data)) int ix, iy, iz;
return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
int ix, iy, iz, nix, niy, niz;
if(interpolation == INTERPOLATION_CLOSEST) {
frac(x*(float)width, &ix); frac(x*(float)width, &ix);
frac(y*(float)height, &iy); frac(y*(float)height, &iy);
frac(z*(float)depth, &iz); frac(z*(float)depth, &iz);
switch(extension) { switch(extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, width);
sergeyUnsubmitted
Not Done
Inline Actions

Indentation.

sergey: Indentation.
iy = wrap_periodic(iy, height); iy = wrap_periodic(iy, height);
iz = wrap_periodic(iz, depth); iz = wrap_periodic(iz, depth);
break; break;
case EXTENSION_CLIP: case EXTENSION_CLIP:
if(x < 0.0f || y < 0.0f || z < 0.0f || if(x < 0.0f || y < 0.0f || z < 0.0f ||
x > 1.0f || y > 1.0f || z > 1.0f) x > 1.0f || y > 1.0f || 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);
} }
/* Fall through. */ /* Fall through. */
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
ix = wrap_clamp(ix, width); ix = wrap_clamp(ix, width);
iy = wrap_clamp(iy, height); iy = wrap_clamp(iy, height);
iz = wrap_clamp(iz, 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 make_float4(0.0f, 0.0f, 0.0f, 0.0f);
} }
return read(data[ix + iy*width + iz*width*height]); return read(data[ix + iy*width + iz*width*height]);
} }
else if(interpolation == INTERPOLATION_LINEAR) {
ccl_always_inline float4 interp_3d_ex_linear(float x, float y, float z)
{
int ix, iy, iz;
int nix, niy, niz;
float tx = frac(x*(float)width - 0.5f, &ix); float tx = frac(x*(float)width - 0.5f, &ix);
float ty = frac(y*(float)height - 0.5f, &iy); float ty = frac(y*(float)height - 0.5f, &iy);
float tz = frac(z*(float)depth - 0.5f, &iz); float tz = frac(z*(float)depth - 0.5f, &iz);
switch(extension) { switch(extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, width);
sergeyUnsubmitted
Done
Inline Actions

We don't use space after keyword in Cycles.

Same applies to the cases below.

sergey: We don't use space after keyword in Cycles. Same applies to the cases below.
iy = wrap_periodic(iy, height); iy = wrap_periodic(iy, height);
sergeyUnsubmitted
Done
Inline Actions

Indentation.

sergey: Indentation.
iz = wrap_periodic(iz, depth); iz = wrap_periodic(iz, depth);
nix = wrap_periodic(ix+1, width); nix = wrap_periodic(ix+1, width);
niy = wrap_periodic(iy+1, height); niy = wrap_periodic(iy+1, height);
niz = wrap_periodic(iz+1, depth); niz = wrap_periodic(iz+1, depth);
break; break;
case EXTENSION_CLIP: case EXTENSION_CLIP:
if(x < 0.0f || y < 0.0f || z < 0.0f || if(x < 0.0f || y < 0.0f || z < 0.0f ||
x > 1.0f || y > 1.0f || z > 1.0f) x > 1.0f || y > 1.0f || 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);
} }
/* Fall through. */ /* Fall through. */
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
nix = wrap_clamp(ix+1, width); nix = wrap_clamp(ix+1, width);
niy = wrap_clamp(iy+1, height); niy = wrap_clamp(iy+1, height);
niz = wrap_clamp(iz+1, depth); niz = wrap_clamp(iz+1, depth);
ix = wrap_clamp(ix, width); ix = wrap_clamp(ix, width);
iy = wrap_clamp(iy, height); iy = wrap_clamp(iy, height);
iz = wrap_clamp(iz, 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 make_float4(0.0f, 0.0f, 0.0f, 0.0f);
} }
float4 r; float4 r;
r = (1.0f - tz)*(1.0f - ty)*(1.0f - tx)*read(data[ix + iy*width + iz*width*height]); r = (1.0f - tz)*(1.0f - ty)*(1.0f - tx)*read(data[ix + iy*width + iz*width*height]);
r += (1.0f - tz)*(1.0f - ty)*tx*read(data[nix + iy*width + iz*width*height]); r += (1.0f - tz)*(1.0f - ty)*tx*read(data[nix + iy*width + iz*width*height]);
r += (1.0f - tz)*ty*(1.0f - tx)*read(data[ix + niy*width + iz*width*height]); r += (1.0f - tz)*ty*(1.0f - tx)*read(data[ix + niy*width + iz*width*height]);
r += (1.0f - tz)*ty*tx*read(data[nix + niy*width + iz*width*height]); r += (1.0f - tz)*ty*tx*read(data[nix + niy*width + iz*width*height]);
r += tz*(1.0f - ty)*(1.0f - tx)*read(data[ix + iy*width + niz*width*height]); r += tz*(1.0f - ty)*(1.0f - tx)*read(data[ix + iy*width + niz*width*height]);
r += tz*(1.0f - ty)*tx*read(data[nix + iy*width + niz*width*height]); r += tz*(1.0f - ty)*tx*read(data[nix + iy*width + niz*width*height]);
r += tz*ty*(1.0f - tx)*read(data[ix + niy*width + niz*width*height]); r += tz*ty*(1.0f - tx)*read(data[ix + niy*width + niz*width*height]);
r += tz*ty*tx*read(data[nix + niy*width + niz*width*height]); r += tz*ty*tx*read(data[nix + niy*width + niz*width*height]);
return r; return r;
} }
else {
ccl_never_inline float4 interp_3d_ex_tricubic(float x, float y, float z)
{
int ix, iy, iz;
int nix, niy, niz;
LazyDodoAuthorUnsubmitted
Not Done
Inline Actions

