| 1 |
|
| |
| 2 |
|
| #line 453 "lib/PDL/PP.pm" |
| 3 |
|
| /* |
| 4 |
|
| * THIS FILE WAS GENERATED BY PDL::PP from lib/PDL/ImageND.pd! Do not modify! |
| 5 |
|
| */ |
| 6 |
|
| |
| 7 |
|
| #define PDL_FREE_CODE(trans, destroy, comp_free_code, ntpriv_free_code) \ |
| 8 |
|
| if (destroy) { \ |
| 9 |
|
| comp_free_code \ |
| 10 |
|
| } \ |
| 11 |
|
| if ((trans)->dims_redone) { \ |
| 12 |
|
| ntpriv_free_code \ |
| 13 |
|
| } |
| 14 |
|
| |
| 15 |
|
| #include "EXTERN.h" |
| 16 |
|
| #include "perl.h" |
| 17 |
|
| #include "XSUB.h" |
| 18 |
|
| #include "pdl.h" |
| 19 |
|
| #include "pdlcore.h" |
| 20 |
|
| #define PDL PDL_ImageND |
| 21 |
|
| extern Core* PDL; /* Structure hold core C functions */ |
| 22 |
|
| |
| 23 |
|
| #line 1877 "lib/PDL/PP.pm" |
| 24 |
|
| typedef struct pdl_params_convolveND { |
| 25 |
|
| #line 26 "lib/PDL/ImageND-pp-convolveND.c" |
| 26 |
|
| pdl *k; |
| 27 |
|
| pdl *aa; |
| 28 |
|
| pdl *a; |
| 29 |
|
| } pdl_params_convolveND; |
| 30 |
|
| |
| 31 |
|
| |
| 32 |
|
| #line 1888 "lib/PDL/PP.pm" |
| 33 |
|
| pdl_error pdl_convolveND_readdata(pdl_trans *__privtrans) { |
| 34 |
|
| pdl_error PDL_err = {0, NULL, 0}; |
| 35 |
|
| #line 36 "lib/PDL/ImageND-pp-convolveND.c" |
| 36 |
3 |
| pdl_params_convolveND *__params = __privtrans->params; (void)__params; |
| 37 |
3 |
▶ | if (!__privtrans->broadcast.incs) return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "broadcast.incs NULL"); |
Branch: 1 branch
| Branch | True | False | Total |
| if (!__privtrans->broadcast.incs) return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "broadcast.incs NULL"); |
0 |
3 |
3 |
|
| 38 |
|
| /* broadcastloop declarations */ |
| 39 |
|
| int __brcloopval; |
| 40 |
|
| register PDL_Indx __tind0,__tind1; /* counters along dim */ |
| 41 |
3 |
| register PDL_Indx __tnpdls = __privtrans->broadcast.npdls; |
| 42 |
|
| /* dims here are how many steps along those dims */ |
| 43 |
3 |
| register PDL_Indx __tinc0_k0 = PDL_BRC_INC(__privtrans->broadcast.incs,__tnpdls,0,0); |
| 44 |
3 |
| register PDL_Indx __tinc1_k0 = PDL_BRC_INC(__privtrans->broadcast.incs,__tnpdls,0,1); |
| 45 |
|
| #define PDL_BROADCASTLOOP_START_convolveND_readdata PDL_BROADCASTLOOP_START( \ |
| 46 |
|
| readdata, \ |
| 47 |
|
| __privtrans->broadcast, \ |
| 48 |
|
| __privtrans->vtable, \ |
| 49 |
|
| k0_datap += __offsp[0]; \ |
| 50 |
|
| , \ |
| 51 |
|
| ( ,k0_datap += __tinc1_k0 - __tinc0_k0 * __tdims0 \ |
| 52 |
|
| ), \ |
| 53 |
|
| ( ,k0_datap += __tinc0_k0 \ |
| 54 |
|
| ) \ |
| 55 |
|
| ) |
| 56 |
|
| #define PDL_BROADCASTLOOP_END_convolveND_readdata PDL_BROADCASTLOOP_END( \ |
| 57 |
|
| __privtrans->broadcast, \ |
| 58 |
|
| k0_datap -= __tinc1_k0 * __tdims1 + __offsp[0]; \ |
| 59 |
|
| ) |
| 60 |
|
| #ifndef PDL_DECLARE_PARAMS_convolveND_1 |
| 61 |
|
| #define PDL_DECLARE_PARAMS_convolveND_1(PDL_TYPE_OP,PDL_PPSYM_OP) \ |
| 62 |
|
| PDL_DECLARE_PARAMETER(PDL_TYPE_OP, k0, (__privtrans->pdls[0]), 1, PDL_PPSYM_OP) |
| 63 |
|
| #endif |
| 64 |
|
| #define PDL_IF_BAD(t,f) f |
| 65 |
3 |
| switch (__privtrans->__datatype) { /* Start generic switch */ |
| 66 |
0 |
| case PDL_SB: { |
| 67 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_SByte,A) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_SByte,A) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_SByte,A) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_SByte,A) |
0 |
0 |
0 |
|
| 68 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 69 |
|
| * Direct convolution |
| 70 |
|
| * |
| 71 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 72 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 73 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 74 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 75 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 76 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 77 |
|
| * |
| 78 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 79 |
|
| * macro. The other three arguments should all have the same type as the |
| 80 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 81 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 82 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 83 |
|
| * |
| 84 |
|
| */ |
| 85 |
|
| |
| 86 |
|
| PDL_Indx i; |
| 87 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 88 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 89 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 90 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 91 |
|
| |
| 92 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 93 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 94 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 95 |
|
| |
| 96 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 97 |
0 |
| PDL_SByte kvals[k->nvals]; |
| 98 |
0 |
| PDL_Indx ivec[ndims]; |
| 99 |
|
| |
| 100 |
|
| /************************************/ |
| 101 |
|
| /* Fill up the koffs & kvals arrays */ |
| 102 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 103 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 104 |
0 |
| PDL_Indx *koff = koffs; |
| 105 |
0 |
| PDL_SByte *kval = kvals; |
| 106 |
0 |
| PDL_SByte *aptr = (PDL_SByte *)k->data + k->nvals - 1; |
| 107 |
|
| |
| 108 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 109 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 110 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 111 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 112 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 113 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 114 |
|
| } |
| 115 |
|
| do { |
| 116 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 117 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 118 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 119 |
|
| } while (1); |
| 120 |
|
| |
| 121 |
|
| /******************************/ |
| 122 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 123 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 124 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 125 |
0 |
| aptr = a->data; |
| 126 |
0 |
| PDL_SByte *aaptr = aa->data; |
| 127 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 128 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 129 |
0 |
| do { |
| 130 |
0 |
| PDL_SByte acc = 0; |
| 131 |
0 |
| koff = koffs; |
| 132 |
0 |
| kval = kvals; |
| 133 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 134 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 135 |
0 |
| aptr[offs[0]] = acc; |
| 136 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 137 |
|
| } while (1); |
