xref: /freebsd/contrib/llvm-project/clang/lib/Headers/__clang_cuda_texture_intrinsics.h (revision bdd1243df58e60e85101c09001d9812a789b6bc4)
1 /*===--- __clang_cuda_texture_intrinsics.h - Device-side texture support ---===
2  *
3  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4  * See https://llvm.org/LICENSE.txt for license information.
5  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6  *
7  *===-----------------------------------------------------------------------===
8  *
9  * This header provides in-header implmentations for NVCC's built-in
10  * __nv_tex_surf_handler() which is used by CUDA's texture-related headers.  The
11  * built-in is unusual as it's actually a set of function overloads that use the
12  * first string literal argument as one of the overload parameters.
13  */
14 #ifndef __CLANG_CUDA_TEXTURE_INTRINSICS_H__
15 #define __CLANG_CUDA_TEXTURE_INTRINSICS_H__
16 #ifndef __CUDA__
17 #error "This file is for CUDA compilation only."
18 #endif
19 
20 // __nv_tex_surf_handler() provided by this header as a macro.
21 #define __nv_tex_surf_handler(__op, __ptr, ...)                                \
22   ::__cuda_tex::__tex_fetch<                                                   \
23       ::__cuda_tex::__Tag<::__cuda_tex::__tex_op_hash(__op)>>(__ptr,           \
24                                                               __VA_ARGS__)
25 
26 #pragma push_macro("__ASM_OUT")
27 #pragma push_macro("__ASM_OUTP")
28 #pragma push_macro("__Args")
29 #pragma push_macro("__ID")
30 #pragma push_macro("__IDV")
31 #pragma push_macro("__IMPL_2DGATHER")
32 #pragma push_macro("__IMPL_ALIAS")
33 #pragma push_macro("__IMPL_ALIASI")
34 #pragma push_macro("__IMPL_F1")
35 #pragma push_macro("__IMPL_F3")
36 #pragma push_macro("__IMPL_F3N")
37 #pragma push_macro("__IMPL_F3S")
38 #pragma push_macro("__IMPL_S")
39 #pragma push_macro("__IMPL_S3")
40 #pragma push_macro("__IMPL_S3I")
41 #pragma push_macro("__IMPL_S3N")
42 #pragma push_macro("__IMPL_S3NI")
43 #pragma push_macro("__IMPL_S3S")
44 #pragma push_macro("__IMPL_S3SI")
45 #pragma push_macro("__IMPL_SI")
46 #pragma push_macro("__L")
47 #pragma push_macro("__STRIP_PARENS")
48 
49 // Put all functions into anonymous namespace so they have internal linkage.
50 // The device-only function here must be internal in order to avoid ODR
51 // violations in case they are used from the files compiled with
52 // -fgpu-rdc. E.g. a library and an app using it may be built with a different
53 // version of this header file.
54 namespace {
55 
56 // Put the implmentation into its own namespace so we don't pollute the TU.
57 namespace __cuda_tex {
58 
59 // First, we need a perfect hash function and a few constexpr helper functions
60 // for converting a string literal into a numeric value which can be used to
61 // parametrize a template. We can not use string literals for that as that would
62 // require C++20.
63 //
64 // The hash function was generated with 'gperf' and then manually converted into
65 // its constexpr equivalent.
66 //
67 // NOTE: the perfect hashing scheme comes with inherent self-test. If the hash
68 // function has a collision for any of the texture operations, the compilation
69 // will fail due to an attempt to redefine a tag with the same value. If the
70 // header compiles, then the hash function is good enough for the job.
71 
__tex_len(const char * s)72 constexpr int __tex_len(const char *s) {
73   return (s[0] == 0)    ? 0
74          : (s[1] == 0)  ? 1
75          : (s[2] == 0)  ? 2
76          : (s[3] == 0)  ? 3
77          : (s[4] == 0)  ? 4
78          : (s[5] == 0)  ? 5
79          : (s[6] == 0)  ? 6
80          : (s[7] == 0)  ? 7
81          : (s[8] == 0)  ? 8
82          : (s[9] == 0)  ? 9
83          : (s[10] == 0) ? 10
84          : (s[11] == 0) ? 11
85          : (s[12] == 0) ? 12
86          : (s[13] == 0) ? 13
87          : (s[14] == 0) ? 14
88          : (s[15] == 0) ? 15
89          : (s[16] == 0) ? 16
90          : (s[17] == 0) ? 17
91          : (s[18] == 0) ? 18
92          : (s[19] == 0) ? 19
93          : (s[20] == 0) ? 20
94          : (s[21] == 0) ? 21
95          : (s[22] == 0) ? 22
96          : (s[23] == 0) ? 23
97          : (s[24] == 0) ? 24
98          : (s[25] == 0) ? 25
99          : (s[26] == 0) ? 26
100          : (s[27] == 0) ? 27
101          : (s[28] == 0) ? 28
102          : (s[29] == 0) ? 29
103          : (s[30] == 0) ? 30
104          : (s[31] == 0) ? 31
105                         : 32;
106 }
107 
__tex_hash_map(int c)108 constexpr int __tex_hash_map(int c) {
109   return (c == 49)    ? 10
110          : (c == 50)  ? 0
111          : (c == 51)  ? 100
112          : (c == 52)  ? 30
113          : (c == 67)  ? 10
114          : (c == 68)  ? 0
115          : (c == 69)  ? 25
116          : (c == 72)  ? 70
117          : (c == 77)  ? 0
118          : (c == 96)  ? 44
119          : (c == 99)  ? 10
120          : (c == 100) ? 5
121          : (c == 101) ? 60
122          : (c == 102) ? 40
123          : (c == 103) ? 70
124          : (c == 104) ? 25
125          : (c == 112) ? 0
126          : (c == 114) ? 45
127          : (c == 117) ? 5
128          : (c == 118) ? 85
129          : (c == 120) ? 20
130                       : 225;
131 }
132 
__tex_op_hash(const char * str)133 constexpr int __tex_op_hash(const char *str) {
134   return __tex_len(str) + __tex_hash_map(str[7] + 1) + __tex_hash_map(str[6]) +
135          __tex_hash_map(str[5]) + __tex_hash_map(str[__tex_len(str) - 1]);
136 }
137 
138 // Tag type to identify particular texture operation.
