Code refactor: make texture code more consistent between devices.

* Use common TextureInfo struct for all devices, except CUDA fermi.
* Move image sampling code to kernels/*/kernel_*_image.h files.
* Use arrays for data textures on Fermi too, so device_vector<Struct> works.

2 files changed, 463 insertions, 147 deletions

diff --git a/intern/cycles/kernel/kernels/cpu/kernel.cpp b/intern/cycles/kernel/kernels/cpu/kernel.cpp
index 998619ac897..7679ab4f111 100644
--- a/intern/cycles/kernel/kernels/cpu/kernel.cpp
+++ b/intern/cycles/kernel/kernels/cpu/kernel.cpp
@@ -84,130 +84,16 @@ void kernel_tex_copy(KernelGlobals *kg,
 	if(0) {
 	}
 
-#define KERNEL_TEX(type, ttype, tname) \
+#define KERNEL_TEX(type, tname) \
 	else if(strcmp(name, #tname) == 0) { \
 		kg->tname.data = (type*)mem; \
 		kg->tname.width = width; \
 	}
-#define KERNEL_IMAGE_TEX(type, ttype, tname)
+#define KERNEL_IMAGE_TEX(type, tname)
 #include "kernel/kernel_textures.h"
-
-	else if(strstr(name, "__tex_image_float4")) {
-		texture_image_float4 *tex = NULL;
-		int id = atoi(name + strlen("__tex_image_float4_"));
-		int array_index = kernel_tex_index(id);
-
-		if(array_index >= 0) {
-			if(array_index >= kg->texture_float4_images.size()) {
-				kg->texture_float4_images.resize(array_index+1);
-			}
-			tex = &kg->texture_float4_images[array_index];
-		}
-
-		if(tex) {
-			tex->data = (float4*)mem;
-			tex->dimensions_set(width, height, depth);
-			tex->interpolation = interpolation;
-			tex->extension = extension;
-		}
-	}
-	else if(strstr(name, "__tex_image_float")) {
-		texture_image_float *tex = NULL;
-		int id = atoi(name + strlen("__tex_image_float_"));
-		int array_index = kernel_tex_index(id);
-
-		if(array_index >= 0) {
-			if(array_index >= kg->texture_float_images.size()) {
-				kg->texture_float_images.resize(array_index+1);
-			}
-			tex = &kg->texture_float_images[array_index];
-		}
-
-		if(tex) {
-			tex->data = (float*)mem;
-			tex->dimensions_set(width, height, depth);
-			tex->interpolation = interpolation;
-			tex->extension = extension;
-		}
-	}
-	else if(strstr(name, "__tex_image_byte4")) {
-		texture_image_uchar4 *tex = NULL;
-		int id = atoi(name + strlen("__tex_image_byte4_"));
-		int array_index = kernel_tex_index(id);
-
-		if(array_index >= 0) {
-			if(array_index >= kg->texture_byte4_images.size()) {
-				kg->texture_byte4_images.resize(array_index+1);
-			}
-			tex = &kg->texture_byte4_images[array_index];
-		}
-
-		if(tex) {
-			tex->data = (uchar4*)mem;
-			tex->dimensions_set(width, height, depth);
-			tex->interpolation = interpolation;
-			tex->extension = extension;
-		}
-	}
-	else if(strstr(name, "__tex_image_byte")) {
-		texture_image_uchar *tex = NULL;
-		int id = atoi(name + strlen("__tex_image_byte_"));
-		int array_index = kernel_tex_index(id);
-
-		if(array_index >= 0) {
-			if(array_index >= kg->texture_byte_images.size()) {
-				kg->texture_byte_images.resize(array_index+1);
-			}
-			tex = &kg->texture_byte_images[array_index];
-		}
-
-		if(tex) {
-			tex->data = (uchar*)mem;
-			tex->dimensions_set(width, height, depth);
-			tex->interpolation = interpolation;
-			tex->extension = extension;
-		}
-	}
-	else if(strstr(name, "__tex_image_half4")) {
-		texture_image_half4 *tex = NULL;
-		int id = atoi(name + strlen("__tex_image_half4_"));
-		int array_index = kernel_tex_index(id);
-
-		if(array_index >= 0) {
-			if(array_index >= kg->texture_half4_images.size()) {
-				kg->texture_half4_images.resize(array_index+1);
-			}
-			tex = &kg->texture_half4_images[array_index];
-		}
-
-		if(tex) {
-			tex->data = (half4*)mem;
-			tex->dimensions_set(width, height, depth);
-			tex->interpolation = interpolation;
-			tex->extension = extension;
-		}
-	}
-	else if(strstr(name, "__tex_image_half")) {
-		texture_image_half *tex = NULL;
-		int id = atoi(name + strlen("__tex_image_half_"));
-		int array_index = kernel_tex_index(id);
-
-		if(array_index >= 0) {
-			if(array_index >= kg->texture_half_images.size()) {
-				kg->texture_half_images.resize(array_index+1);
-			}
-			tex = &kg->texture_half_images[array_index];
-		}
-
-		if(tex) {
-			tex->data = (half*)mem;
-			tex->dimensions_set(width, height, depth);
-			tex->interpolation = interpolation;
-			tex->extension = extension;
-		}
-	}
-	else
+	else {
 		assert(0);
+	}
 }
 
