… | |
… | |
7 | typedef cl_platform_id OpenCL__Platform; |
7 | typedef cl_platform_id OpenCL__Platform; |
8 | typedef cl_device_id OpenCL__Device; |
8 | typedef cl_device_id OpenCL__Device; |
9 | typedef cl_context OpenCL__Context; |
9 | typedef cl_context OpenCL__Context; |
10 | typedef cl_command_queue OpenCL__Queue; |
10 | typedef cl_command_queue OpenCL__Queue; |
11 | typedef cl_mem OpenCL__Memory; |
11 | typedef cl_mem OpenCL__Memory; |
|
|
12 | typedef cl_mem OpenCL__Buffer; |
|
|
13 | typedef cl_mem OpenCL__Image; |
|
|
14 | typedef cl_mem OpenCL__Image2D; |
|
|
15 | typedef cl_mem OpenCL__Image3D; |
|
|
16 | typedef cl_mem OpenCL__Memory_ornull; |
|
|
17 | typedef cl_mem OpenCL__Buffer_ornull; |
|
|
18 | typedef cl_mem OpenCL__Image_ornull; |
|
|
19 | typedef cl_mem OpenCL__Image2D_ornull; |
|
|
20 | typedef cl_mem OpenCL__Image3D_ornull; |
12 | typedef cl_sampler OpenCL__Sampler; |
21 | typedef cl_sampler OpenCL__Sampler; |
13 | typedef cl_program OpenCL__Program; |
22 | typedef cl_program OpenCL__Program; |
14 | typedef cl_kernel OpenCL__Kernel; |
23 | typedef cl_kernel OpenCL__Kernel; |
15 | typedef cl_event OpenCL__Event; |
24 | typedef cl_event OpenCL__Event; |
|
|
25 | typedef cl_event OpenCL__UserEvent; |
16 | |
26 | |
17 | static const struct { |
27 | typedef SV *FUTURE; |
|
|
28 | |
|
|
29 | /*****************************************************************************/ |
|
|
30 | |
|
|
31 | /* up to two temporary buffers */ |
|
|
32 | static void * |
|
|
33 | tmpbuf (size_t size) |
|
|
34 | { |
|
|
35 | static int idx; |
|
|
36 | static void *buf [2]; |
|
|
37 | static size_t len [2]; |
|
|
38 | |
|
|
39 | idx ^= 1; |
|
|
40 | |
|
|
41 | if (len [idx] < size) |
|
|
42 | { |
|
|
43 | free (buf [idx]); |
|
|
44 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
|
|
45 | buf [idx] = malloc (len [idx]); |
|
|
46 | } |
|
|
47 | |
|
|
48 | return buf [idx]; |
|
|
49 | } |
|
|
50 | |
|
|
51 | /*****************************************************************************/ |
|
|
52 | |
|
|
53 | typedef struct |
|
|
54 | { |
18 | IV iv; |
55 | IV iv; |
19 | const char *name; |
56 | const char *name; |
20 | } cl_error[] = { |
|
|
21 | #define def_error(name) { (IV)CL_ ## name, # name }, |
57 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
22 | #include "invalid.h" |
58 | } ivstr; |
23 | }; |
|
|
24 | |
59 | |
25 | static const char * |
60 | static const char * |
26 | clstrerror (cl_int res) |
61 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
27 | { |
62 | { |
28 | int i; |
63 | int i; |
29 | static char numbuf [32]; |
64 | static char strbuf [32]; |
30 | |
65 | |
31 | for (i = sizeof (cl_error) / sizeof (cl_error [0]); i--; ) |
66 | for (i = count; i--; ) |
32 | if (cl_error [i].iv == res) |
67 | if (base [i].iv == value) |
33 | return cl_error [i].name; |
68 | return base [i].name; |
34 | |
69 | |
35 | snprintf (numbuf, sizeof (numbuf), "ERROR(%d)", res); |
70 | snprintf (strbuf, sizeof (strbuf), fallback, (int)value); |
36 | |
71 | |
37 | return numbuf; |
72 | return strbuf; |
38 | } |
73 | } |
39 | |
74 | |
|
|
75 | static const char * |
|
|
76 | enum2str (cl_uint value) |
|
|
77 | { |
|
|
78 | static const ivstr enumstr[] = { |
|
|
79 | #include "enumstr.h" |
|
|
80 | }; |
|
|
81 | |
|
|
82 | return iv2str (value, enumstr, sizeof (enumstr) / sizeof (enumstr [0]), "ENUM(0x%04x)"); |
|
|
83 | } |
|
|
84 | |
|
|
85 | static const char * |
|
|
86 | err2str (cl_int err) |
|
|
87 | { |
|
|
88 | static const ivstr errstr[] = { |
|
|
89 | #include "errstr.h" |
|
|
90 | }; |
|
|
91 | |
|
|
92 | return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); |
|
|
93 | } |
|
|
94 | |
|
|
95 | /*****************************************************************************/ |
|
|
96 | |
|
|
97 | static cl_int last_error; |
|
|
98 | |