| 138 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 139 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 140 |
0 |
| } break; |
| 141 |
0 |
| case PDL_B: { |
| 142 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_Byte,B) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Byte,B) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Byte,B) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Byte,B) |
0 |
0 |
0 |
|
| 143 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 144 |
|
| * Direct convolution |
| 145 |
|
| * |
| 146 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 147 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 148 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 149 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 150 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 151 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 152 |
|
| * |
| 153 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 154 |
|
| * macro. The other three arguments should all have the same type as the |
| 155 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 156 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 157 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 158 |
|
| * |
| 159 |
|
| */ |
| 160 |
|
| |
| 161 |
|
| PDL_Indx i; |
| 162 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 163 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 164 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 165 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 166 |
|
| |
| 167 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 168 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 169 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 170 |
|
| |
| 171 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 172 |
0 |
| PDL_Byte kvals[k->nvals]; |
| 173 |
0 |
| PDL_Indx ivec[ndims]; |
| 174 |
|
| |
| 175 |
|
| /************************************/ |
| 176 |
|
| /* Fill up the koffs & kvals arrays */ |
| 177 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 178 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 179 |
0 |
| PDL_Indx *koff = koffs; |
| 180 |
0 |
| PDL_Byte *kval = kvals; |
| 181 |
0 |
| PDL_Byte *aptr = (PDL_Byte *)k->data + k->nvals - 1; |
| 182 |
|
| |
| 183 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 184 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 185 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 186 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 187 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 188 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 189 |
|
| } |
| 190 |
|
| do { |
| 191 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 192 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 193 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 194 |
|
| } while (1); |
| 195 |
|
| |
| 196 |
|
| /******************************/ |
| 197 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 198 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 199 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 200 |
0 |
| aptr = a->data; |
| 201 |
0 |
| PDL_Byte *aaptr = aa->data; |
| 202 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 203 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 204 |
0 |
| do { |
| 205 |
0 |
| PDL_Byte acc = 0; |
| 206 |
0 |
| koff = koffs; |
| 207 |
0 |
| kval = kvals; |
| 208 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 209 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 210 |
0 |
| aptr[offs[0]] = acc; |
| 211 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 212 |
|
| } while (1); |
| 213 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 214 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 215 |
0 |
| } break; |
| 216 |
0 |
| case PDL_S: { |
| 217 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_Short,S) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Short,S) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Short,S) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Short,S) |
0 |
0 |
0 |
|
| 218 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 219 |
|
| * Direct convolution |
| 220 |
|
| * |
| 221 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 222 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 223 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 224 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 225 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 226 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 227 |
|
| * |
| 228 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 229 |
|
| * macro. The other three arguments should all have the same type as the |
| 230 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 231 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 232 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 233 |
|
| * |
| 234 |
|
| */ |
| 235 |
|
| |
| 236 |
|
| PDL_Indx i; |
| 237 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 238 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 239 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 240 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 241 |
|
| |
| 242 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 243 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 244 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 245 |
|
| |
| 246 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 247 |
0 |
| PDL_Short kvals[k->nvals]; |
| 248 |
0 |
| PDL_Indx ivec[ndims]; |
| 249 |
|
| |
| 250 |
|
| /************************************/ |
| 251 |
|
| /* Fill up the koffs & kvals arrays */ |
| 252 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 253 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 254 |
0 |
| PDL_Indx *koff = koffs; |
| 255 |
0 |
| PDL_Short *kval = kvals; |
| 256 |
0 |
| PDL_Short *aptr = (PDL_Short *)k->data + k->nvals - 1; |
| 257 |
|
| |
| 258 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 259 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 260 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 261 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 262 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 263 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 264 |
|
| } |
| 265 |
|
| do { |
| 266 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 267 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 268 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 269 |
|
| } while (1); |
| 270 |
|
| |
| 271 |
|
| /******************************/ |