139 template <int N> struct __Tag;
140 #define __ID(__op) __Tag<__tex_op_hash(__op)>
141 // Tags for variants of particular operation. E.g. tex2Dgather can translate
142 // into 4 different instructions.
143 #define __IDV(__op, __variant)                                                 \
144   __Tag<10000 + __tex_op_hash(__op) * 100 + __variant>
145 
146 // Helper classes for figuring out key data types for derived types.
147 // E.g. char2 has __base_t = char, __fetch_t = char4
148 template <class> struct __TypeInfoT;
149 // Type info for the fundamental types.
150 template <> struct __TypeInfoT<float> {
151   using __base_t = float;
152   using __fetch_t = float4;
153 };
154 template <> struct __TypeInfoT<char> {
155   using __base_t = char;
156   using __fetch_t = int4;
157 };
158 template <> struct __TypeInfoT<signed char> {
159   using __base_t = signed char;
160   using __fetch_t = int4;
161 };
162 template <> struct __TypeInfoT<unsigned char> {
163   using __base_t = unsigned char;
164   using __fetch_t = uint4;
165 };
166 template <> struct __TypeInfoT<short> {
167   using __base_t = short;
168   using __fetch_t = int4;
169 };
170 template <> struct __TypeInfoT<unsigned short> {
171   using __base_t = unsigned short;
172   using __fetch_t = uint4;
173 };
174 template <> struct __TypeInfoT<int> {
175   using __base_t = int;
176   using __fetch_t = int4;
177 };
178 template <> struct __TypeInfoT<unsigned int> {
179   using __base_t = unsigned int;
180   using __fetch_t = uint4;
181 };
182 
183 // Derived base/fetch types for N-element vectors.
184 template <class __T> struct __TypeInfoT {
185   using __base_t = decltype(__T::x);
186   using __fetch_t = typename __TypeInfoT<__base_t>::__fetch_t;
187 };
188 
189 // Classes that implement specific texture ops.
190 template <class __op> struct __tex_fetch_v4;
191 
192 // Helper macros to strip parens from a macro argument.
193 #define __Args(...) __VA_ARGS__
194 #define __STRIP_PARENS(__X) __X
195 #define __L(__X) __STRIP_PARENS(__Args __X)
196 
197 // Construct inline assembly output args.
198 // Results are stored in a temp var __r.
199 // isResident bool is pointed to by __ir
200 // Asm args for return values. It's a 4-element vector
201 #define __ASM_OUT(__t)                                                         \
202   ("=" __t(__r.x), "=" __t(__r.y), "=" __t(__r.z), "=" __t(__r.w))
203 // .. possibly combined with a predicate.
204 #define __ASM_OUTP(__t) (__L(__ASM_OUT(__t)), "=h"(*__ir))
205 
206 // Implements a single variant of texture fetch instruction.
207 #define __IMPL_F1(__rt, __dt, __args, __asm_op, __asm_outs, __asm_args)        \
208   template <>                                                                  \
209   __device__ __rt __run<__dt>(cudaTextureObject_t __obj, __L(__args)) {        \
210     __rt __r;                                                                  \
211     asm(__asm_op : __L(__asm_outs) : "l"(__obj), __L(__asm_args));             \
212     return __r;                                                                \
213   }
214 
215 // Implements texture fetch instructions for int4/uint4/float4 data types.
216 #define __IMPL_F3(__args, __asm_op, __ctype, __asm_op_args, __asm_args)        \
217   __IMPL_F1(int4, int4, __args, __asm_op ".s32." __ctype "\t" __asm_op_args,   \
218             __ASM_OUT("r"), __asm_args)                                        \
219   __IMPL_F1(uint4, uint4, __args, __asm_op ".u32." __ctype "\t" __asm_op_args, \
220             __ASM_OUT("r"), __asm_args)                                        \
221   __IMPL_F1(float4, float4, __args,                                            \
222             __asm_op ".f32." __ctype "\t" __asm_op_args, __ASM_OUT("f"),       \
223             __asm_args)
224 // Implements 'sparse' texture fetch instructions for int4/uint4/float4 data
225 // types. Similar to above, but returns a boolean 'isPresent' value in addition
226 // to texture data,
227 #define __IMPL_F3S(__args, __asm_op, __ctype, __asm_op_args, __asm_args)       \
228   __IMPL_F1(int4, int4, __args, __asm_op ".s32." __ctype "\t" __asm_op_args,   \
229             __ASM_OUTP("r"), __asm_args)                                       \
230   __IMPL_F1(uint4, uint4, __args, __asm_op ".u32." __ctype "\t" __asm_op_args, \
231             __ASM_OUTP("r"), __asm_args)                                       \
232   __IMPL_F1(float4, float4, __args,                                            \
233             __asm_op ".f32." __ctype "\t" __asm_op_args, __ASM_OUTP("f"),      \
234             __asm_args)
235 
236 // Similar to F3, but for integer data which is returned as normalized floats.