40 | #define FAIL(name,res) \ |
99 | #define FAIL(name,err) \ |
41 | croak ("cl" # name ": %s", clstrerror (res)); |
100 | croak ("cl" # name ": %s", err2str (last_error = err)); |
42 | |
101 | |
43 | #define NEED_SUCCESS(name,args) \ |
102 | #define NEED_SUCCESS(name,args) \ |
44 | do { \ |
103 | do { \ |
45 | cl_int res = cl ## name args; \ |
104 | cl_int res = cl ## name args; \ |
46 | \ |
105 | \ |
47 | if (res) \ |
106 | if (res) \ |
48 | FAIL (name, res); \ |
107 | FAIL (name, res); \ |
49 | } while (0) |
108 | } while (0) |
50 | |
109 | |
|
|
110 | /*****************************************************************************/ |
|
|
111 | |
51 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
112 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
52 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
113 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
53 | |
114 | |
54 | /*TODO*/ |
115 | static void * |
55 | #define EVENT_LIST(items,count) cl_uint event_list_count = 0; cl_event *event_list_ptr = 0 |
116 | SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
|
|
117 | { |
|
|
118 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
119 | return (void *)SvIV (SvRV (sv)); |
|
|
120 | |
|
|
121 | croak ("%s: %s is not of type %s", func, svname, pkg); |
|
|
122 | } |
|
|
123 | |
|
|
124 | /*****************************************************************************/ |
|
|
125 | |
|
|
126 | static cl_event * |
|
|
127 | event_list (SV **items, int count) |
|
|
128 | { |
|
|
129 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
|
|
130 | |
|
|
131 | while (count--) |
|
|
132 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
|
|
133 | |
|
|
134 | return list; |
|
|
135 | } |
|
|
136 | |
|
|
137 | #define EVENT_LIST(items,count) \ |
|
|
138 | cl_uint event_list_count = (count); \ |
|
|
139 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
56 | |
140 | |
57 | #define INFO(class) \ |
141 | #define INFO(class) \ |
58 | { \ |
142 | { \ |
59 | size_t size; \ |
143 | size_t size; \ |
60 | SV *sv; \ |
144 | SV *sv; \ |
… | |
… | |
73 | PROTOTYPES: ENABLE |
157 | PROTOTYPES: ENABLE |
74 | |
158 | |
75 | BOOT: |
159 | BOOT: |
76 | { |
160 | { |
77 | HV *stash = gv_stashpv ("OpenCL", 1); |
161 | HV *stash = gv_stashpv ("OpenCL", 1); |
78 | static const struct { |
|
|
79 | const char *name; |
|
|
80 | IV iv; |
|
|
81 | } *civ, const_iv[] = { |
162 | static const ivstr *civ, const_iv[] = { |
82 | #define const_iv(name) { # name, (IV)CL_ ## name }, |
163 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
|
|
164 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
|
|
165 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
|
|
166 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
|
|
167 | { sizeof (cl_int ), "SIZEOF_INT" }, |
|
|
168 | { sizeof (cl_uint ), "SIZEOF_UINT" }, |
|
|
169 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
|
|
170 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
|
|
171 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
|
|
172 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
|
|
173 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
83 | #include "constiv.h" |
174 | #include "constiv.h" |
84 | }; |
175 | }; |
85 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
176 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
86 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
177 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
87 | } |
178 | } |
88 | |
179 | |
|
|
180 | cl_int |
|
|
181 | errno () |
|
|
182 | CODE: |
|
|
183 | errno = last_error; |
|
|
184 | |
|
|
185 | const char * |
|
|
186 | err2str (cl_int err) |
|
|
187 | |
|
|
188 | const char * |
|
|
189 | enum2str (cl_uint value) |
|
|
190 | |
89 | void |
191 | void |
90 | platforms () |
192 | platforms () |
91 | PPCODE: |
193 | PPCODE: |