| 272 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 273 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 274 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 275 |
0 |
| aptr = a->data; |
| 276 |
0 |
| PDL_Short *aaptr = aa->data; |
| 277 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 278 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 279 |
0 |
| do { |
| 280 |
0 |
| PDL_Short acc = 0; |
| 281 |
0 |
| koff = koffs; |
| 282 |
0 |
| kval = kvals; |
| 283 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 284 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 285 |
0 |
| aptr[offs[0]] = acc; |
| 286 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 287 |
|
| } while (1); |
| 288 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 289 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 290 |
0 |
| } break; |
| 291 |
0 |
| case PDL_US: { |
| 292 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_Ushort,U) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Ushort,U) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Ushort,U) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Ushort,U) |
0 |
0 |
0 |
|
| 293 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 294 |
|
| * Direct convolution |
| 295 |
|
| * |
| 296 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 297 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 298 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 299 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 300 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 301 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 302 |
|
| * |
| 303 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 304 |
|
| * macro. The other three arguments should all have the same type as the |
| 305 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 306 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 307 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 308 |
|
| * |
| 309 |
|
| */ |
| 310 |
|
| |
| 311 |
|
| PDL_Indx i; |
| 312 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 313 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 314 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 315 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 316 |
|
| |
| 317 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 318 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 319 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 320 |
|
| |
| 321 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 322 |
0 |
| PDL_Ushort kvals[k->nvals]; |
| 323 |
0 |
| PDL_Indx ivec[ndims]; |
| 324 |
|
| |
| 325 |
|
| /************************************/ |
| 326 |
|
| /* Fill up the koffs & kvals arrays */ |
| 327 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 328 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 329 |
0 |
| PDL_Indx *koff = koffs; |
| 330 |
0 |
| PDL_Ushort *kval = kvals; |
| 331 |
0 |
| PDL_Ushort *aptr = (PDL_Ushort *)k->data + k->nvals - 1; |
| 332 |
|
| |
| 333 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 334 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 335 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 336 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 337 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 338 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 339 |
|
| } |
| 340 |
|
| do { |
| 341 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 342 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 343 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 344 |
|
| } while (1); |
| 345 |
|
| |
| 346 |
|
| /******************************/ |
| 347 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 348 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 349 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 350 |
0 |
| aptr = a->data; |
| 351 |
0 |
| PDL_Ushort *aaptr = aa->data; |
| 352 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 353 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 354 |
0 |
| do { |
| 355 |
0 |
| PDL_Ushort acc = 0; |
| 356 |
0 |
| koff = koffs; |
| 357 |
0 |
| kval = kvals; |
| 358 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 359 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 360 |
0 |
| aptr[offs[0]] = acc; |
| 361 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 362 |
|
| } while (1); |
| 363 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 364 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 365 |
0 |
| } break; |
| 366 |
0 |
| case PDL_L: { |
| 367 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_Long,L) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Long,L) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Long,L) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Long,L) |
0 |
0 |
0 |
|
| 368 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 369 |
|
| * Direct convolution |
| 370 |
|
| * |
| 371 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 372 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 373 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 374 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 375 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 376 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 377 |
|
| * |
| 378 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 379 |
|
| * macro. The other three arguments should all have the same type as the |
| 380 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 381 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 382 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 383 |
|
| * |
| 384 |
|
| */ |
| 385 |
|
| |
| 386 |
|
| PDL_Indx i; |
| 387 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 388 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 389 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 390 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 391 |
|
| |
| 392 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 393 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 394 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 395 |
|
| |
| 396 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 397 |
0 |
| PDL_Long kvals[k->nvals]; |
| 398 |
0 |
| PDL_Indx ivec[ndims]; |
| 399 |
|
| |
| 400 |
|
| /************************************/ |
| 401 |
|
| /* Fill up the koffs & kvals arrays */ |
| 402 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 403 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 404 |
0 |
| PDL_Indx *koff = koffs; |
| 405 |
0 |
| PDL_Long *kval = kvals; |
| 406 |
0 |