237 // Only instantiates fetch functions for int4/uint4.
238 #define __IMPL_F3N(__args, __asm_op, __ctype, __asm_op_args, __asm_args)       \
239   __IMPL_F1(float4, int4, __args, __asm_op ".s32." __ctype "\t" __asm_op_args, \
240             __ASM_OUT("r"), __asm_args)                                        \
241   __IMPL_F1(float4, uint4, __args,                                             \
242             __asm_op ".u32." __ctype "\t" __asm_op_args, __ASM_OUT("r"),       \
243             __asm_args)
244 
245 // Instantiates __tex_fetch_v4 with regular fetch functions.
246 #define __IMPL_S3I(__op, __args, __asm_op, __ctype, __asm_op_args, __asm_args) \
247   template <> struct __tex_fetch_v4<__op> {                                    \
248     template <class T>                                                         \
249     __device__ static T __run(cudaTextureObject_t __obj, __L(__args));         \
250     __IMPL_F3(__args, __asm_op, __ctype, __asm_op_args, __asm_args)            \
251   }
252 
253 // Same, but for sparse ops. Only available on sm_60+
254 #if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 600)
255 #define __IMPL_S3SI(__op, __args, __asm_op, __ctype, __asm_op_args,            \
256                     __asm_args)                                                \
257   template <> struct __tex_fetch_v4<__op> {                                    \
258     template <class T>                                                         \
259     __device__ static T __run(cudaTextureObject_t __obj, __L(__args));         \
260     __IMPL_F3S(__args, __asm_op, __ctype, __asm_op_args, __asm_args)           \
261   }
262 #else
263 #define __IMPL_S3SI(__op, __args, __asm_op, __ctype, __asm_op_args, __asm_args)
264 #endif
265 
266 // Same, but for normalized float ops.
267 #define __IMPL_S3NI(__op, __args, __asm_op, __ctype, __asm_op_args,            \
268                     __asm_args)                                                \
269   template <> struct __tex_fetch_v4<__op> {                                    \
270     template <class T>                                                         \
271     __device__ static float4 __run(cudaTextureObject_t __obj, __L(__args));    \
272     __IMPL_F3N(__args, __asm_op, __ctype, __asm_op_args, __asm_args)           \
273   }
274 
275 // Regular and normalized float ops share a lot of similarities.  This macro
276 // instantiates both variants -- normal for __op and normalized for __opn.
277 #define __IMPL_SI(__op, __opn, __args, __asm_op, __ctype, __asm_op_args,       \
278                   __asm_args)                                                  \
279   __IMPL_S3I(__op, __args, __asm_op, __ctype, __asm_op_args, __asm_args);      \
280   __IMPL_S3NI(__opn, __args, __asm_op, __ctype, __asm_op_args, __asm_args)
281 
282 // Convenience macros which converts string literal __op into a __Tag,
283 #define __IMPL_S3(__op, __args, __asm_op, __ctype, __asm_op_args, __asm_args)  \
284   __IMPL_S3I(__ID(__op), __args, __asm_op, __ctype, __asm_op_args, __asm_args)
285 #define __IMPL_S3S(__op, __args, __asm_op, __ctype, __asm_op_args, __asm_args) \
286   __IMPL_S3SI(__ID(__op), __args, __asm_op, __ctype, __asm_op_args, __asm_args)
287 #define __IMPL_S3N(__op, __args, __asm_op, __ctype, __asm_op_args, __asm_args) \
288   __IMPL_S3NI(__ID(__op), __args, __asm_op, __ctype, __asm_op_args, __asm_args)
289 #define __IMPL_S(__op, __opn, __args, __asm_op, __ctype, __asm_op_args,        \
290                  __asm_args)                                                   \
291   __IMPL_SI(__ID(__op), __ID(__opn), __args, __asm_op, __ctype, __asm_op_args, \
292             __asm_args)
293 
294 // CUDA headers have some 'legacy' texture oprerations that duplicate
295 // functionality. So, we just inherit it, instead of refining a copy.
296 #define __IMPL_ALIASI(__op, __opn)                                             \
297   template <> struct __tex_fetch_v4<__op> : __tex_fetch_v4<__opn> {}
298 #define __IMPL_ALIAS(__op, __opn) __IMPL_ALIASI(__ID(__op), __ID(__opn))
299 
300 // Now we can instantiate everything we need for each specific texture fetch
301 // variant.