92 | { |
194 | { |
93 | cl_platform_id *list; |
195 | cl_platform_id *list; |
94 | cl_uint count; |
196 | cl_uint count; |
95 | int i; |
197 | int i; |
96 | |
198 | |
97 | NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
199 | NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
98 | Newx (list, count, cl_platform_id); |
200 | list = tmpbuf (sizeof (*list) * count); |
99 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
201 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
100 | |
202 | |
101 | EXTEND (SP, count); |
203 | EXTEND (SP, count); |
102 | for (i = 0; i < count; ++i) |
204 | for (i = 0; i < count; ++i) |
103 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
205 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
104 | |
|
|
105 | Safefree (list); |
|
|
106 | } |
206 | } |
107 | |
207 | |
108 | void |
208 | void |
109 | context_from_type_simple (cl_device_type type = CL_DEVICE_TYPE_DEFAULT) |
209 | context_from_type (FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
110 | PPCODE: |
210 | PPCODE: |
111 | { |
211 | { |
112 | cl_int res; |
212 | cl_int res; |
113 | cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); |
213 | cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); |
114 | |
214 | |
… | |
… | |
142 | cl_device_id *list; |
242 | cl_device_id *list; |
143 | cl_uint count; |
243 | cl_uint count; |
144 | int i; |
244 | int i; |
145 | |
245 | |
146 | NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); |
246 | NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); |
147 | Newx (list, count, cl_device_id); |
247 | list = tmpbuf (sizeof (*list) * count); |
148 | NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); |
248 | NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); |
149 | |
249 | |
150 | EXTEND (SP, count); |
250 | EXTEND (SP, count); |
151 | for (i = 0; i < count; ++i) |
251 | for (i = 0; i < count; ++i) |
152 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
252 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
153 | |
|
|
154 | Safefree (list); |
|
|
155 | } |
253 | } |
156 | |
254 | |
157 | void |
255 | void |
158 | context_from_type_simple (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_DEFAULT) |
256 | context_from_type (OpenCL::Platform this, FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
159 | PPCODE: |
257 | PPCODE: |
160 | { |
258 | { |
161 | cl_int res; |
259 | cl_int res; |
162 | cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; |
260 | cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; |
163 | cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); |
261 | cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); |
… | |
… | |
166 | FAIL (CreateContextFromType, res); |
264 | FAIL (CreateContextFromType, res); |
167 | |
265 | |
168 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
266 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
169 | } |
267 | } |
170 | |
268 | |
171 | void |
|
|
172 | unload_compiler () |
|
|
173 | CODE: |
|
|
174 | NEED_SUCCESS (UnloadCompiler, ()); |
|
|
175 | |
|
|
176 | MODULE = OpenCL PACKAGE = OpenCL::Device |
269 | MODULE = OpenCL PACKAGE = OpenCL::Device |
177 | |
270 | |
178 | void |
271 | void |
179 | info (OpenCL::Device this, cl_device_info name) |
272 | info (OpenCL::Device this, cl_device_info name) |
180 | PPCODE: |
273 | PPCODE: |
181 | INFO (Device) |
274 | INFO (Device) |
182 | |
275 | |
183 | void |
276 | void |
184 | context_simple (OpenCL::Device this) |
277 | context (OpenCL::Device this, FUTURE properties = 0, FUTURE notify = 0) |
185 | PPCODE: |
278 | PPCODE: |
186 | { |
279 | { |
187 | cl_int res; |
280 | cl_int res; |
188 | cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); |
281 | cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); |
189 | |
282 | |
… | |
… | |
