builtin.h 16.3 KB
Newer Older
1
2
3
4
5
6
7
8
9
/*!
 * \file tl/op/builtin.h
 * \brief Builtin intrinsics.
 *
 */

#ifndef TVM_TL_OP_BUILTIN_H_
#define TVM_TL_OP_BUILTIN_H_

10
#include "operator.h"
11
#include <tvm/ir/transform.h>
12
13

namespace tvm {
14
15
16
17
18
19
20
21
/*!
 * \brief Create the TVM intrinsic that initializes a PTX fence barrier.
 *
 * Initializes a PTX fence-style barrier used to coordinate asynchronous memory
 * operations (for example, TMA/TMA_STORE). Returns the Op representing this
 * intrinsic for use in TIR lowering and code generation.
 *
 */
22
namespace tl {
23
24

namespace attr {
25
static constexpr const char *kSafeValueMap = "safe_value_map";
26
27
static constexpr const char *kWarpSpecializationScope =
    "kWarpSpecializationScope";
28
29
static constexpr const char *kCustomWarpSpecialization =
    "kCustomWarpSpecialization";
30
static constexpr const char *kLocalVarInit = "tl.local_var_init";
31
32
33
34
// A PrimFunc-level attribute carrying a list of handle Vars
// that must NOT be marked with the restrict qualifier in codegen.
// Type: Array<tir::Var>
static constexpr const char *kNonRestrictParams = "tl.non_restrict_params";
35
36
} // namespace attr

37
38
static constexpr const char *kDebugMergeSharedMemoryAllocations =
    "tl.debug_merge_shared_memory_allocations";
39
static constexpr const char *kDisableTMALower = "tl.disable_tma_lower";
40
41
static constexpr const char *kDisableSafeMemoryLegalize =
    "tl.disable_safe_memory_legalize";
42
43
static constexpr const char *kDisableWarpSpecialized =
    "tl.disable_warp_specialized";
44
static constexpr const char *kConfigIndexBitwidth = "tl.config_index_bitwidth";
45
46
static constexpr const char *kEnableAggressiveSharedMemoryMerge =
    "tl.enable_aggressive_shared_memory_merge";
47
static constexpr const char *kDisableFastMath = "tl.disable_fast_math";
48
static constexpr const char *kEnableFastMath = "tl.enable_fast_math";
49
50
static constexpr const char *kPtxasRegisterUsageLevel =
    "tl.ptxas_register_usage_level";
51
52
static constexpr const char *kEnablePTXASVerboseOutput =
    "tl.enable_ptxas_verbose_output";
53
static constexpr const char *kDisableVectorize256 = "tl.disable_vectorize_256";
54
static constexpr const char *kDisableWGMMA = "tl.disable_wgmma";
55
static constexpr const char *kDisableShuffleElect = "tl.disable_shuffle_elect";
56
57
static constexpr const char *kStorageRewriteDetectInplace =
    "tl.storage_rewrite_detect_inplace";
58
59
60
61
static constexpr const char *kLayoutVisualizationEnable =
    "tl.layout_visualization_enable";
static constexpr const char *kLayoutVisualizationFormats =
    "tl.layout_visualization_formats";
62
static constexpr const char *kDeviceCompileFlags = "tl.device_compile_flags";
63

64
65
66
67
68
69
70
71
72
73
74
75
76
77
/*!
 * \brief Whether to disable thread storage synchronization
 *
 * When enabled, disables the automatic insertion of thread synchronization
 * barriers (e.g., __syncthreads()) for shared memory access coordination.
 * This can be useful for performance optimization in cases where manual
 * synchronization is preferred or when synchronization is not needed.
 *
 * kDisableThreadStorageSync = "tl.disable_thread_storage_sync"
 *
 */
static constexpr const char *kDisableThreadStorageSync =
    "tl.disable_thread_storage_sync";