302 __IMPL_S("__tex1D_v2", "__tex1D_rmnf_v2", (float __x), "tex.1d.v4", "f32",
303          "{%0, %1, %2, %3}, [%4, {%5}];", ("f"(__x)));
304 __IMPL_S("__tex1Dfetch_v2", "__tex1Dfetch_rmnf_v2", (int __x), "tex.1d.v4",
305          "s32", "{%0, %1, %2, %3}, [%4, {%5}];", ("r"(__x)));
306 __IMPL_ALIAS("__itex1D", "__tex1D_v2");
307 __IMPL_ALIAS("__itex1Dfetch", "__tex1Dfetch_v2");
308 
309 __IMPL_S("__tex1DGrad_v2", "__tex1DGrad_rmnf_v2",
310          (float __x, float __dPdx, float __dPdy), "tex.grad.1d.v4", "f32",
311          "{%0, %1, %2, %3}, [%4, {%5}], {%6}, {%7};",
312          ("f"(__x), "f"(__dPdx), "f"(__dPdy)));
313 __IMPL_ALIAS("__itex1DGrad", "__tex1DGrad_v2");
314 
315 __IMPL_S("__tex1DLayered_v2", "__tex1DLayered_rmnf_v2",
316          (float __x, int __layer), "tex.a1d.v4", "f32",
317          "{%0, %1, %2, %3}, [%4, {%5, %6}];", ("r"(__layer), "f"(__x)));
318 __IMPL_ALIAS("__itex1DLayered", "__tex1DLayered_v2");
319 
320 __IMPL_S("__tex1DLayeredGrad_v2", "__tex1DLayeredGrad_rmnf_v2",
321          (float __x, int __layer, float __dPdx, float __dPdy),
322          "tex.grad.a1d.v4", "f32",
323          "{%0, %1, %2, %3}, [%4, {%5, %6}], {%7}, {%8};",
324          ("r"(__layer), "f"(__x), "f"(__dPdx), "f"(__dPdy)));
325 __IMPL_ALIAS("__itex1DLayeredGrad", "__tex1DLayeredGrad_v2");
326 
327 __IMPL_S("__tex1DLayeredLod_v2", "__tex1DLayeredLod_rmnf_v2",
328          (float __x, int __layer, float __level), "tex.level.a1d.v4", "f32",
329          "{%0, %1, %2, %3}, [%4, {%5, %6}], %7;",
330          ("r"(__layer), "f"(__x), "f"(__level)));
331 __IMPL_ALIAS("__itex1DLayeredLod", "__tex1DLayeredLod_v2");
332 
333 __IMPL_S("__tex1DLod_v2", "__tex1DLod_rmnf_v2", (float __x, float __level),
334          "tex.level.1d.v4", "f32", "{%0, %1, %2, %3}, [%4, {%5}], %6;",
335          ("f"(__x), "f"(__level)));
336 __IMPL_ALIAS("__itex1DLod", "__tex1DLod_v2");
337 
338 // 2D
339 __IMPL_S("__tex2D_v2", "__tex2D_rmnf_v2", (float __x, float __y), "tex.2d.v4",
340          "f32", "{%0, %1, %2, %3}, [%4, {%5, %6}];", ("f"(__x), "f"(__y)));
341 __IMPL_ALIAS("__itex2D", "__tex2D_v2");
342 
343 __IMPL_S3S("__itex2D_sparse", (float __x, float __y, unsigned char *__ir),
344            "{.reg .pred %%p0;\n\t"
345            "tex.2d.v4",
346            "f32",
347            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7}];\n\t"
348            " selp.u16 %4, 1, 0, %%p0; }",
349            ("f"(__x), "f"(__y)));
350 
351 __IMPL_S("__tex2DGrad_v2", "__tex2DGrad_rmnf_v2",
352          (float __x, float __y, const float2 *__dPdx, const float2 *__dPdy),
353          "tex.grad.2d.v4", "f32",
354          "{%0, %1, %2, %3}, [%4, {%5, %6}], {%7, %8}, {%9, %10};",
355          ("f"(__x), "f"(__y), "f"(__dPdx->x), "f"(__dPdx->y), "f"(__dPdy->x),
356           "f"(__dPdy->y)));
357 __IMPL_ALIAS("__itex2DGrad_v2", "__tex2DGrad_v2");
358 
359 __IMPL_S3S("__itex2DGrad_sparse",
360            (float __x, float __y, const float2 *__dPdx, const float2 *__dPdy,
361             unsigned char *__ir),
362            "{.reg .pred %%p0;\n\t"
363            "tex.grad.2d.v4",
364            "f32",
365            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7}], {%8, %9}, {%10, %11};\n\t"
366            "selp.u16 %4, 1, 0, %%p0; }",
367            ("f"(__x), "f"(__y), "f"(__dPdx->x), "f"(__dPdx->y), "f"(__dPdy->x),
368             "f"(__dPdy->y)));
369 
370 __IMPL_S("__tex2DLayered_v2", "__tex2DLayered_rmnf_v2",
371          (float __x, float __y, int __layer), "tex.a2d.v4", "f32",
372          "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}];",
373          ("r"(__layer), "f"(__x), "f"(__y)));
374 __IMPL_ALIAS("__itex2DLayered", "__tex2DLayered_v2");
375 
376 __IMPL_S3S("__itex2DLayered_sparse",
377            (float __x, float __y, int __layer, unsigned char *__ir),
378            "{.reg .pred %%p0;\n\t"
379            "tex.a2d.v4",
380            "f32",
381            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7, %8, %8}];\n\t"
382            "selp.u16 %4, 1, 0, %%p0; }",
383            ("r"(__layer), "f"(__x), "f"(__y)));
384 