 CCL_NAMESPACE_END
diff --git a/intern/cycles/kernel/kernels/cpu/kernel_cpu_image.h b/intern/cycles/kernel/kernels/cpu/kernel_cpu_image.h
index f6bb4c25012..b2ad60f08c1 100644
--- a/intern/cycles/kernel/kernels/cpu/kernel_cpu_image.h
+++ b/intern/cycles/kernel/kernels/cpu/kernel_cpu_image.h
@@ -17,70 +17,500 @@
 #ifndef __KERNEL_CPU_IMAGE_H__
 #define __KERNEL_CPU_IMAGE_H__
 
-#ifdef __KERNEL_CPU__
-
 CCL_NAMESPACE_BEGIN
 
-ccl_device float4 kernel_tex_image_interp_impl(KernelGlobals *kg, int tex, float x, float y)
+template<typename T> struct TextureInterpolator  {
+#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[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[3] = (1.0f / 6.0f) * t * t * t; \
+	} (void)0
+
+	static ccl_always_inline float4 read(float4 r)
+	{
+		return r;
+	}
+
+	static ccl_always_inline float4 read(uchar4 r)
+	{
+		float f = 1.0f/255.0f;
+		return make_float4(r.x*f, r.y*f, r.z*f, r.w*f);
+	}
+
+	static ccl_always_inline float4 read(uchar r)
+	{
+		float f = r*(1.0f/255.0f);
+		return make_float4(f, f, f, 1.0f);
+	}
+
+	static ccl_always_inline float4 read(float r)
+	{
+		/* TODO(dingto): Optimize this, so interpolation
+		 * happens on float instead of float4 */
+		return make_float4(r, r, r, 1.0f);
+	}
+
+	static ccl_always_inline float4 read(half4 r)
+	{
+		return half4_to_float4(r);
+	}
+
+	static ccl_always_inline float4 read(half r)
+	{
+		float f = half_to_float(r);
+		return make_float4(f, f, f, 1.0f);
+	}
+
+	static ccl_always_inline int wrap_periodic(int x, int width)
+	{
+		x %= width;
+		if(x < 0)
+			x += width;
+		return x;
+	}
+
+	static ccl_always_inline int wrap_clamp(int x, int width)
+	{
+		return clamp(x, 0, width-1);
+	}
+
+	static ccl_always_inline float frac(float x, int *ix)
+	{
+		int i = float_to_int(x) - ((x < 0.0f)? 1: 0);
+		*ix = i;
+		return x - (float)i;
+	}
+
+	static ccl_always_inline float4 interp(const TextureInfo& info, float x, float y)
+	{
+		if(UNLIKELY(!info.data))
+			return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+
+		const T *data = (const T*)info.data;
+		int width = info.width;
+		int height = info.height;
+		int ix, iy, nix, niy;
+
+		if(info.interpolation == INTERPOLATION_CLOSEST) {
+			frac(x*(float)width, &ix);
+			frac(y*(float)height, &iy);
+			switch(info.extension) {
+				case EXTENSION_REPEAT:
+					ix = wrap_periodic(ix, width);
+					iy = wrap_periodic(iy, height);
+					break;
+				case EXTENSION_CLIP:
+					if(x < 0.0f || y < 0.0f || x > 1.0f || y > 1.0f) {
+						return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+					}
+					ATTR_FALLTHROUGH;
+				case EXTENSION_EXTEND:
+					ix = wrap_clamp(ix, width);
+					iy = wrap_clamp(iy, height);
+					break;
+				default:
+					kernel_assert(0);
+					return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+			}
+			return read(data[ix + iy*width]);
+		}
+		else if(info.interpolation == INTERPOLATION_LINEAR) {
+			float tx = frac(x*(float)width - 0.5f, &ix);
+			float ty = frac(y*(float)height - 0.5f, &iy);
+
+			switch(info.extension) {
+				case EXTENSION_REPEAT:
+					ix = wrap_periodic(ix, width);
+					iy = wrap_periodic(iy, height);
+
+					nix = wrap_periodic(ix+1, width);
+					niy = wrap_periodic(iy+1, height);
+					break;
+				case EXTENSION_CLIP:
+					if(x < 0.0f || y < 0.0f || x > 1.0f || y > 1.0f) {
+						return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+					}
+					ATTR_FALLTHROUGH;
+				case EXTENSION_EXTEND:
+					nix = wrap_clamp(ix+1, width);
+					niy = wrap_clamp(iy+1, height);
+
+					ix = wrap_clamp(ix, width);
+					iy = wrap_clamp(iy, height);
+					break;
+				default:
+					kernel_assert(0);
+					return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+			}
+
+			float4 r = (1.0f - ty)*(1.0f - tx)*read(data[ix + iy*width]);
+			r += (1.0f - ty)*tx*read(data[nix + iy*width]);
+			r += ty*(1.0f - tx)*read(data[ix + niy*width]);
+			r += ty*tx*read(data[nix + niy*width]);
+
+			return r;
+		}
+		else {
+			/* Bicubic b-spline interpolation. */
+			float tx = frac(x*(float)width - 0.5f, &ix);
+			float ty = frac(y*(float)height - 0.5f, &iy);
+			int pix, piy, nnix, nniy;
+			switch(info.extension) {
+				case EXTENSION_REPEAT:
+					ix = wrap_periodic(ix, width);
+					iy = wrap_periodic(iy, height);
+
+					pix = wrap_periodic(ix-1, width);