204 | info (OpenCL::Context this, cl_context_info name) |
297 | info (OpenCL::Context this, cl_context_info name) |
205 | PPCODE: |
298 | PPCODE: |
206 | INFO (Context) |
299 | INFO (Context) |
207 | |
300 | |
208 | void |
301 | void |
209 | command_queue_simple (OpenCL::Context this, OpenCL::Device device) |
302 | queue (OpenCL::Context this, OpenCL::Device device, cl_command_queue_properties properties = 0) |
210 | PPCODE: |
303 | PPCODE: |
211 | { |
304 | { |
212 | cl_int res; |
305 | cl_int res; |
213 | cl_command_queue queue = clCreateCommandQueue (this, device, 0, &res); |
306 | cl_command_queue queue = clCreateCommandQueue (this, device, properties, &res); |
214 | |
307 | |
215 | if (res) |
308 | if (res) |
216 | FAIL (CreateCommandQueue, res); |
309 | FAIL (CreateCommandQueue, res); |
217 | |
310 | |
218 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
311 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
219 | } |
312 | } |
220 | |
313 | |
221 | void |
314 | void |
|
|
315 | user_event (OpenCL::Context this) |
|
|
316 | PPCODE: |
|
|
317 | { |
|
|
318 | cl_int res; |
|
|
319 | cl_event ev = clCreateUserEvent (this, &res); |
|
|
320 | |
|
|
321 | if (res) |
|
|
322 | FAIL (CreateUserevent, res); |
|
|
323 | |
|
|
324 | XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
|
|
325 | } |
|
|
326 | |
|
|
327 | void |
222 | buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) |
328 | buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) |
223 | PPCODE: |
329 | PPCODE: |
224 | { |
330 | { |
225 | cl_int res; |
331 | cl_int res; |
|
|
332 | cl_mem mem; |
|
|
333 | |
|
|
334 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
|
|
335 | croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
|
|
336 | |
226 | cl_mem mem = clCreateBuffer (this, flags, len, 0, &res); |
337 | mem = clCreateBuffer (this, flags, len, 0, &res); |
227 | |
338 | |
228 | if (res) |
339 | if (res) |
229 | FAIL (CreateBuffer, res); |
340 | FAIL (CreateBuffer, res); |
230 | |
341 | |
231 | XPUSH_NEW_OBJ ("OpenCL::Memory", mem); |
342 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
232 | } |
343 | } |
233 | |
344 | |
234 | void |
345 | void |
235 | buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) |
346 | buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) |
236 | PPCODE: |
347 | PPCODE: |
237 | { |
348 | { |
238 | STRLEN len; |
349 | STRLEN len; |
239 | char *ptr = SvPVbyte (data, len); |
350 | char *ptr = SvPVbyte (data, len); |
240 | cl_int res; |
351 | cl_int res; |
|
|
352 | cl_mem mem; |
|
|
353 | |
|
|
354 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
|
|
355 | croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
|
|
356 | |
241 | cl_mem mem = clCreateBuffer (this, flags, len, ptr, &res); |
357 | mem = clCreateBuffer (this, flags, len, ptr, &res); |
242 | |
358 | |
243 | if (res) |
359 | if (res) |
244 | FAIL (CreateBuffer, res); |
360 | FAIL (CreateBuffer, res); |
245 | |
361 | |
246 | XPUSH_NEW_OBJ ("OpenCL::Memory", mem); |
362 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
|
|
363 | } |
|
|
364 | |
|
|
365 | void |
|
|
366 | image2d (OpenCL::Context this, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, SV *data) |
|
|
367 | PPCODE: |
|
|
368 | { |
|
|
369 | STRLEN len; |
|
|
370 | char *ptr = SvPVbyte (data, len); |
|
|
371 | const cl_image_format format = { channel_order, channel_type }; |
|
|
372 | cl_int res; |
|
|
373 | cl_mem mem = clCreateImage2D (this, flags, &format, width, height, len / height, ptr, &res); |
|
|
374 | |
|
|
375 | if (res) |
|
|
376 | FAIL (CreateImage2D, res); |
|
|
377 | |
|
|
378 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
379 | } |
|
|
380 | |
|
|
381 | void |
|
|
382 | image3d (OpenCL::Context this, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t depth, size_t slice_pitch, SV *data) |
|
|