385 __IMPL_S("__tex2DLayeredGrad_v2", "__tex2DLayeredGrad_rmnf_v2",
386          (float __x, float __y, int __layer, const float2 *__dPdx,
387           const float2 *__dPdy),
388          "tex.grad.a2d.v4", "f32",
389          "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}], {%8, %9}, {%10, %11};",
390          ("r"(__layer), "f"(__x), "f"(__y), "f"(__dPdx->x), "f"(__dPdx->y),
391           "f"(__dPdy->x), "f"(__dPdy->y)));
392 __IMPL_ALIAS("__itex2DLayeredGrad_v2", "__tex2DLayeredGrad_v2");
393 
394 __IMPL_S3S(
395     "__itex2DLayeredGrad_sparse",
396     (float __x, float __y, int __layer, const float2 *__dPdx,
397      const float2 *__dPdy, unsigned char *__ir),
398     "{.reg .pred %%p0;\n\t"
399     "tex.grad.a2d.v4",
400     "f32",
401     "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7, %8, %8}], {%9, %10}, {%11, %12};\n\t"
402     "selp.u16 %4, 1, 0, %%p0; }",
403     ("r"(__layer), "f"(__x), "f"(__y), "f"(__dPdx->x), "f"(__dPdx->y),
404      "f"(__dPdy->x), "f"(__dPdy->y)));
405 
406 __IMPL_S("__tex2DLayeredLod_v2", "__tex2DLayeredLod_rmnf_v2",
407          (float __x, float __y, int __layer, float __level), "tex.level.a2d.v4",
408          "f32", "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}], %8;",
409          ("r"(__layer), "f"(__x), "f"(__y), "f"(__level)));
410 __IMPL_ALIAS("__itex2DLayeredLod", "__tex2DLayeredLod_v2");
411 
412 __IMPL_S3S("__itex2DLayeredLod_sparse",
413            (float __x, float __y, int __layer, float __level,
414             unsigned char *__ir),
415            "{.reg .pred %%p0;\n\t"
416            "tex.level.a2d.v4",
417            "f32",
418            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7, %8, %8}], %9;\n\t"
419            "selp.u16 %4, 1, 0, %%p0; }",
420            ("r"(__layer), "f"(__x), "f"(__y), "f"(__level)));
421 
422 __IMPL_S("__tex2DLod_v2", "__tex2DLod_rmnf_v2",
423          (float __x, float __y, float __level), "tex.level.2d.v4", "f32",
424          "{%0, %1, %2, %3}, [%4, {%5, %6}], %7;",
425          ("f"(__x), "f"(__y), "f"(__level)));
426 __IMPL_ALIAS("__itex2DLod", "__tex2DLod_v2");
427 
428 __IMPL_S3S("__itex2DLod_sparse",
429            (float __x, float __y, float __level, unsigned char *__ir),
430            "{.reg .pred %%p0;\n\t"
431            "tex.level.2d.v4",
432            "f32",
433            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7}], %8;\n\t"
434            "selp.u16 %4, 1, 0, %%p0; }",
435            ("f"(__x), "f"(__y), "f"(__level)));
436 
437 // 2D gather is special. Unlike other variants that translate into exactly one
438 // asm instruction, it uses one of the four different instructions selected by
439 // __comp.  We implement each instruction variant separately, and dispatch the
440 // right one from the manually implemented 'umbrella' fetch.
441 #define __IMPL_2DGATHER(variant, instr)                                        \
442   __IMPL_SI(__IDV("__tex2Dgather_v2", variant),                                \
443             __IDV("__tex2Dgather_rmnf_v2", variant),                           \
444             (float __x, float __y, int __comp), instr, "f32",                  \
445             "{%0, %1, %2, %3}, [%4, {%5, %6}];", ("f"(__x), "f"(__y)));        \
446   __IMPL_ALIASI(__IDV("__itex2Dgather", variant),                              \
447                 __IDV("__tex2Dgather_v2", variant));                           \
448   __IMPL_S3SI(__IDV("__itex2Dgather_sparse", variant),                         \
449               (float __x, float __y, unsigned char *__ir, int __comp),         \
450               "{.reg .pred %%p0;\n\t" instr, "f32",                            \
451               "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7}];\n\t"                     \
452               "selp.u16 %4, 1, 0, %%p0; }",                                    \
453               ("f"(__x), "f"(__y)));
454 __IMPL_2DGATHER(0, "tld4.r.2d.v4");
455 __IMPL_2DGATHER(1, "tld4.g.2d.v4");
456 __IMPL_2DGATHER(2, "tld4.b.2d.v4");
457 __IMPL_2DGATHER(3, "tld4.a.2d.v4");
458 
459 // Umbrella dispatcher -- calls into specific 2Dgather variant.