+					piy = wrap_periodic(iy-1, height);
+
+					nix = wrap_periodic(ix+1, width);
+					niy = wrap_periodic(iy+1, height);
+
+					nnix = wrap_periodic(ix+2, width);
+					nniy = wrap_periodic(iy+2, height);
+					break;
+				case EXTENSION_CLIP:
+					if(x < 0.0f || y < 0.0f || x > 1.0f || y > 1.0f) {
+						return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+					}
+					ATTR_FALLTHROUGH;
+				case EXTENSION_EXTEND:
+					pix = wrap_clamp(ix-1, width);
+					piy = wrap_clamp(iy-1, height);
+
+					nix = wrap_clamp(ix+1, width);
+					niy = wrap_clamp(iy+1, height);
+
+					nnix = wrap_clamp(ix+2, width);
+					nniy = wrap_clamp(iy+2, height);
+
+					ix = wrap_clamp(ix, width);
+					iy = wrap_clamp(iy, height);
+					break;
+				default:
+					kernel_assert(0);
+					return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+			}
+
+			const int xc[4] = {pix, ix, nix, nnix};
+			const int yc[4] = {width * piy,
+			                   width * iy,
+			                   width * niy,
+			                   width * nniy};
+			float u[4], v[4];
+			/* Some helper macro to keep code reasonable size,
+			 * let compiler to inline all the matrix multiplications.
+			 */
+#define DATA(x, y) (read(data[xc[x] + yc[y]]))
+#define TERM(col) \
+			(v[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(v, ty);
+
+			/* Actual interpolation. */
+			return TERM(0) + TERM(1) + TERM(2) + TERM(3);
+
+#undef TERM
+#undef DATA
+		}
+	}
+
+	static ccl_always_inline float4 interp_3d_closest(const TextureInfo& info, float x, float y, float z)
+	{
+		int width = info.width;
+		int height = info.height;
+		int depth = info.depth;
+		int ix, iy, iz;
+
+		frac(x*(float)width, &ix);
+		frac(y*(float)height, &iy);
+		frac(z*(float)depth, &iz);
+
+		switch(info.extension) {
+			case EXTENSION_REPEAT:
+				ix = wrap_periodic(ix, width);
+				iy = wrap_periodic(iy, height);
+				iz = wrap_periodic(iz, depth);
+				break;
+			case EXTENSION_CLIP:
+				if(x < 0.0f || y < 0.0f || z < 0.0f ||
+				   x > 1.0f || y > 1.0f || z > 1.0f)
+				{
+					return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+				}
+				ATTR_FALLTHROUGH;
+			case EXTENSION_EXTEND:
+				ix = wrap_clamp(ix, width);
+				iy = wrap_clamp(iy, height);
+				iz = wrap_clamp(iz, depth);
+				break;
+			default:
+				kernel_assert(0);
+				return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+		}
+
+		const T *data = (const T*)info.data;
+		return read(data[ix + iy*width + iz*width*height]);
+	}
+
+	static ccl_always_inline float4 interp_3d_linear(const TextureInfo& info, float x, float y, float z)
+	{
+		int width = info.width;
+		int height = info.height;
+		int depth = info.depth;
+		int ix, iy, iz;
+		int nix, niy, niz;
+
+		float tx = frac(x*(float)width - 0.5f, &ix);
+		float ty = frac(y*(float)height - 0.5f, &iy);
+		float tz = frac(z*(float)depth - 0.5f, &iz);
+
+		switch(info.extension) {
+			case EXTENSION_REPEAT:
+				ix = wrap_periodic(ix, width);
+				iy = wrap_periodic(iy, height);
+				iz = wrap_periodic(iz, depth);
+
+				nix = wrap_periodic(ix+1, width);
+				niy = wrap_periodic(iy+1, height);
+				niz = wrap_periodic(iz+1, depth);
+				break;
+			case EXTENSION_CLIP:
+				if(x < 0.0f || y < 0.0f || z < 0.0f ||
+				   x > 1.0f || y > 1.0f || z > 1.0f)
+				{
+					return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+				}
+				ATTR_FALLTHROUGH;
+			case EXTENSION_EXTEND:
+				nix = wrap_clamp(ix+1, width);
+				niy = wrap_clamp(iy+1, height);
+				niz = wrap_clamp(iz+1, depth);
+
+				ix = wrap_clamp(ix, width);
+				iy = wrap_clamp(iy, height);
+				iz = wrap_clamp(iz, depth);
+				break;
+			default:
+				kernel_assert(0);
+				return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+		}
+
+		const T *data = (const T*)info.data;
+		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)*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*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)*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*tx*read(data[nix + niy*width + niz*width*height]);
+
+		return r;
+	}
+
+	/* 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.
+	 *
+	 * Only happens for AVX2 kernel and global __KERNEL_SSE__ vectorization
+	 * enabled.
+	 */
+#ifdef __GNUC__
+	static ccl_always_inline
+#else