| PDL_Long *aptr = (PDL_Long *)k->data + k->nvals - 1; |
| 407 |
|
| |
| 408 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 409 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 410 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 411 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 412 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 413 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 414 |
|
| } |
| 415 |
|
| do { |
| 416 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 417 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 418 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 419 |
|
| } while (1); |
| 420 |
|
| |
| 421 |
|
| /******************************/ |
| 422 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 423 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 424 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 425 |
0 |
| aptr = a->data; |
| 426 |
0 |
| PDL_Long *aaptr = aa->data; |
| 427 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 428 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 429 |
0 |
| do { |
| 430 |
0 |
| PDL_Long acc = 0; |
| 431 |
0 |
| koff = koffs; |
| 432 |
0 |
| kval = kvals; |
| 433 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 434 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 435 |
0 |
| aptr[offs[0]] = acc; |
| 436 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 437 |
|
| } while (1); |
| 438 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 439 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 440 |
0 |
| } break; |
| 441 |
0 |
| case PDL_UL: { |
| 442 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_ULong,K) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_ULong,K) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_ULong,K) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_ULong,K) |
0 |
0 |
0 |
|
| 443 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 444 |
|
| * Direct convolution |
| 445 |
|
| * |
| 446 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 447 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 448 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 449 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 450 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 451 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 452 |
|
| * |
| 453 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 454 |
|
| * macro. The other three arguments should all have the same type as the |
| 455 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 456 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 457 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 458 |
|
| * |
| 459 |
|
| */ |
| 460 |
|
| |
| 461 |
|
| PDL_Indx i; |
| 462 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 463 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 464 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 465 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 466 |
|
| |
| 467 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 468 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 469 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 470 |
|
| |
| 471 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 472 |
0 |
| PDL_ULong kvals[k->nvals]; |
| 473 |
0 |
| PDL_Indx ivec[ndims]; |
| 474 |
|
| |
| 475 |
|
| /************************************/ |
| 476 |
|
| /* Fill up the koffs & kvals arrays */ |
| 477 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 478 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 479 |
0 |
| PDL_Indx *koff = koffs; |
| 480 |
0 |
| PDL_ULong *kval = kvals; |
| 481 |
0 |
| PDL_ULong *aptr = (PDL_ULong *)k->data + k->nvals - 1; |
| 482 |
|
| |
| 483 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 484 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 485 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 486 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 487 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 488 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 489 |
|
| } |
| 490 |
|
| do { |
| 491 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 492 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 493 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 494 |
|
| } while (1); |
| 495 |
|
| |
| 496 |
|
| /******************************/ |
| 497 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 498 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 499 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 500 |
0 |
| aptr = a->data; |
| 501 |
0 |
| PDL_ULong *aaptr = aa->data; |
| 502 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 503 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 504 |
0 |
| do { |
| 505 |
0 |
| PDL_ULong acc = 0; |
| 506 |
0 |
| koff = koffs; |
| 507 |
0 |
| kval = kvals; |
| 508 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 509 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 510 |
0 |
| aptr[offs[0]] = acc; |
| 511 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 512 |
|
| } while (1); |
| 513 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 514 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 515 |
0 |
| } break; |
| 516 |
0 |
| case PDL_IND: { |
| 517 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_Indx,N) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Indx,N) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Indx,N) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Indx,N) |
0 |
0 |
0 |
|
| 518 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 519 |
|
| * Direct convolution |
| 520 |
|
| * |
| 521 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 522 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 523 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 524 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 525 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 526 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 527 |
|
| * |
| 528 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 529 |
|
| * macro. The other three arguments should all have the same type as the |
| 530 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 531 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 532 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 533 |
|
| * |
| 534 |
|
| */ |
| 535 |
|
| |
| 536 |
|
| PDL_Indx i; |
| 537 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 538 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 539 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 540 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 541 |