383 | PPCODE: |
|
|
384 | { |
|
|
385 | STRLEN len; |
|
|
386 | char *ptr = SvPVbyte (data, len); |
|
|
387 | const cl_image_format format = { channel_order, channel_type }; |
|
|
388 | cl_int res; |
|
|
389 | cl_mem mem = clCreateImage3D (this, flags, &format, width, height, |
|
|
390 | depth, len / (height * slice_pitch), slice_pitch, ptr, &res); |
|
|
391 | |
|
|
392 | if (res) |
|
|
393 | FAIL (CreateImage3D, res); |
|
|
394 | |
|
|
395 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
|
|
396 | } |
|
|
397 | |
|
|
398 | void |
|
|
399 | supported_image_formats (OpenCL::Context this, cl_mem_flags flags, cl_mem_object_type image_type) |
|
|
400 | PPCODE: |
|
|
401 | { |
|
|
402 | cl_uint count; |
|
|
403 | cl_image_format *list; |
|
|
404 | int i; |
|
|
405 | |
|
|
406 | NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, 0, 0, &count)); |
|
|
407 | Newx (list, count, cl_image_format); |
|
|
408 | NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, count, list, 0)); |
|
|
409 | |
|
|
410 | EXTEND (SP, count); |
|
|
411 | for (i = 0; i < count; ++i) |
|
|
412 | { |
|
|
413 | AV *av = newAV (); |
|
|
414 | av_store (av, 1, newSVuv (list [i].image_channel_data_type)); |
|
|
415 | av_store (av, 0, newSVuv (list [i].image_channel_order)); |
|
|
416 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
417 | } |
247 | } |
418 | } |
248 | |
419 | |
249 | void |
420 | void |
250 | sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
421 | sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
251 | PPCODE: |
422 | PPCODE: |
… | |
… | |
289 | info (OpenCL::Queue this, cl_command_queue_info name) |
460 | info (OpenCL::Queue this, cl_command_queue_info name) |
290 | PPCODE: |
461 | PPCODE: |
291 | INFO (CommandQueue) |
462 | INFO (CommandQueue) |
292 | |
463 | |
293 | void |
464 | void |
294 | enqueue_read_buffer (OpenCL::Queue this, OpenCL::Memory mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
465 | enqueue_read_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
295 | PPCODE: |
466 | PPCODE: |
296 | { |
467 | { |
297 | cl_event ev = 0; |
468 | cl_event ev = 0; |
298 | EVENT_LIST (6, items - 6); |
469 | EVENT_LIST (6, items - 6); |
299 | |
470 | |
… | |
… | |
306 | if (ev) |
477 | if (ev) |
307 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
478 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
308 | } |
479 | } |
309 | |
480 | |
310 | void |
481 | void |
311 | enqueue_write_buffer (OpenCL::Queue this, OpenCL::Memory mem, cl_bool blocking, size_t offset, SV *data, ...) |
482 | enqueue_write_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
312 | PPCODE: |
483 | PPCODE: |
313 | { |
484 | { |
314 | cl_event ev = 0; |
485 | cl_event ev = 0; |
315 | STRLEN len; |
486 | STRLEN len; |
316 | char *ptr = SvPVbyte (data, len); |
487 | char *ptr = SvPVbyte (data, len); |
… | |
… | |
321 | if (ev) |
492 | if (ev) |
322 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
493 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
323 | } |
494 | } |
324 | |
495 | |
325 | void |
496 | void |
326 | enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Memory src, OpenCL::Memory dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
497 | enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
327 | PPCODE: |
498 | PPCODE: |
328 | { |
499 | { |
329 | cl_event ev = 0; |
500 | cl_event ev = 0; |
330 | EVENT_LIST (6, items - 6); |
501 | EVENT_LIST (6, items - 6); |
331 | |
502 | |
332 | NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
503 | NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
504 | |
|
|
505 | if (ev) |
|
|
506 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
507 | } |
|
|
508 | |
|
|
509 | /*TODO http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clEnqueueReadBufferRect.html */ |
|
|