+	static ccl_never_inline
+#endif
+	float4 interp_3d_tricubic(const TextureInfo& info, float x, float y, float z)
+	{
+		int width = info.width;
+		int height = info.height;
+		int depth = info.depth;
+		int ix, iy, iz;
+		int nix, niy, niz;
+		/* Tricubic b-spline interpolation. */
+		const float tx = frac(x*(float)width - 0.5f, &ix);
+		const float ty = frac(y*(float)height - 0.5f, &iy);
+		const float tz = frac(z*(float)depth - 0.5f, &iz);
+		int pix, piy, piz, nnix, nniy, nniz;
+
+		switch(info.extension) {
+			case EXTENSION_REPEAT:
+				ix = wrap_periodic(ix, width);
+				iy = wrap_periodic(iy, height);
+				iz = wrap_periodic(iz, depth);
+
+				pix = wrap_periodic(ix-1, width);
+				piy = wrap_periodic(iy-1, height);
+				piz = wrap_periodic(iz-1, depth);
+
+				nix = wrap_periodic(ix+1, width);
+				niy = wrap_periodic(iy+1, height);
+				niz = wrap_periodic(iz+1, depth);
+
+				nnix = wrap_periodic(ix+2, width);
+				nniy = wrap_periodic(iy+2, height);
+				nniz = wrap_periodic(iz+2, depth);
+				break;
+			case EXTENSION_CLIP:
+				if(x < 0.0f || y < 0.0f || z < 0.0f ||
+				   x > 1.0f || y > 1.0f || z > 1.0f)
+				{
+					return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+				}
+				ATTR_FALLTHROUGH;
+			case EXTENSION_EXTEND:
+				pix = wrap_clamp(ix-1, width);
+				piy = wrap_clamp(iy-1, height);
+				piz = wrap_clamp(iz-1, depth);
+
+				nix = wrap_clamp(ix+1, width);
+				niy = wrap_clamp(iy+1, height);
+				niz = wrap_clamp(iz+1, depth);
+
+				nnix = wrap_clamp(ix+2, width);
+				nniy = wrap_clamp(iy+2, height);
+				nniz = wrap_clamp(iz+2, depth);
+
+				ix = wrap_clamp(ix, width);
+				iy = wrap_clamp(iy, height);
+				iz = wrap_clamp(iz, depth);
+				break;
+			default:
+				kernel_assert(0);
+				return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+		}
+
+		const int xc[4] = {pix, ix, nix, nnix};
+		const int yc[4] = {width * piy,
+		                   width * iy,
+		                   width * niy,
+		                   width * nniy};
+		const int zc[4] = {width * height * piz,
+		                   width * height * iz,
+		                   width * height * niz,
+		                   width * height * nniz};
+		float u[4], v[4], w[4];
+
+		/* Some helper macro to keep code reasonable size,
+		 * let compiler to inline all the matrix multiplications.
+		 */
+#define DATA(x, y, z) (read(data[xc[x] + yc[y] + zc[z]]))
+#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. */
+		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 interp_3d(const TextureInfo& info,
+	                                          float x, float y, float z,
+	                                          int interpolation = INTERPOLATION_LINEAR)
+	{
+		if(UNLIKELY(!info.data))
+			return make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+
+		switch(interpolation) {
+			case INTERPOLATION_CLOSEST:
+				return interp_3d_closest(info, x, y, z);
+			case INTERPOLATION_LINEAR:
+				return interp_3d_linear(info, x, y, z);
+			default:
+				return interp_3d_tricubic(info, x, y, z);
+		}
+	}
+#undef SET_CUBIC_SPLINE_WEIGHTS
+};
+
+ccl_device float4 kernel_tex_image_interp(KernelGlobals *kg, int id, float x, float y)
 {
-	switch(kernel_tex_type(tex)) {
+	const TextureInfo& info = kernel_tex_fetch(__texture_info, id);
+
+	switch(kernel_tex_type(id)) {
 		case IMAGE_DATA_TYPE_HALF:
-			return kg->texture_half_images[kernel_tex_index(tex)].interp(x, y);
+			return TextureInterpolator<half>::interp(info, x, y);
 		case IMAGE_DATA_TYPE_BYTE:
-			return kg->texture_byte_images[kernel_tex_index(tex)].interp(x, y);
+			return TextureInterpolator<uchar>::interp(info, x, y);
 		case IMAGE_DATA_TYPE_FLOAT:
-			return kg->texture_float_images[kernel_tex_index(tex)].interp(x, y);
+			return TextureInterpolator<float>::interp(info, x, y);
 		case IMAGE_DATA_TYPE_HALF4:
-			return kg->texture_half4_images[kernel_tex_index(tex)].interp(x, y);
+			return TextureInterpolator<half4>::interp(info, x, y);
 		case IMAGE_DATA_TYPE_BYTE4:
-			return kg->texture_byte4_images[kernel_tex_index(tex)].interp(x, y);
+			return TextureInterpolator<uchar4>::interp(info, x, y);
 		case IMAGE_DATA_TYPE_FLOAT4:
 		default:
-			return kg->texture_float4_images[kernel_tex_index(tex)].interp(x, y);
+			return TextureInterpolator<float4>::interp(info, x, y);
 	}
 }
 