|
| |
| 542 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 543 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 544 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 545 |
|
| |
| 546 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 547 |
0 |
| PDL_Indx kvals[k->nvals]; |
| 548 |
0 |
| PDL_Indx ivec[ndims]; |
| 549 |
|
| |
| 550 |
|
| /************************************/ |
| 551 |
|
| /* Fill up the koffs & kvals arrays */ |
| 552 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 553 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 554 |
0 |
| PDL_Indx *koff = koffs; |
| 555 |
0 |
| PDL_Indx *kval = kvals; |
| 556 |
0 |
| PDL_Indx *aptr = (PDL_Indx *)k->data + k->nvals - 1; |
| 557 |
|
| |
| 558 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 559 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 560 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 561 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 562 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 563 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 564 |
|
| } |
| 565 |
|
| do { |
| 566 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 567 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 568 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 569 |
|
| } while (1); |
| 570 |
|
| |
| 571 |
|
| /******************************/ |
| 572 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 573 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 574 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 575 |
0 |
| aptr = a->data; |
| 576 |
0 |
| PDL_Indx *aaptr = aa->data; |
| 577 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 578 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 579 |
0 |
| do { |
| 580 |
0 |
| PDL_Indx acc = 0; |
| 581 |
0 |
| koff = koffs; |
| 582 |
0 |
| kval = kvals; |
| 583 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 584 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 585 |
0 |
| aptr[offs[0]] = acc; |
| 586 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 587 |
|
| } while (1); |
| 588 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 589 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 590 |
0 |
| } break; |
| 591 |
0 |
| case PDL_ULL: { |
| 592 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_ULongLong,P) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_ULongLong,P) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_ULongLong,P) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_ULongLong,P) |
0 |
0 |
0 |
|
| 593 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 594 |
|
| * Direct convolution |
| 595 |
|
| * |
| 596 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 597 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 598 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 599 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 600 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 601 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 602 |
|
| * |
| 603 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 604 |
|
| * macro. The other three arguments should all have the same type as the |
| 605 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 606 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 607 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 608 |
|
| * |
| 609 |
|
| */ |
| 610 |
|
| |
| 611 |
|
| PDL_Indx i; |
| 612 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 613 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 614 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 615 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 616 |
|
| |
| 617 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 618 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 619 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 620 |
|
| |
| 621 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 622 |
0 |
| PDL_ULongLong kvals[k->nvals]; |
| 623 |
0 |
| PDL_Indx ivec[ndims]; |
| 624 |
|
| |
| 625 |
|
| /************************************/ |
| 626 |
|
| /* Fill up the koffs & kvals arrays */ |
| 627 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 628 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 629 |
0 |
| PDL_Indx *koff = koffs; |
| 630 |
0 |
| PDL_ULongLong *kval = kvals; |
| 631 |
0 |
| PDL_ULongLong *aptr = (PDL_ULongLong *)k->data + k->nvals - 1; |
| 632 |
|
| |
| 633 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 634 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 635 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 636 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 637 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 638 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 639 |
|
| } |
| 640 |
|
| do { |
| 641 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 642 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 643 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 644 |
|
| } while (1); |
| 645 |
|
| |
| 646 |
|
| /******************************/ |
| 647 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 648 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 649 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 650 |
0 |
| aptr = a->data; |
| 651 |
0 |
| PDL_ULongLong *aaptr = aa->data; |
| 652 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 653 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 654 |
0 |
| do { |
| 655 |
0 |
| PDL_ULongLong acc = 0; |
| 656 |
0 |
| koff = koffs; |
| 657 |
0 |
| kval = kvals; |
| 658 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 659 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 660 |
0 |
| aptr[offs[0]] = acc; |
| 661 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 662 |
|
| } while (1); |
| 663 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 664 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 665 |
0 |
| } break; |
| 666 |
0 |
| case PDL_LL: { |
| 667 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_LongLong,Q) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_LongLong,Q) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_LongLong,Q) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_LongLong,Q) |
0 |
0 |
0 |
|
| 668 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 669 |
|
| * Direct convolution |
| 670 |
|
| * |
| 671 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 672 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 673 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 674 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 675 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 676 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 677 |