510 | /*TODO http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clEnqueueWriteBufferRect.html */ |
|
|
511 | |
|
|
512 | void |
|
|
513 | enqueue_read_image (OpenCL::Queue this, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
|
|
514 | PPCODE: |
|
|
515 | { |
|
|
516 | cl_event ev = 0; |
|
|
517 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
518 | const size_t region[3] = { width, height, depth }; |
|
|
519 | size_t len = row_pitch * slice_pitch * depth; |
|
|
520 | EVENT_LIST (11, items - 11); |
|
|
521 | |
|
|
522 | SvUPGRADE (data, SVt_PV); |
|
|
523 | SvGROW (data, len); |
|
|
524 | SvPOK_only (data); |
|
|
525 | SvCUR_set (data, len); |
|
|
526 | NEED_SUCCESS (EnqueueReadImage, (this, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
527 | |
|
|
528 | if (ev) |
|
|
529 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
530 | } |
|
|
531 | |
|
|
532 | void |
|
|
533 | enqueue_write_image (OpenCL::Queue this, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, SV *data, ...) |
|
|
534 | PPCODE: |
|
|
535 | { |
|
|
536 | cl_event ev = 0; |
|
|
537 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
538 | const size_t region[3] = { width, height, depth }; |
|
|
539 | STRLEN len; |
|
|
540 | char *ptr = SvPVbyte (data, len); |
|
|
541 | size_t slice_pitch = len / (row_pitch * height); |
|
|
542 | EVENT_LIST (11, items - 11); |
|
|
543 | |
|
|
544 | NEED_SUCCESS (EnqueueWriteImage, (this, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
545 | |
|
|
546 | if (ev) |
|
|
547 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
548 | } |
|
|
549 | |
|
|
550 | void |
|
|
551 | enqueue_copy_buffer_rect (OpenCL::Queue this, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) |
|
|
552 | PPCODE: |
|
|
553 | { |
|
|
554 | cl_event ev = 0; |
|
|
555 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
556 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
557 | const size_t region[3] = { width, height, depth }; |
|
|
558 | EVENT_LIST (16, items - 16); |
|
|
559 | |
|
|
560 | NEED_SUCCESS (EnqueueCopyBufferRect, (this, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
561 | |
|
|
562 | if (ev) |
|
|
563 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
564 | } |
|
|
565 | |
|
|
566 | void |
|
|
567 | enqueue_copy_buffer_to_image (OpenCL::Queue this, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
568 | PPCODE: |
|
|
569 | { |
|
|
570 | cl_event ev = 0; |
|
|
571 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
572 | const size_t region[3] = { width, height, depth }; |
|
|
573 | EVENT_LIST (10, items - 10); |
|
|
574 | |
|
|
575 | NEED_SUCCESS (EnqueueCopyBufferToImage, (this, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
576 | |
|
|
577 | if (ev) |
|
|
578 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
579 | } |
|
|
580 | |
|
|
581 | void |
|
|
582 | enqueue_copy_image (OpenCL::Queue this, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
583 | PPCODE: |
|
|
584 | { |
|
|
585 | cl_event ev = 0; |
|
|
586 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
587 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
588 | const size_t region[3] = { width, height, depth }; |
|
|
589 | EVENT_LIST (12, items - 12); |
|
|
590 | |
|
|
591 | NEED_SUCCESS (EnqueueCopyImage, (this, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
592 | |
|
|
593 | if (ev) |
|
|
594 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
595 | } |
|
|
596 | |
|
|
597 | void |
|
|
598 | enqueue_copy_image_to_buffer (OpenCL::Queue this, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) |
|
|
599 | PPCODE: |
|
|
600 | { |
|
|
601 | cl_event ev = 0; |
|
|
602 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
603 | const size_t region[3] = { width, height, depth }; |
|
|
604 | EVENT_LIST (10, items - 10); |
|
|
605 | |
|
|