-ccl_device float4 kernel_tex_image_interp_3d_impl(KernelGlobals *kg, int tex, float x, float y, float z)
+ccl_device float4 kernel_tex_image_interp_3d(KernelGlobals *kg, int id, float x, float y, float z)
 {
-	switch(kernel_tex_type(tex)) {
+	const TextureInfo& info = kernel_tex_fetch(__texture_info, id);
+	InterpolationType interp = (InterpolationType)info.interpolation;
+
+	switch(kernel_tex_type(id)) {
 		case IMAGE_DATA_TYPE_HALF:
-			return kg->texture_half_images[kernel_tex_index(tex)].interp_3d(x, y, z);
+			return TextureInterpolator<half>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_BYTE:
-			return kg->texture_byte_images[kernel_tex_index(tex)].interp_3d(x, y, z);
+			return TextureInterpolator<uchar>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_FLOAT:
-			return kg->texture_float_images[kernel_tex_index(tex)].interp_3d(x, y, z);
+			return TextureInterpolator<float>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_HALF4:
-			return kg->texture_half4_images[kernel_tex_index(tex)].interp_3d(x, y, z);
+			return TextureInterpolator<half4>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_BYTE4:
-			return kg->texture_byte4_images[kernel_tex_index(tex)].interp_3d(x, y, z);
+			return TextureInterpolator<uchar4>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_FLOAT4:
 		default:
-			return kg->texture_float4_images[kernel_tex_index(tex)].interp_3d(x, y, z);
+			return TextureInterpolator<float4>::interp_3d(info, x, y, z, interp);
 	}
 }
 