|
| * |
| 678 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 679 |
|
| * macro. The other three arguments should all have the same type as the |
| 680 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 681 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 682 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 683 |
|
| * |
| 684 |
|
| */ |
| 685 |
|
| |
| 686 |
|
| PDL_Indx i; |
| 687 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 688 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 689 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 690 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 691 |
|
| |
| 692 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 693 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 694 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 695 |
|
| |
| 696 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 697 |
0 |
| PDL_LongLong kvals[k->nvals]; |
| 698 |
0 |
| PDL_Indx ivec[ndims]; |
| 699 |
|
| |
| 700 |
|
| /************************************/ |
| 701 |
|
| /* Fill up the koffs & kvals arrays */ |
| 702 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 703 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 704 |
0 |
| PDL_Indx *koff = koffs; |
| 705 |
0 |
| PDL_LongLong *kval = kvals; |
| 706 |
0 |
| PDL_LongLong *aptr = (PDL_LongLong *)k->data + k->nvals - 1; |
| 707 |
|
| |
| 708 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 709 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 710 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 711 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 712 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 713 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 714 |
|
| } |
| 715 |
|
| do { |
| 716 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 717 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 718 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 719 |
|
| } while (1); |
| 720 |
|
| |
| 721 |
|
| /******************************/ |
| 722 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 723 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 724 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 725 |
0 |
| aptr = a->data; |
| 726 |
0 |
| PDL_LongLong *aaptr = aa->data; |
| 727 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 728 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 729 |
0 |
| do { |
| 730 |
0 |
| PDL_LongLong acc = 0; |
| 731 |
0 |
| koff = koffs; |
| 732 |
0 |
| kval = kvals; |
| 733 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 734 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 735 |
0 |
| aptr[offs[0]] = acc; |
| 736 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 737 |
|
| } while (1); |
| 738 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 739 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 740 |
0 |
| } break; |
| 741 |
0 |
| case PDL_F: { |
| 742 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_Float,F) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Float,F) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Float,F) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Float,F) |
0 |
0 |
0 |
|
| 743 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 744 |
|
| * Direct convolution |
| 745 |
|
| * |
| 746 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 747 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 748 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 749 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 750 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 751 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 752 |
|
| * |
| 753 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 754 |
|
| * macro. The other three arguments should all have the same type as the |
| 755 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 756 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 757 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 758 |
|
| * |
| 759 |
|
| */ |
| 760 |
|
| |
| 761 |
|
| PDL_Indx i; |
| 762 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 763 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 764 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 765 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 766 |
|
| |
| 767 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 768 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 769 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 770 |
|
| |
| 771 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 772 |
0 |
| PDL_Float kvals[k->nvals]; |
| 773 |
0 |
| PDL_Indx ivec[ndims]; |
| 774 |
|
| |
| 775 |
|
| /************************************/ |
| 776 |
|
| /* Fill up the koffs & kvals arrays */ |
| 777 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 778 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 779 |
0 |
| PDL_Indx *koff = koffs; |
| 780 |
0 |
| PDL_Float *kval = kvals; |
| 781 |
0 |
| PDL_Float *aptr = (PDL_Float *)k->data + k->nvals - 1; |
| 782 |
|
| |
| 783 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 784 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 785 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 786 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 787 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 788 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 789 |
|
| } |
| 790 |
|
| do { |
| 791 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 792 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 793 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 794 |
|
| } while (1); |
| 795 |
|
| |
| 796 |
|
| /******************************/ |
| 797 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 798 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 799 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 800 |
0 |
| aptr = a->data; |
| 801 |
0 |
| PDL_Float *aaptr = aa->data; |
| 802 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 803 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 804 |
0 |
| do { |
| 805 |
0 |
| PDL_Float acc = 0; |
| 806 |
0 |
| koff = koffs; |
| 807 |
0 |
| kval = kvals; |
| 808 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 809 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 810 |
0 |
| aptr[offs[0]] = acc; |
| 811 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 812 |
|