606 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (this, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
607 | |
|
|
608 | if (ev) |
|
|
609 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
610 | } |
|
|
611 | |
|
|
612 | void |
|
|
613 | enqueue_task (OpenCL::Queue this, OpenCL::Kernel kernel, ...) |
|
|
614 | PPCODE: |
|
|
615 | { |
|
|
616 | cl_event ev = 0; |
|
|
617 | EVENT_LIST (2, items - 2); |
|
|
618 | |
|
|
619 | NEED_SUCCESS (EnqueueTask, (this, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
620 | |
|
|
621 | if (ev) |
|
|
622 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
623 | } |
|
|
624 | |
|
|
625 | void |
|
|
626 | enqueue_nd_range_kernel (OpenCL::Queue this, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
627 | PPCODE: |
|
|
628 | { |
|
|
629 | cl_event ev = 0; |
|
|
630 | size_t *gwo = 0, *gws, *lws = 0; |
|
|
631 | int gws_len; |
|
|
632 | size_t *lists; |
|
|
633 | int i; |
|
|
634 | EVENT_LIST (5, items - 5); |
|
|
635 | |
|
|
636 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
|
|
637 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
|
|
638 | |
|
|
639 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
|
|
640 | |
|
|
641 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
|
|
642 | |
|
|
643 | gws = lists + gws_len * 0; |
|
|
644 | for (i = 0; i < gws_len; ++i) |
|
|
645 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
646 | |
|
|
647 | if (SvOK (global_work_offset)) |
|
|
648 | { |
|
|
649 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
|
|
650 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
|
|
651 | |
|
|
652 | if (AvFILLp (SvRV (global_work_size)) + 1 != gws_len) |
|
|
653 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array of same size as global_work_size"); |
|
|
654 | |
|
|
655 | gwo = lists + gws_len * 1; |
|
|
656 | for (i = 0; i < gws_len; ++i) |
|
|
657 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
|
|
658 | } |
|
|
659 | |
|
|
660 | if (SvOK (local_work_size)) |
|
|
661 | { |
|
|
662 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
|
|
663 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
|
|
664 | |
|
|
665 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
|
|
666 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
|
|
667 | |
|
|
668 | lws = lists + gws_len * 2; |
|
|
669 | for (i = 0; i < gws_len; ++i) |
|
|
670 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
671 | } |
|
|
672 | |
|
|
673 | NEED_SUCCESS (EnqueueNDRangeKernel, (this, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
333 | |
674 | |
334 | if (ev) |
675 | if (ev) |
335 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
676 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
336 | } |
677 | } |
337 | |
678 | |
… | |
… | |
354 | |
695 | |
355 | void |
696 | void |
356 | enqueue_barrier (OpenCL::Queue this) |
697 | enqueue_barrier (OpenCL::Queue this) |
357 | CODE: |
698 | CODE: |
358 | NEED_SUCCESS (EnqueueBarrier, (this)); |
699 | NEED_SUCCESS (EnqueueBarrier, (this)); |
|
|
700 | |
|
|
701 | void |
|
|
702 | flush (OpenCL::Queue this) |
|
|
703 | CODE: |
|
|
704 | NEED_SUCCESS (Flush, (this)); |
|
|
705 | |
|
|
706 | void |
|
|
707 | finish (OpenCL::Queue this) |
|
|
708 | CODE: |
|
|
709 | NEED_SUCCESS (Finish, (this)); |
359 | |
710 | |
360 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
711 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
361 | |
712 | |
362 | void |
713 | void |
363 | DESTROY (OpenCL::Memory this) |
714 | DESTROY (OpenCL::Memory this) |
… | |
… | |
438 | info (OpenCL::Kernel this, cl_kernel_info name) |
789 | info (OpenCL::Kernel this, cl_kernel_info name) |
439 | PPCODE: |
790 | PPCODE: |
440 | INFO (Kernel) |
791 | INFO (Kernel) |
441 | |
792 | |
442 | void |
793 | void |
443 | set_bool (OpenCL::Kernel this, cl_uint idx, cl_bool value) |