460 template <> struct __tex_fetch_v4<__ID("__tex2Dgather_v2")> {
461   template <class __T>
462   __device__ static __T __run(cudaTextureObject_t __obj, float __x, float __y,
463                               int __comp) {
464     switch (__comp) {
465     case 0:
466       return __tex_fetch_v4<__IDV("__tex2Dgather_v2", 0)>::__run<__T>(
467           __obj, __x, __y, __comp);
468     case 1:
469       return __tex_fetch_v4<__IDV("__tex2Dgather_v2", 1)>::__run<__T>(
470           __obj, __x, __y, __comp);
471     case 2:
472       return __tex_fetch_v4<__IDV("__tex2Dgather_v2", 2)>::__run<__T>(
473           __obj, __x, __y, __comp);
474     case 3:
475       return __tex_fetch_v4<__IDV("__tex2Dgather_v2", 3)>::__run<__T>(
476           __obj, __x, __y, __comp);
477     }
478   }
479 };
480 __IMPL_ALIAS("__itex2Dgather", "__tex2Dgather_v2");
481 
482 template <> struct __tex_fetch_v4<__ID("__tex2Dgather_rmnf_v2")> {
483   template <class __T>
484   __device__ static float4 __run(cudaTextureObject_t __obj, float __x,
485                                  float __y, int __comp) {
486     switch (__comp) {
487     case 0:
488       return __tex_fetch_v4<__IDV("__tex2Dgather_rmnf_v2", 0)>::__run<__T>(
489           __obj, __x, __y, __comp);
490     case 1:
491       return __tex_fetch_v4<__IDV("__tex2Dgather_rmnf_v2", 1)>::__run<__T>(
492           __obj, __x, __y, __comp);
493     case 2:
494       return __tex_fetch_v4<__IDV("__tex2Dgather_rmnf_v2", 2)>::__run<__T>(
495           __obj, __x, __y, __comp);
496     case 3:
497       return __tex_fetch_v4<__IDV("__tex2Dgather_rmnf_v2", 3)>::__run<__T>(
498           __obj, __x, __y, __comp);
499     }
500   }
501 };
502 
503 #if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__ >= 600)
504 template <> struct __tex_fetch_v4<__ID("__itex2Dgather_sparse")> {
505   template <class __T>
506   __device__ static __T __run(cudaTextureObject_t __obj, float __x, float __y,
507                               unsigned char *__ir, int __comp) {
508     switch (__comp) {
509     case 0:
510       return __tex_fetch_v4<__IDV("__itex2Dgather_sparse", 0)>::__run<__T>(
511           __obj, __x, __y, __ir, __comp);
512     case 1:
513       return __tex_fetch_v4<__IDV("__itex2Dgather_sparse", 1)>::__run<__T>(
514           __obj, __x, __y, __ir, __comp);
515     case 2:
516       return __tex_fetch_v4<__IDV("__itex2Dgather_sparse", 2)>::__run<__T>(
517           __obj, __x, __y, __ir, __comp);
518     case 3:
519       return __tex_fetch_v4<__IDV("__itex2Dgather_sparse", 3)>::__run<__T>(
520           __obj, __x, __y, __ir, __comp);
521     }
522   }
523 };
524 #endif
525 
526 // 3D
527 __IMPL_S("__tex3D_v2", "__tex3D_rmnf_v2", (float __x, float __y, float __z),
528          "tex.3d.v4", "f32", "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}];",
529          ("f"(__x), "f"(__y), "f"(__z)));
530 __IMPL_ALIAS("__itex3D", "__tex3D_v2");
531 
532 __IMPL_S3S("__itex3D_sparse",
533            (float __x, float __y, float __z, unsigned char *__ir),
534            "{.reg .pred %%p0;\n\t"
535            "tex.3d.v4",
536            "f32",
537            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7, %8, %8}];\n\t"
538            "selp.u16 %4, 1, 0, %%p0; }",
539            ("f"(__x), "f"(__y), "f"(__z)));
540 
541 __IMPL_S("__tex3DGrad_v2", "__tex3DGrad_rmnf_v2",
542          (float __x, float __y, float __z, const float4 *__dPdx,
543           const float4 *__dPdy),
544          "tex.grad.3d.v4", "f32",
545          "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}], "
546          "{%8, %9, %10, %10}, {%11, %12, %13, %13};",
547          ("f"(__x), "f"(__y), "f"(__z), "f"(__dPdx->x), "f"(__dPdx->y),
548           "f"(__dPdx->z), "f"(__dPdy->x), "f"(__dPdy->y), "f"(__dPdy->z)));
549 __IMPL_ALIAS("__itex3DGrad_v2", "__tex3DGrad_v2");
550 
551 __IMPL_S3S("__itex3DGrad_sparse",
552            (float __x, float __y, float __z, const float4 *__dPdx,
553             const float4 *__dPdy, unsigned char *__ir),
554            "{.reg .pred %%p0;\n\t"
555            "tex.grad.3d.v4",
556            "f32",
557            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7, %8, %8}], "
558            "{%9, %10, %11, %11}, {%12, %13, %14, %14};\n\t"
559            "selp.u16 %4, 1, 0, %%p0; }",
560            ("f"(__x), "f"(__y), "f"(__z), "f"(__dPdx->x), "f"(__dPdx->y),
561             "f"(__dPdx->z), "f"(__dPdy->x), "f"(__dPdy->y), "f"(__dPdy->z)));
562 
563 __IMPL_S("__tex3DLod_v2", "__tex3DLod_rmnf_v2",
564          (float __x, float __y, float __z, float __level), "tex.level.3d.v4",
565          "f32", "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}], %8;",
566          ("f"(__x), "f"(__y), "f"(__z), "f"(__level)));