-ccl_device float4 kernel_tex_image_interp_3d_ex_impl(KernelGlobals *kg, int tex, float x, float y, float z, int interpolation)
+ccl_device float4 kernel_tex_image_interp_3d_ex(KernelGlobals *kg, int id, float x, float y, float z, int interp)
 {
-	switch(kernel_tex_type(tex)) {
+	const TextureInfo& info = kernel_tex_fetch(__texture_info, id);
+
+	switch(kernel_tex_type(id)) {
 		case IMAGE_DATA_TYPE_HALF:
-			return kg->texture_half_images[kernel_tex_index(tex)].interp_3d_ex(x, y, z, interpolation);
+			return TextureInterpolator<half>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_BYTE:
-			return kg->texture_byte_images[kernel_tex_index(tex)].interp_3d_ex(x, y, z, interpolation);
+			return TextureInterpolator<uchar>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_FLOAT:
-			return kg->texture_float_images[kernel_tex_index(tex)].interp_3d_ex(x, y, z, interpolation);
+			return TextureInterpolator<float>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_HALF4:
-			return kg->texture_half4_images[kernel_tex_index(tex)].interp_3d_ex(x, y, z, interpolation);
+			return TextureInterpolator<half4>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_BYTE4:
-			return kg->texture_byte4_images[kernel_tex_index(tex)].interp_3d_ex(x, y, z, interpolation);
+			return TextureInterpolator<uchar4>::interp_3d(info, x, y, z, interp);
 		case IMAGE_DATA_TYPE_FLOAT4:
 		default:
-			return kg->texture_float4_images[kernel_tex_index(tex)].interp_3d_ex(x, y, z, interpolation);
+			return TextureInterpolator<float4>::interp_3d(info, x, y, z, interp);
 	}
 }
 
 CCL_NAMESPACE_END
 
-#endif  // __KERNEL_CPU__
-
-
 #endif // __KERNEL_CPU_IMAGE_H__