| } while (1); |
| 813 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 814 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 815 |
0 |
| } break; |
| 816 |
3 |
| case PDL_D: { |
| 817 |
3 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_Double,D) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Double,D) |
0 |
3 |
3 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Double,D) |
3 |
0 |
3 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_Double,D) |
0 |
3 |
3 |
|
| 818 |
12 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
3 |
3 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
3 |
3 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
3 |
3 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
3 |
3 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
3 |
3 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
3 |
3 |
6 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
3 |
3 |
6 |
|
| 819 |
|
| * Direct convolution |
| 820 |
|
| * |
| 821 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 822 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 823 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 824 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 825 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 826 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 827 |
|
| * |
| 828 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 829 |
|
| * macro. The other three arguments should all have the same type as the |
| 830 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 831 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 832 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 833 |
|
| * |
| 834 |
|
| */ |
| 835 |
|
| |
| 836 |
|
| PDL_Indx i; |
| 837 |
3 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 838 |
3 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
3 |
3 |
|
| 839 |
3 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
3 |
3 |
|
| 840 |
3 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
3 |
3 |
|
| 841 |
|
| |
| 842 |
3 |
| PDL_Indx ndims = aa->ndims; |
| 843 |
3 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
3 |
0 |
3 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
3 |
3 |
|
| 844 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 845 |
|
| |
| 846 |
3 |
| PDL_Indx koffs[k->nvals]; |
| 847 |
3 |
| PDL_Double kvals[k->nvals]; |
| 848 |
3 |
| PDL_Indx ivec[ndims]; |
| 849 |
|
| |
| 850 |
|
| /************************************/ |
| 851 |
|
| /* Fill up the koffs & kvals arrays */ |
| 852 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 853 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 854 |
3 |
| PDL_Indx *koff = koffs; |
| 855 |
3 |
| PDL_Double *kval = kvals; |
| 856 |
3 |
| PDL_Double *aptr = (PDL_Double *)k->data + k->nvals - 1; |
| 857 |
|
| |
| 858 |
9 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
6 |
3 |
9 |
|
| 859 |
3 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 860 |
9 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
6 |
3 |
9 |
|
| 861 |
9 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
6 |
3 |
9 |
|
| 862 |
6 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 863 |
6 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 864 |
|
| } |
| 865 |
|
| do { |
| 866 |
12 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 867 |
12 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 868 |
12 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
3 |
9 |
12 |
|
| 869 |
|
| } while (1); |
| 870 |
|
| |
| 871 |
|
| /******************************/ |
| 872 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 873 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 874 |
9 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
6 |
3 |
9 |
|
| 875 |
3 |
| aptr = a->data; |
| 876 |
3 |
| PDL_Double *aaptr = aa->data; |
| 877 |
9 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
6 |
3 |
9 |
|
| 878 |
9 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
6 |
3 |
9 |
|
| 879 |
105 |
| do { |
| 880 |
108 |
| PDL_Double acc = 0; |
| 881 |
108 |
| koff = koffs; |
| 882 |
108 |
| kval = kvals; |
| 883 |
540 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
432 |
108 |
540 |
|
| 884 |
432 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 885 |
108 |
| aptr[offs[0]] = acc; |
| 886 |
108 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
3 |
105 |
108 |
|
| 887 |
|
| } while (1); |
| 888 |
3 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 889 |
3 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
3 |
3 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
3 |
3 |
|
| 890 |
3 |
| } break; |
| 891 |
0 |
| case PDL_LD: { |
| 892 |
0 |
▶ | PDL_DECLARE_PARAMS_convolveND_1(PDL_LDouble,E) |
Branch: 3 branches
| Branch | True | False | Total |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_LDouble,E) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_LDouble,E) |
0 |
0 |
0 |
| PDL_DECLARE_PARAMS_convolveND_1(PDL_LDouble,E) |
0 |
0 |
0 |
|
| 893 |
0 |
▶ | PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
Branch: 7 branches
| Branch | True | False | Total |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
| PDL_BROADCASTLOOP_START_convolveND_readdata {/* |
0 |
0 |
0 |
|
| 894 |
|
| * Direct convolution |
| 895 |
|
| * |
| 896 |
|
| * Because the kernel is usually the smaller of the two arrays to be convolved, |
| 897 |
|
| * we broadcast kernel-first to keep it in the processor's cache. The strategy: |
| 898 |
|
| * work on a padded copy of the original image, so that (even with boundary |
| 899 |
|
| * conditions) the geometry of the kernel is linearly related to the input |
| 900 |
|
| * array. Otherwise, follow the path blazed by Karl in convolve(): keep track |
| 901 |
|
| * of the offsets for each kernel element in a flattened original PDL. |
| 902 |
|
| * |
| 903 |
|
| * The first (PP) argument is a dummy that's only used to set the GENERIC() |
| 904 |
|
| * macro. The other three arguments should all have the same type as the |
| 905 |
|
| * first arguments, and are all passed in as SVs. They are: the kernel, |
| 906 |
|
| * the padded copy of the input PDL, and a pre-allocated output PDL. The |
| 907 |
|
| * input PDL should be padded by the dimensionality of the kernel. |
| 908 |
|
| * |
| 909 |
|
| */ |
| 910 |
|
| |
| 911 |
|
| PDL_Indx i; |
| 912 |
0 |
| pdl *k = __params->k, *a = __params->a, *aa = __params->aa; |
| 913 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(aa)); |
0 |
0 |
0 |
|
| 914 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(a)); |