78
79
80
81
82
83
84
85
/*!
 * \brief Force inline Let bindings during simplification.
 *
 * kForceLetInline = "tl.force_let_inline"
 *
 */
static constexpr const char *kForceLetInline = "tl.force_let_inline";

86
87
88
89
90
91
92
93
/*!
 * \brief Get the type of the CUDA tensor map
 *
 * DataType cuTensorMapType()
 *
 */
DataType cuTensorMapType();

94
// fast math related op
95
// __exp(x) - fast exponential
96
TVM_DLL const Op &__exp();
97
// __exp10(x) - fast base-10 exponential
98
TVM_DLL const Op &__exp10();
99
// __log(x) - fast natural logarithm
100
TVM_DLL const Op &__log();
101
// __log2(x) - fast base-2 logarithm
102
TVM_DLL const Op &__log2();
103
// __log10(x) - fast base-10 logarithm
104
TVM_DLL const Op &__log10();
105
// __tan(x) - fast tangent
106
TVM_DLL const Op &__tan();
107
// __cos(x) - fast cosine
108
TVM_DLL const Op &__cos();
109
// __sin(x) - fast sine
110
111
TVM_DLL const Op &__sin();

112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
// high precision with IEEE-compliant.
// ieee_add(x, y, rounding_mode) - IEEE-compliant addition
TVM_DLL const Op &ieee_add();
// ieee_sub(x, y, rounding_mode) - IEEE-compliant subtraction
TVM_DLL const Op &ieee_sub();
// ieee_mul(x, y, rounding_mode) - IEEE-compliant multiplication
TVM_DLL const Op &ieee_mul();
// ieee_fmaf(x, y, z, rounding_mode) - IEEE-compliant fused multiply-add
TVM_DLL const Op &ieee_fmaf();
// ieee_frcp(x, rounding_mode) - IEEE-compliant reciprocal
TVM_DLL const Op &ieee_frcp();
// ieee_fsqrt(x, rounding_mode) - IEEE-compliant square root
TVM_DLL const Op &ieee_fsqrt();
// ieee_frsqrt(x) - IEEE-compliant reciprocal square root (rn only)
TVM_DLL const Op &ieee_frsqrt();
// ieee_fdiv(x, y, rounding_mode) - IEEE-compliant division
TVM_DLL const Op &ieee_fdiv();

130
131
132
133
// random op
TVM_DLL const Op &rng_init();
TVM_DLL const Op &rng_rand();

134
135
136
/*!
 * \brief tvm intrinsics for TMADescriptor creation for tiled load
 *
137
 * CuTensorMap* create_tma_descriptor(data_type, rank, global_addr,
138
139
 * global_shape..., global_stride..., smem_box..., smem_stride..., interleave,
 * swizzle, l2_promotion, oob_fill)
140
141
 *
 */
142
TVM_DLL const Op &create_tma_descriptor();
143
144
145
146

/*!
 * \brief tvm intrinsics for TMADescriptor creation for image to column load
 *
147
 * CuTensorMap* create_tma_im2col_descriptor(data_type, rank, global_addr,
148
149
150
 * global_shape..., global_stride..., elem_stride..., lower_corner...,
 * upper_corner..., smme_box_pixel, smem_box_channel, interleave, swizzle,
 * l2_promotion, oob_fill)
151
152
 *
 */
153
TVM_DLL const Op &create_tma_im2col_descriptor();
154
155
156
157

/*!
 * \brief Create a list of mbarrier with num_threads
 *
158
 * create_list_of_mbarrier(num_threads0, num_threads1, ...)
159
160
 *
 */
161
TVM_DLL const Op &create_list_of_mbarrier();
162
163
164
165
166
167
168

/*!
 * \brief Get the mbarrier with barrier_id
 *
 * int64_t* GetMBarrier(barrier_id)
 *
 */
169
TVM_DLL const Op &get_mbarrier();
170
171

/*!
172
173
 * \brief tvm intrinsics for loading data from global tensor descriptor to
 * shared memory
174
 *
175
 * tma_load(descriptor, mbarrier, smem_data, coord_0, coord_1, ...)
176
177
 *
 */
178
TVM_DLL const Op &tma_load();
179
180