given the problem is very msvc specific, what are the thoughts of an ccl_msvc_never_inline ? i mean no need to punish gcc for msvc's weirdness?

LazyDodo: given the problem is very msvc specific, what are the thoughts of an ccl_msvc_never_inline ? i…
sergeyUnsubmitted
Not Done
Inline Actions

Don't do that. Otherwise you'll end up with nasty things like ccl_inline_for_cuda_unline_for_gcc_forceinline_for_msvc.

This is a big enough function which doesn't make much sense to inline in any of compilers.

sergey: Don't do that. Otherwise you'll end up with nasty things like…
/* Tricubic b-spline interpolation. */ /* Tricubic b-spline interpolation. */
const float tx = frac(x*(float)width - 0.5f, &ix); const float tx = frac(x*(float)width - 0.5f, &ix);
const float ty = frac(y*(float)height - 0.5f, &iy); const float ty = frac(y*(float)height - 0.5f, &iy);
const float tz = frac(z*(float)depth - 0.5f, &iz); const float tz = frac(z*(float)depth - 0.5f, &iz);
int pix, piy, piz, nnix, nniy, nniz; int pix, piy, piz, nnix, nniy, nniz;
switch(extension) { switch(extension) {
case EXTENSION_REPEAT: case EXTENSION_REPEAT:
ix = wrap_periodic(ix, width); ix = wrap_periodic(ix, width);
iy = wrap_periodic(iy, height); iy = wrap_periodic(iy, height);
iz = wrap_periodic(iz, depth); iz = wrap_periodic(iz, depth);
sergeyUnsubmitted
Not Done
Inline Actions

Indentation.

sergey: Indentation.
pix = wrap_periodic(ix-1, width); pix = wrap_periodic(ix-1, width);
piy = wrap_periodic(iy-1, height); piy = wrap_periodic(iy-1, height);
piz = wrap_periodic(iz-1, depth); piz = wrap_periodic(iz-1, depth);
nix = wrap_periodic(ix+1, width); nix = wrap_periodic(ix+1, width);
niy = wrap_periodic(iy+1, height); niy = wrap_periodic(iy+1, height);
niz = wrap_periodic(iz+1, depth); niz = wrap_periodic(iz+1, depth);
nnix = wrap_periodic(ix+2, width); nnix = wrap_periodic(ix+2, width);
nniy = wrap_periodic(iy+2, height); nniy = wrap_periodic(iy+2, height);
nniz = wrap_periodic(iz+2, depth); nniz = wrap_periodic(iz+2, depth);
break; break;
case EXTENSION_CLIP: case EXTENSION_CLIP:
if(x < 0.0f || y < 0.0f || z < 0.0f || if(x < 0.0f || y < 0.0f || z < 0.0f ||
x > 1.0f || y > 1.0f || z > 1.0f) x > 1.0f || y > 1.0f || 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);
} }
sergeyUnsubmitted
Not Done
Inline Actions

Indentation (x was aligned to the same column in original code).

