diff options
Diffstat (limited to 'src/core/CL/cl_kernels/warp_helpers.h')
| -rw-r--r-- | src/core/CL/cl_kernels/warp_helpers.h | 65 |
1 files changed, 32 insertions, 33 deletions
diff --git a/src/core/CL/cl_kernels/warp_helpers.h b/src/core/CL/cl_kernels/warp_helpers.h index 005861ddfa..6595bd1981 100644 --- a/src/core/CL/cl_kernels/warp_helpers.h +++ b/src/core/CL/cl_kernels/warp_helpers.h @@ -31,11 +31,13 @@ * @param[in] border_size Border size of the image * */ -inline const float8 clamp_to_border_with_size(float8 coords, const float width, const float height, const float border_size) +inline const float8 +clamp_to_border_with_size(float8 coords, const float width, const float height, const float border_size) { const float4 clamped_x = clamp(coords.even, 0.0f - border_size, width - 1 + border_size); const float4 clamped_y = clamp(coords.odd, 0.0f - border_size, height - 1 + border_size); - return (float8)(clamped_x.s0, clamped_y.s0, clamped_x.s1, clamped_y.s1, clamped_x.s2, clamped_y.s2, clamped_x.s3, clamped_y.s3); + return (float8)(clamped_x.s0, clamped_y.s0, clamped_x.s1, clamped_y.s1, clamped_x.s2, clamped_y.s2, clamped_x.s3, + clamped_y.s3); } /** Clamps the given coordinates to the borders. @@ -63,12 +65,6 @@ inline const VEC_DATA_TYPE(DATA_TYPE, 4) read_texels4(const Image *in, const int *((__global DATA_TYPE *)offset(in, coords.s6, coords.s7))); } -/** Returns the current thread coordinates. */ -inline const float2 get_current_coords() -{ - return (float2)(get_global_id(0) * 4, get_global_id(1)); -} - /** Given a texel coordinates this function will return the following array of coordinates: * [ P, right neighbour, below neighbour, below right neighbour ] * @@ -80,7 +76,8 @@ inline const float2 get_current_coords() */ inline const float8 get_neighbour_coords(const float2 coord) { - return (float8)(/*tl*/ coord.s0, coord.s1, /*tr*/ coord.s0 + 1, coord.s1, /*bl*/ coord.s0, coord.s1 + 1, /*br*/ coord.s0 + 1, coord.s1 + 1); + return (float8)(/*tl*/ coord.s0, coord.s1, /*tr*/ coord.s0 + 1, coord.s1, /*bl*/ coord.s0, coord.s1 + 1, + /*br*/ coord.s0 + 1, coord.s1 + 1); } /** Computes the bilinear interpolation for each set of coordinates in the vector coords and returns the values @@ -91,37 +88,38 @@ inline const float8 get_neighbour_coords(const float2 coord) * @param[in] height Height of the image * @param[in] border_size Border size */ -inline const VEC_DATA_TYPE(DATA_TYPE, 4) bilinear_interpolate_with_border(const Image *in, const float8 coords, const float width, const float height, const float border_size) +inline const VEC_DATA_TYPE(DATA_TYPE, 4) bilinear_interpolate_with_border( + const Image *in, const float8 coords, const float width, const float height, const float border_size) { // If any of the 4 texels is out of the image's boundaries we use the border value (REPLICATE or CONSTANT) for any texel out of the image. // Sets the 4x4 coordinates for each of the four input texels const float8 fc = floor(coords); - const float16 c1 = (float16)( - clamp_to_border_with_size(get_neighbour_coords((float2)(fc.s0, fc.s1)), width, height, border_size), - clamp_to_border_with_size(get_neighbour_coords((float2)(fc.s2, fc.s3)), width, height, border_size)); - const float16 c2 = (float16)( - clamp_to_border_with_size(get_neighbour_coords((float2)(fc.s4, fc.s5)), width, height, border_size), - clamp_to_border_with_size(get_neighbour_coords((float2)(fc.s6, fc.s7)), width, height, border_size)); + const float16 c1 = + (float16)(clamp_to_border_with_size(get_neighbour_coords((float2)(fc.s0, fc.s1)), width, height, border_size), + clamp_to_border_with_size(get_neighbour_coords((float2)(fc.s2, fc.s3)), width, height, border_size)); + const