567 __IMPL_ALIAS("__itex3DLod", "__tex3DLod_v2");
568 
569 __IMPL_S3S("__itex3DLod_sparse",
570            (float __x, float __y, float __z, float __level,
571             unsigned char *__ir),
572            "{.reg .pred %%p0;\n\t"
573            "tex.level.3d.v4",
574            "f32",
575            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7, %8, %8}], %9;\n\t"
576            "selp.u16 %4, 1, 0, %%p0; }",
577            ("f"(__x), "f"(__y), "f"(__z), "f"(__level)));
578 
579 // Cubemap
580 __IMPL_S("__texCubemap_v2", "__texCubemap_rmnf_v2",
581          (float __x, float __y, float __z), "tex.cube.v4", "f32",
582          "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}];",
583          ("f"(__x), "f"(__y), "f"(__z)));
584 __IMPL_ALIAS("__itexCubemap", "__texCubemap_v2");
585 
586 __IMPL_S3S("__itexCubemap_sparse",
587            (float __x, float __y, float __z, unsigned char *__ir),
588            "{.reg .pred %%p0;\n\t"
589            "tex.cube.v4",
590            "f32",
591            "{%0, %1, %2, %3}|%%p0, [%5, {%6, %7, %8, %8}];\n\t"
592            "selp.u16 %4, 1, 0, %%p0; }",
593            ("f"(__x), "f"(__y), "f"(__z)));
594 
595 __IMPL_S("__texCubemapGrad_v2", "__texCubemapGrad_rmnf_v2",
596          (float __x, float __y, float __z, const float4 *__dPdx,
597           const float4 *__dPdy),
598          "tex.grad.cube.v4", "f32",
599          "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}], "
600          "{%8, %9, %10, %10}, {%11, %12, %13, %13};",
601          ("f"(__x), "f"(__y), "f"(__z), "f"(__dPdx->x), "f"(__dPdx->y),
602           "f"(__dPdx->z), "f"(__dPdy->x), "f"(__dPdy->y), "f"(__dPdy->z)));
603 __IMPL_ALIAS("__itexCubemapGrad_v2", "__texCubemapGrad_v2");
604 
605 __IMPL_S("__texCubemapLayered_v2", "__texCubemapLayered_rmnf_v2",
606          (float __x, float __y, float __z, int __layer), "tex.acube.v4", "f32",
607          "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %8}];",
608          ("r"(__layer), "f"(__x), "f"(__y), "f"(__z)));
609 __IMPL_ALIAS("__itexCubemapLayered", "__texCubemapLayered_v2");
610 
611 __IMPL_S("__texCubemapLayeredGrad_v2", "__texCubemapLayeredGrad_rmnf_v2",
612          (float __x, float __y, float __z, int __layer, const float4 *__dPdx,
613           const float4 *__dPdy),
614          "tex.grad.acube.v4", "f32",
615          "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %8}], "
616          "{%9, %10, %11, %11}, {%12, %13, %14, %14};",
617          ("r"(__layer), "f"(__x), "f"(__y), "f"(__z), "f"(__dPdx->x),
618           "f"(__dPdx->y), "f"(__dPdx->z), "f"(__dPdy->x), "f"(__dPdy->y),
619           "f"(__dPdy->z)));
620 __IMPL_ALIAS("__itexCubemapLayeredGrad_v2", "__texCubemapLayeredGrad_v2");
621 
622 __IMPL_S("__texCubemapLayeredLod_v2", "__texCubemapLayeredLod_rmnf_v2",
623          (float __x, float __y, float __z, int __layer, float __level),
624          "tex.level.acube.v4", "f32",
625          "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %8}], %9;",
626          ("r"(__layer), "f"(__x), "f"(__y), "f"(__z), "f"(__level)));
627 __IMPL_ALIAS("__itexCubemapLayeredLod", "__texCubemapLayeredLod_v2");
628 
629 __IMPL_S("__texCubemapLod_v2", "__texCubemapLod_rmnf_v2",
630          (float __x, float __y, float __z, float __level), "tex.level.cube.v4",
631          "f32", "{%0, %1, %2, %3}, [%4, {%5, %6, %7, %7}], %8;",
632          ("f"(__x), "f"(__y), "f"(__z), "f"(__level)));
633 __IMPL_ALIAS("__itexCubemapLod", "__texCubemapLod_v2");
634 
635 // Helper class for extracting slice of data from V4 fetch results.
636 template <class __DestT, class __SrcT> struct __convert {
637   template <int __NElements = sizeof(__DestT) /
638                               sizeof(typename __TypeInfoT<__DestT>::__base_t)>
639   __device__ static __DestT __run(__SrcT __v);
640   template <> __device__ static __DestT __run<1>(__SrcT __v) { return {__v.x}; }
641   template <> __device__ static __DestT __run<2>(__SrcT __v) {
642     return {__v.x, __v.y};
643   }
644   template <> __device__ static __DestT __run<3>(__SrcT __v) {
645     return {__v.x, __v.y, __v.z};
646   }
647   template <> __device__ static __DestT __run<4>(__SrcT __v) {
648     return {__v.x, __v.y, __v.z, __v.w};
649   }
650 };
651 
652 // These are the top-level function overloads the __nv_tex_surf_handler expands
653 // to.  Each overload deals with one of the several ways __nv_tex_surf_handler
654 // is called by CUDA headers. In the end, each of the overloads does the same
655 // job -- it figures out which `__tex_fetch_v4::run` variant should be used to
656 // fetch texture data and which `__convert::run` is needed to convert it into
657 // appropriate return type.