794 | set_char (OpenCL::Kernel this, cl_uint idx, cl_char value) |
444 | CODE: |
795 | CODE: |
445 | clKernelSetArg (this, idx, sizeof (value), &value); |
796 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
797 | |
|
|
798 | void |
|
|
799 | set_uchar (OpenCL::Kernel this, cl_uint idx, cl_uchar value) |
|
|
800 | CODE: |
|
|
801 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
802 | |
|
|
803 | void |
|
|
804 | set_short (OpenCL::Kernel this, cl_uint idx, cl_short value) |
|
|
805 | CODE: |
|
|
806 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
807 | |
|
|
808 | void |
|
|
809 | set_ushort (OpenCL::Kernel this, cl_uint idx, cl_ushort value) |
|
|
810 | CODE: |
|
|
811 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
812 | |
|
|
813 | void |
|
|
814 | set_int (OpenCL::Kernel this, cl_uint idx, cl_int value) |
|
|
815 | CODE: |
|
|
816 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
817 | |
|
|
818 | void |
|
|
819 | set_uint (OpenCL::Kernel this, cl_uint idx, cl_uint value) |
|
|
820 | CODE: |
|
|
821 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
822 | |
|
|
823 | void |
|
|
824 | set_long (OpenCL::Kernel this, cl_uint idx, cl_long value) |
|
|
825 | CODE: |
|
|
826 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
827 | |
|
|
828 | void |
|
|
829 | set_ulong (OpenCL::Kernel this, cl_uint idx, cl_ulong value) |
|
|
830 | CODE: |
|
|
831 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
832 | |
|
|
833 | void |
|
|
834 | set_half (OpenCL::Kernel this, cl_uint idx, cl_half value) |
|
|
835 | CODE: |
|
|
836 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
837 | |
|
|
838 | void |
|
|
839 | set_float (OpenCL::Kernel this, cl_uint idx, cl_float value) |
|
|
840 | CODE: |
|
|
841 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
842 | |
|
|
843 | void |
|
|
844 | set_double (OpenCL::Kernel this, cl_uint idx, cl_double value) |
|
|
845 | CODE: |
|
|
846 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
847 | |
|
|
848 | void |
|
|
849 | set_memory (OpenCL::Kernel this, cl_uint idx, OpenCL::Memory_ornull value) |
|
|
850 | CODE: |
|
|
851 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
852 | |
|
|
853 | void |
|
|
854 | set_buffer (OpenCL::Kernel this, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
855 | CODE: |
|
|
856 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
857 | |
|
|
858 | void |
|
|
859 | set_image2d (OpenCL::Kernel this, cl_uint idx, OpenCL::Image2D_ornull value) |
|
|
860 | CODE: |
|
|
861 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
862 | |
|
|
863 | void |
|
|
864 | set_image3d (OpenCL::Kernel this, cl_uint idx, OpenCL::Image3D_ornull value) |
|
|
865 | CODE: |
|
|
866 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
867 | |
|
|
868 | void |
|
|
869 | set_sampler (OpenCL::Kernel this, cl_uint idx, OpenCL::Sampler value) |
|
|
870 | CODE: |
|
|
871 | clSetKernelArg (this, idx, sizeof (value), &value); |
|
|
872 | |
|
|
873 | void |
|
|
874 | set_event (OpenCL::Kernel this, cl_uint idx, OpenCL::Event value) |
|
|
875 | CODE: |
|
|
876 | clSetKernelArg (this, idx, sizeof (value), &value); |
446 | |
877 | |
447 | MODULE = OpenCL PACKAGE = OpenCL::Event |
878 | MODULE = OpenCL PACKAGE = OpenCL::Event |
448 | |
879 | |
449 | void |
880 | void |
450 | DESTROY (OpenCL::Event this) |
881 | DESTROY (OpenCL::Event this) |
… | |
… | |
459 | void |
890 | void |
460 | wait (OpenCL::Event this) |
891 | wait (OpenCL::Event this) |
461 | CODE: |
892 | CODE: |
462 | clWaitForEvents (1, &this); |
893 | clWaitForEvents (1, &this); |
463 | |
894 | |
|
|
895 | MODULE = OpenCL PACKAGE = OpenCL::UserEvent |
|
|
896 | |
|
|
897 | void |
|
|
898 | set_status (OpenCL::UserEvent this, cl_int execution_status) |
|
|
899 | CODE: |
|
|
900 | clSetUserEventStatus (this, execution_status); |
|
|
901 | |