float16 c2 = + (float16)(clamp_to_border_with_size(get_neighbour_coords((float2)(fc.s4, fc.s5)), width, height, border_size), + clamp_to_border_with_size(get_neighbour_coords((float2)(fc.s6, fc.s7)), width, height, border_size)); // Loads the values from the input image const float16 t = (float16)( - /* tl, tr, bl, br */ - * ((__global DATA_TYPE *)offset(in, c1.s0, c1.s1)), *((__global DATA_TYPE *)offset(in, c1.s2, c1.s3)), - *((__global DATA_TYPE *)offset(in, c1.s4, c1.s5)), *((__global DATA_TYPE *)offset(in, c1.s6, c1.s7)), - *((__global DATA_TYPE *)offset(in, c1.s8, c1.s9)), *((__global DATA_TYPE *)offset(in, c1.sa, c1.sb)), - *((__global DATA_TYPE *)offset(in, c1.sc, c1.sd)), *((__global DATA_TYPE *)offset(in, c1.se, c1.sf)), - *((__global DATA_TYPE *)offset(in, c2.s0, c2.s1)), *((__global DATA_TYPE *)offset(in, c2.s2, c2.s3)), - *((__global DATA_TYPE *)offset(in, c2.s4, c2.s5)), *((__global DATA_TYPE *)offset(in, c2.s6, c2.s7)), - *((__global DATA_TYPE *)offset(in, c2.s8, c2.s9)), *((__global DATA_TYPE *)offset(in, c2.sa, c2.sb)), - *((__global DATA_TYPE *)offset(in, c2.sc, c2.sd)), *((__global DATA_TYPE *)offset(in, c2.se, c2.sf))); - const float8 a = coords - fc; - const float8 b = ((float8)(1.f)) - a; - const float4 fr = (float4)( - ((t.s0 * b.s0 * b.s1) + (t.s1 * a.s0 * b.s1) + (t.s2 * b.s0 * a.s1) + (t.s3 * a.s0 * a.s1)), - ((t.s4 * b.s2 * b.s3) + (t.s5 * a.s2 * b.s3) + (t.s6 * b.s2 * a.s3) + (t.s7 * a.s2 * a.s3)), - ((t.s8 * b.s4 * b.s5) + (t.s9 * a.s4 * b.s5) + (t.sa * b.s4 * a.s5) + (t.sb * a.s4 * a.s5)), - ((t.sc * b.s6 * b.s7) + (t.sd * a.s6 * b.s7) + (t.se * b.s6 * a.s7) + (t.sf * a.s6 * a.s7))); + /* tl, tr, bl, br */ + *((__global DATA_TYPE *)offset(in, c1.s0, c1.s1)), *((__global DATA_TYPE *)offset(in, c1.s2, c1.s3)), + *((__global DATA_TYPE *)offset(in, c1.s4, c1.s5)), *((__global DATA_TYPE *)offset(in, c1.s6, c1.s7)), + *((__global DATA_TYPE *)offset(in, c1.s8, c1.s9)), *((__global DATA_TYPE *)offset(in, c1.sa, c1.sb)), + *((__global DATA_TYPE *)offset(in, c1.sc, c1.sd)), *((__global DATA_TYPE *)offset(in, c1.se, c1.sf)), + *((__global DATA_TYPE *)offset(in, c2.s0, c2.s1)), *((__global DATA_TYPE *)offset(in, c2.s2, c2.s3)), + *((__global DATA_TYPE *)offset(in, c2.s4, c2.s5)), *((__global DATA_TYPE *)offset(in, c2.s6, c2.s7)), + *((__global DATA_TYPE *)offset(in, c2.s8, c2.s9)), *((__global DATA_TYPE *)offset(in, c2.sa, c2.sb)), + *((__global DATA_TYPE *)offset(in, c2.sc, c2.sd)), *((__global DATA_TYPE *)offset(in, c2.se, c2.sf))); + const float8 a = coords - fc; + const float8 b = ((float8)(1.f)) - a; + const float4 fr = + (float4)(((t.s0 * b.s0 * b.s1) + (t.s1 * a.s0 * b.s1) + (t.s2 * b.s0 * a.s1) + (t.s3 * a.s0 * a.s1)), + ((t.s4 * b.s2 * b.s3) + (t.s5 * a.s2 * b.s3) + (t.s6 * b.s2 * a.s3) + (t.s7 * a.s2 * a.s3)), + ((t.s8 * b.s4 * b.s5) + (t.s9 * a.s4 * b.s5) + (t.sa * b.s4 * a.s5) + (t.sb * a.s4 * a.s5)), + ((t.sc * b.s6 * b.s7) + (t.sd * a.s6 * b.s7) + (t.se * b.s6 * a.s7) + (t.sf * a.s6 * a.s7))); return CONVERT(fr, VEC_DATA_TYPE(DATA_TYPE, 4)); } @@ -132,7 +130,8 @@ inline const VEC_DATA_TYPE(DATA_TYPE, 4) bilinear_interpolate_with_border(const * @param[in] width Width of the image * @param[in] height Height of the image */ -inline const VEC_DATA_TYPE(DATA_TYPE, 4) bilinear_interpolate(const Image *in, const float8 coords, const float width, const float height) +inline const VEC_DATA_TYPE(DATA_TYPE, 4) + bilinear_interpolate(const Image *in, const float8 coords, const float width, const float height) { return bilinear_interpolate_with_border(in, coords, width, height, 1); } |