658 
659 // __nv_tex_surf_handler("__tex...", &ret, cudaTextureObject_t handle, args...);
660 //   Data type and return type are based on ret.
661 template <class __op, class __T, class... __Args>
662 __device__ static void __tex_fetch(__T *__ptr, cudaTextureObject_t __handle,
663                                    __Args... __args) {
664   using __FetchT = typename __TypeInfoT<__T>::__fetch_t;
665   *__ptr = __convert<__T, __FetchT>::__run(
666       __tex_fetch_v4<__op>::template __run<__FetchT>(__handle, __args...));
667 }
668 
669 #if CUDA_VERSION < 12000
670 // texture<> objects get magically converted into a texture reference.  However,
671 // there's no way to convert them to cudaTextureObject_t on C++ level. So, we
672 // cheat a bit and use inline assembly to do it. It costs us an extra register
673 // and a move, but that is easy for ptxas to optimize away.
674 template <class __T>
675 __device__ cudaTextureObject_t __tex_handle_to_obj(__T __handle) {
676   cudaTextureObject_t __obj;
677   asm("mov.b64 %0, %1; " : "=l"(__obj) : "l"(__handle));
678   return __obj;
679 }
680 
681 // __nv_tex_surf_handler ("__tex...", &ret, textureReference, args...);
682 //   Data type and return type is based on ret.
683 template <class __op, class __T, class __HandleT, class... __Args>
684 __device__ static void __tex_fetch(__T *__ptr, __HandleT __handle,
685                                    __Args... __args) {
686   using __FetchT = typename __TypeInfoT<__T>::__fetch_t;
687   *__ptr = __convert<__T, __FetchT>::__run(
688       __tex_fetch_v4<__op>::template __run<__FetchT>(
689           __tex_handle_to_obj(__handle), __args...));
690 }
691 
692 // __nv_tex_surf_handler ("__tex...", &type_dummy, &ret, texture<...>, args...);
693 // cudaReadModeNormalizedFloat fetches always return float4.
694 template <class __op, class __DataT, class __RetT, int __TexT, class... __Args>
695 __device__ static void
696 __tex_fetch(__DataT *, __RetT *__ptr,
697             texture<__DataT, __TexT, cudaReadModeNormalizedFloat> __handle,
698             __Args... __args) {
699   using __FetchT = typename __TypeInfoT<__DataT>::__fetch_t;
700   *__ptr = __convert<__RetT, float4>::__run(
701       __tex_fetch_v4<__op>::template __run<__FetchT>(
702           __tex_handle_to_obj(__handle), __args...));
703 }
704 
705 // __nv_tex_surf_handler ("__tex...", &type_dummy, &ret, texture<...>, args...);
706 // For cudaReadModeElementType fetch return type is based on type_dummy.
707 template <class __op, class __DataT, class __RetT, int __TexT, class... __Args>
708 __device__ static void
709 __tex_fetch(__DataT *, __RetT *__ptr,
710             texture<__DataT, __TexT, cudaReadModeElementType> __handle,
711             __Args... __args) {
712   using __FetchT = typename __TypeInfoT<__DataT>::__fetch_t;
713   *__ptr = __convert<__RetT, __FetchT>::__run(
714       __tex_fetch_v4<__op>::template __run<__FetchT>(
715           __tex_handle_to_obj(__handle), __args...));
716 }
717 #endif // CUDA_VERSION
718 } // namespace __cuda_tex
719 } // namespace
720 #pragma pop_macro("__ASM_OUT")
721 #pragma pop_macro("__ASM_OUTP")
722 #pragma pop_macro("__Args")
723 #pragma pop_macro("__ID")
724 #pragma pop_macro("__IDV")
725 #pragma pop_macro("__IMPL_2DGATHER")
726 #pragma pop_macro("__IMPL_ALIAS")
727 #pragma pop_macro("__IMPL_ALIASI")
728 #pragma pop_macro("__IMPL_F1")
729 #pragma pop_macro("__IMPL_F3")
730 #pragma pop_macro("__IMPL_F3N")
731 #pragma pop_macro("__IMPL_F3S")
732 #pragma pop_macro("__IMPL_S")
733 #pragma pop_macro("__IMPL_S3")
734 #pragma pop_macro("__IMPL_S3I")
735 #pragma pop_macro("__IMPL_S3N")
736 #pragma pop_macro("__IMPL_S3NI")
737 #pragma pop_macro("__IMPL_S3S")
738 #pragma pop_macro("__IMPL_S3SI")
739 #pragma pop_macro("__IMPL_SI")
740 #pragma pop_macro("__L")
741 #pragma pop_macro("__STRIP_PARENS")
742 #endif // __CLANG_CUDA_TEXTURE_INTRINSICS_H__
743