sergey: Indentation (`x` was aligned to the same column in original code).
/* Fall through. */ /* Fall through. */
case EXTENSION_EXTEND: case EXTENSION_EXTEND:
pix = wrap_clamp(ix-1, width); pix = wrap_clamp(ix-1, width);
piy = wrap_clamp(iy-1, height); piy = wrap_clamp(iy-1, height);
piz = wrap_clamp(iz-1, depth); piz = wrap_clamp(iz-1, depth);
nix = wrap_clamp(ix+1, width); nix = wrap_clamp(ix+1, width);
niy = wrap_clamp(iy+1, height); niy = wrap_clamp(iy+1, height);
niz = wrap_clamp(iz+1, depth); niz = wrap_clamp(iz+1, depth);
nnix = wrap_clamp(ix+2, width); nnix = wrap_clamp(ix+2, width);
nniy = wrap_clamp(iy+2, height); nniy = wrap_clamp(iy+2, height);
nniz = wrap_clamp(iz+2, depth); nniz = wrap_clamp(iz+2, depth);
ix = wrap_clamp(ix, width); ix = wrap_clamp(ix, width);
iy = wrap_clamp(iy, height); iy = wrap_clamp(iy, height);
iz = wrap_clamp(iz, 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 make_float4(0.0f, 0.0f, 0.0f, 0.0f);
} }
const int xc[4] = {pix, ix, nix, nnix}; const int xc[4] = {pix, ix, nix, nnix};
const int yc[4] = {width * piy, const int yc[4] = {width * piy,
width * iy, width * iy,
width * niy, width * niy,
width * nniy}; width * nniy};
const int zc[4] = {width * height * piz, const int zc[4] = {width * height * piz,
sergeyUnsubmitted
Not Done
Inline Actions

Indentation and space around {}. It used to be

const int yc[4] = {width * piy,
                   width * iy,
                   width * niy,
                   width * nniy};
sergey: Indentation and space around {}. It used to be const int yc[4] = {width * piy…
width * height * iz, width * height * iz,
width * height * niz, width * height * niz,
width * height * nniz}; width * height * 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 macro to keep code reasonable size,
* let compiler to inline all the matrix multiplications. * let compiler to inline all the matrix multiplications.
*/ */
#define DATA(x, y, z) (read(data[xc[x] + yc[y] + zc[z]])) #define DATA(x, y, z) (read(data[xc[x] + yc[y] + zc[z]]))
#define COL_TERM(col, row) \ #define COL_TERM(col, row) \
sergeyUnsubmitted
Not Done
Inline Actions

Indentation:

/* Note thew asterisk are at the same column.
 * (they became at the beginning of the line).
 */
sergey: Indentation: /* Note thew asterisk are at the same column. * (they became at the…
(v[col] * (u[0] * DATA(0, col, row) + \ (v[col] * (u[0] * DATA(0, col, row) + \
u[1] * DATA(1, col, row) + \ u[1] * DATA(1, col, row) + \
u[2] * DATA(2, 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) + \ (w[row] * (COL_TERM(0, row) + \
COL_TERM(1, row) + \ COL_TERM(1, row) + \
COL_TERM(2, row) + \ COL_TERM(2, row) + \
COL_TERM(3, row))) COL_TERM(3, row)))
SET_CUBIC_SPLINE_WEIGHTS(u, tx); SET_CUBIC_SPLINE_WEIGHTS(u, tx);
SET_CUBIC_SPLINE_WEIGHTS(v, ty); SET_CUBIC_SPLINE_WEIGHTS(v, ty);
SET_CUBIC_SPLINE_WEIGHTS(w, tz); SET_CUBIC_SPLINE_WEIGHTS(w, tz);
/* Actual interpolation. */ /* Actual interpolation. */
return ROW_TERM(0) + ROW_TERM(1) + ROW_TERM(2) + ROW_TERM(3); return ROW_TERM(0) + ROW_TERM(1) + ROW_TERM(2) + ROW_TERM(3);
#undef COL_TERM #undef COL_TERM
#undef ROW_TERM #undef ROW_TERM
sergeyUnsubmitted
Done
Inline Actions

Seems to be remained from some experiments.

sergey: Seems to be remained from some experiments.
#undef DATA #undef DATA
} }
ccl_always_inline float4 interp_3d_ex(float x, float y, float z,
int interpolation = INTERPOLATION_LINEAR)
{
if(UNLIKELY(!data))
return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
switch(interpolation) {
case INTERPOLATION_CLOSEST:
return interp_3d_ex_closest(x, y, z);
case INTERPOLATION_LINEAR:
return interp_3d_ex_linear(x, y, z);
default:
return interp_3d_ex_tricubic(x, y, z);
}
} }
ccl_always_inline void dimensions_set(int width_, int height_, int depth_) ccl_always_inline void dimensions_set(int width_, int height_, int depth_)
{ {
width = width_; width = width_;
height = height_; height = height_;
depth = depth_; depth = depth_;
} }
▲ Show 20 Lines • Show All 69 Lines • Show Last 20 Lines