/*!
181
182
 * \brief tvm intrinsics for loading image from global tensor to columns in
 * shared memory
183
 *
184
 * tma_load(descriptor, mbarrier, smem_data, coord_0, coord_1, ...,
185
 * image_offset, ...)
186
187
 *
 */
188
TVM_DLL const Op &tma_load_im2col();
189
190

/*!
191
192
 * \brief tvm intrinsics for storing data from shared memory to global tensor
 * descriptor
193
 *
194
 * tma_store(descriptor, smem_data, coord_0, coord_1, ...)
195
196
 *
 */
197
TVM_DLL const Op &tma_store();
198

199
200
201
202
203
204
205
206
/*!
 * \brief tvm intrinsics for barrier initialization fence
 *
 * ptx_fence_barrier_init()
 *
 */
const Op &ptx_fence_barrier_init();

207
208
209
/*!
 * \brief tvm intrinsics for mbarrier wait with parity bit
 *
210
 * mbarrier_wait_parity(mbarrier, parity)
211
212
 *
 */
213
TVM_DLL const Op &mbarrier_wait_parity();
214
215
216
217

/*!
 * \brief tvm intrinsics for mbarrier expect tx
 *
218
 * mbarrier_expect_tx(mbarrier, transaction_bytes)
219
220
 *
 */
221
TVM_DLL const Op &mbarrier_expect_tx();
222

223
224
225
226
227
228
229
230
231
232
233
234
235
236
/*!
 * \brief tvm intrinsic for ptx tensor core wgmma instructions.
 *
 *  void ptx_wgmma_ss(StringImm accum_dtype, StringImm wgmma_prefix, bool
 * a_is_k_major, bool b_is_k_major, StringImm a_dtype_abbrv, StringImm
 * b_dtype_abbrv, StringImm accum_dtype_abbrv, Var A_descriptor, PrimExpr
 * A_offset, Var B_descriptor, Var B_offset, Var C_data, Var C_offset, bool
 * scale_out, bool scale_in_a, bool scale_in_b);
 */
TVM_DLL const Op &ptx_wgmma_ss();

/*!
 * \brief tvm intrinsics for ptx tensor core wgmma instructions.
 *
237
238
239
240
241
 *  void ptx_wgmma_rs(StringImm accum_dtype, StringImm wgmma_prefix,
 * bool b_is_k_major, StringImm a_dtype_abbrv, StringImm b_dtype_abbrv,
 * StringImm accum_dtype_abbrv, Var A_descriptor, PrimExpr A_offset, Var
 * B_descriptor, Var B_offset, Var C_data, Var C_offset, bool scale_out,
 * bool scale_in_a, bool scale_in_b);
242
243
244
 */
TVM_DLL const Op &ptx_wgmma_rs();

245
246
247
248
249
250
251
252
253
254
/*!
 * \brief tvm intrinsic for tcgen05 mma shared-shared instructions.
 */
TVM_DLL const Op &ptx_tcgen05_mma_ss();

/*!
 * \brief tvm intrinsic for tcgen05 mma tensor-shared instructions.
 */
TVM_DLL const Op &ptx_tcgen05_mma_ts();

255
256
257
258
259
260
/*!
 * \brief tvm intrinsics for initializing tensor memory
 *
 * ptx_init_tensor_memory(tmem_buffer, num_cols)
 *
 */
261
TVM_DLL const Op &ptx_init_tensor_memory();
262
263
264
265
266
267
268

/*!
 * \brief tvm intrinsics for deallocating tensor memory
 *
 * tmem_deallocate(tmem_buffer)
 *
 */
269
TVM_DLL const Op &ptx_deallocate_tensor_memory();
270