0 |
0 |
0 |
|
| 915 |
0 |
▶ | PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_physical(k)); |
0 |
0 |
0 |
|
| 916 |
|
| |
| 917 |
0 |
| PDL_Indx ndims = aa->ndims; |
| 918 |
0 |
▶ | if(ndims != k->ndims || ndims != aa->ndims) |
Branch: 2 branches
| Branch | True | False | Total |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
| if(ndims != k->ndims || ndims != aa->ndims) |
0 |
0 |
0 |
|
| 919 |
0 |
| return PDL->make_error(PDL_EUSERERROR, "Error in convolveND:" "convolveND: dims don't agree - should never happen\n"); |
| 920 |
|
| |
| 921 |
0 |
| PDL_Indx koffs[k->nvals]; |
| 922 |
0 |
| PDL_LDouble kvals[k->nvals]; |
| 923 |
0 |
| PDL_Indx ivec[ndims]; |
| 924 |
|
| |
| 925 |
|
| /************************************/ |
| 926 |
|
| /* Fill up the koffs & kvals arrays */ |
| 927 |
|
| /* koffs gets relative offsets into aa for each kernel value; */ |
| 928 |
|
| /* kvals gets the kernel values in the same order (flattened) */ |
| 929 |
0 |
| PDL_Indx *koff = koffs; |
| 930 |
0 |
| PDL_LDouble *kval = kvals; |
| 931 |
0 |
| PDL_LDouble *aptr = (PDL_LDouble *)k->data + k->nvals - 1; |
| 932 |
|
| |
| 933 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 934 |
0 |
| PDL_Indx npdls = 2, incs[npdls*ndims], offs[npdls]; |
| 935 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 936 |
0 |
▶ | for (i=0; i < ndims; i++) { |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) { |
0 |
0 |
0 |
|
| 937 |
0 |
| incs[i*npdls + 0] = k->dimincs[i]; |
| 938 |
0 |
| incs[i*npdls + 1] = aa->dimincs[i]; |
| 939 |
|
| } |
| 940 |
|
| do { |
| 941 |
0 |
| *kval++ = aptr[-offs[0]]; /* Copy kernel value into kernel list */ |
| 942 |
0 |
| *koff++ = offs[1]; /* Copy current aa offset into koffs list */ |
| 943 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, k->dims, ivec)) break; |
0 |
0 |
0 |
|
| 944 |
|
| } while (1); |
| 945 |
|
| |
| 946 |
|
| /******************************/ |
| 947 |
|
| /* Now do the actual convolution: for each vector in a, */ |
| 948 |
|
| /* accumulate the appropriate aa-sum and stick it into a. */ |
| 949 |
0 |
▶ | for (i=0; i < ndims; i++) ivec[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) ivec[i] = 0; |
0 |
0 |
0 |
|
| 950 |
0 |
| aptr = a->data; |
| 951 |
0 |
| PDL_LDouble *aaptr = aa->data; |
| 952 |
0 |
▶ | for (i=0; i < npdls; i++) offs[i] = 0; |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < npdls; i++) offs[i] = 0; |
0 |
0 |
0 |
|
| 953 |
0 |
▶ | for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0; i < ndims; i++) incs[i*npdls + 0] = a->dimincs[i]; /* got aa already */ |
0 |
0 |
0 |
|
| 954 |
0 |
| do { |
| 955 |
0 |
| PDL_LDouble acc = 0; |
| 956 |
0 |
| koff = koffs; |
| 957 |
0 |
| kval = kvals; |
| 958 |
0 |
▶ | for (i=0;i<k->nvals;i++) |
Branch: 1 branch
| Branch | True | False | Total |
| for (i=0;i<k->nvals;i++) |
0 |
0 |
0 |
|
| 959 |
0 |
| acc += aaptr[offs[1] + *koff++] * (*kval++); |
| 960 |
0 |
| aptr[offs[0]] = acc; |
| 961 |
0 |
▶ | if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!pdl_broadcast_nd_step(npdls, offs, 0, ndims, incs, a->dims, ivec)) break; |
0 |
0 |
0 |
|
| 962 |
|
| } while (1); |
| 963 |
0 |
| PDL->changed(a, PDL_PARENTDATACHANGED, 0); |
| 964 |
0 |
▶ | }PDL_BROADCASTLOOP_END_convolveND_readdata |
Branch: 2 branches
| Branch | True | False | Total |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
| }PDL_BROADCASTLOOP_END_convolveND_readdata |
0 |
0 |
0 |
|
| 965 |
0 |
| } break; |
| 966 |
0 |
| default: return PDL->make_error(PDL_EUSERERROR, "PP INTERNAL ERROR in convolveND: unhandled datatype(%d), only handles (ABSULKNPQFDE)! PLEASE MAKE A BUG REPORT\n", __privtrans->__datatype); |
| 967 |
|
| } |
| 968 |
|
| #undef PDL_IF_BAD |
| 969 |
3 |
| return PDL_err; |
| 970 |
|
| } |
| 971 |
|
| |
| 972 |
|
| static pdl_datatypes pdl_convolveND_vtable_gentypes[] = { PDL_SB, PDL_B, PDL_S, PDL_US, PDL_L, PDL_UL, PDL_IND, PDL_ULL, PDL_LL, PDL_F, PDL_D, PDL_LD, -1 }; |
| 973 |
|
| static PDL_Indx pdl_convolveND_vtable_realdims[] = { 0 }; |
| 974 |
|
| static char *pdl_convolveND_vtable_parnames[] = { "k0" }; |
| 975 |
|
| static short pdl_convolveND_vtable_parflags[] = { |
| 976 |
|
| 0 |
| 977 |
|
| }; |
| 978 |
|
| static pdl_datatypes pdl_convolveND_vtable_partypes[] = { -1 }; |
| 979 |
|
| static PDL_Indx pdl_convolveND_vtable_realdims_starts[] = { 0 }; |
| 980 |
|
| static PDL_Indx pdl_convolveND_vtable_realdims_ind_ids[] = { 0 }; |
| 981 |
|
| static char *pdl_convolveND_vtable_indnames[] = { "" }; |
| 982 |
|
| pdl_transvtable pdl_convolveND_vtable = { |
| 983 |
|
| PDL_TRANS_DO_BROADCAST, 0, pdl_convolveND_vtable_gentypes, 1, 1, NULL /*CORE21*/, |
| 984 |
|
| pdl_convolveND_vtable_realdims, pdl_convolveND_vtable_parnames, |
| 985 |
|
| pdl_convolveND_vtable_parflags, pdl_convolveND_vtable_partypes, |
| 986 |
|
| pdl_convolveND_vtable_realdims_starts, pdl_convolveND_vtable_realdims_ind_ids, 0, |
| 987 |
|
| 0, pdl_convolveND_vtable_indnames, |
| 988 |
|
| NULL, pdl_convolveND_readdata, NULL, |
| 989 |
|
| NULL, |
| 990 |
|
| sizeof(pdl_params_convolveND),"PDL::ImageND::convolveND" |
| 991 |
|
| }; |
| 992 |
|
| |
| 993 |
|
| |
| 994 |
3 |
| pdl_error pdl_run_convolveND(pdl *k0,pdl *k,pdl *aa,pdl *a) { |
| 995 |
3 |
| pdl_error PDL_err = {0, NULL, 0}; |
| 996 |
3 |
▶ | if (!PDL) return (pdl_error){PDL_EFATAL, "PDL core struct is NULL, can't continue",0}; |
Branch: 1 branch
| Branch | True | False | Total |
| if (!PDL) return (pdl_error){PDL_EFATAL, "PDL core struct is NULL, can't continue",0}; |
0 |
3 |
3 |
|
| 997 |
3 |
| pdl_trans *__privtrans = PDL->create_trans(&pdl_convolveND_vtable); |
| 998 |
3 |
▶ | if (!__privtrans) return PDL->make_error_simple(PDL_EFATAL, "Couldn't create trans"); |
Branch: 1 branch
| Branch | True | False | Total |
| if (!__privtrans) return PDL->make_error_simple(PDL_EFATAL, "Couldn't create trans"); |
0 |
3 |
3 |
|
| 999 |
3 |
| pdl_params_convolveND *__params = __privtrans->params; |
| 1000 |
3 |
| __privtrans->pdls[0] = k0; |
| 1001 |
3 |
▶ | PDL_RETERROR(PDL_err, PDL->type_coerce(__privtrans)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->type_coerce(__privtrans)); |
0 |
3 |
3 |
|
| 1002 |
3 |
| (__params->k) = (k); /* CType.get_copy */ |
| 1003 |
3 |
| (__params->aa) = (aa); /* CType.get_copy */ |
| 1004 |
3 |
| (__params->a) = (a); /* CType.get_copy */ |
| 1005 |
3 |
▶ | PDL_RETERROR(PDL_err, PDL->make_trans_mutual(__privtrans)); |
Branch: 1 branch
| Branch | True | False | Total |
| PDL_RETERROR(PDL_err, PDL->make_trans_mutual(__privtrans)); |
0 |
3 |
3 |
|
| 1006 |
3 |
| return PDL_err; |
| 1007 |
|
| } |