271
272
273
274
275
276
277
278
279
280
281
/*!
 * \brief tvm intrinsic for ptx tensor core mma instructions on SM70.
 *
 *  void ptx_mma_sm70(StringImm shape, StringImm A_layout, StringImm B_layout,
 *                    StringImm A_dtype, StringImm B_dtype, StringImm C_dtype,
 *                    Var multiplicand_a, Expr a_index,
 *                    Var multiplicand_b, Expr b_index,
 *                    Var accumulator, Expr c_index, bool saturate);
 */
TVM_DLL const Op &ptx_mma_sm70();

282
283
284
/*!
 * \brief tvm intrinsics for ldmatrix
 *
285
 * ptx_ldmatrix(transposed, num, shared_addr, local_addr)
286
287
 *
 */
288
TVM_DLL const Op &ptx_ldmatrix();
289
290
291
292

/*!
 * \brief tvm intrinsics for stmatrix
 *
293
 * ptx_ldmatrix(transposed, num, shared_addr, int32_values...)
294
295
 *
 */
296
TVM_DLL const Op &ptx_stmatrix();
297

298
299
300
301
302
303
304
305
/*!
 * \brief tvm intrinsic for ptx async copy barrier using
 * cp.async.mbarrier.arrive.noinc
 *
 *  This op is used to represent a ptx async copy barrier operation in tilelang.
 */
TVM_DLL const Op &ptx_cp_async_barrier_noinc();

306
307
308
/*!
 * \brief Pack two b16 value into a b32 value
 *
309
 * int32 pack_b16(b16_value, b16_value)
310
311
 *
 */
312
TVM_DLL const Op &pack_b16();
313
314
315
316
317
318
319

/*!
 * \brief Issue a shared memory fence for async operations
 *
 * FenceProxyAsync()
 *
 */
320
TVM_DLL const Op &fence_proxy_async();
321

322
323
324
/*!
 * \brief Indicate arrival of warp issuing TMA_STORE
 *
325
 * tma_store_arrive()
326
327
 *
 */
328
TVM_DLL const Op &tma_store_arrive();
329
330
331
332

/*!
 * \brief Wait for TMA_STORE to finish
 *
333
 * tma_store_wait()
334
335
 *
 */
336
TVM_DLL const Op &tma_store_wait();
337

338
339
340
341
342
343
/*!
 * \brief Set reg hint for warp-specialized branched
 *
 * SetMaxNRegInc(num_reg, is_inc)
 *
 */
344
TVM_DLL const Op &set_max_nreg();
345

346
347
348
/*!
 * \brief No set reg hint for warp-specialized branched
 *
349
 * no_set_max_nreg()
350
351
 *
 */
352
TVM_DLL const Op &no_set_max_nreg();
353

354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
/*!
 * \brief Arrive at a warpgroup fence for WGMMA sequences
 *
 * warpgroup_arrive()
 *
 */
TVM_DLL const Op &warpgroup_arrive();

/*!
 * \brief Commit the current warpgroup batch for WGMMA sequences
 *
 * warpgroup_commit_batch()
 *
 */
TVM_DLL const Op &warpgroup_commit_batch();

/*!
 * \brief Wait for the warpgroup batch identified by num_mma
 *
 * warpgroup_wait(num_mma)
 *
 */
TVM_DLL const Op &warpgroup_wait();

378
379
380
381
382
383
384
385
/*!
 * \brief Fence accumulator operand registers for upcoming WGMMA operations
 *
 * warpgroup_fence_operand(dtype, ptr, offset, num_regs)
 *
 */
TVM_DLL const Op &warpgroup_fence_operand();

386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
/*!
 * \brief Return the canonical lane index for the calling thread.
 *
 * get_lane_idx([warp_size])
 *
 */
TVM_DLL const Op &get_lane_idx();

/*!
 * \brief Return the canonical warp index, assuming converged threads.
 *
 * get_warp_idx_sync([warp_size])
 *
 */
TVM_DLL const Op &get_warp_idx_sync();

/*!
 * \brief Return the canonical warp index without synchronizing the warp.
 *
 * get_warp_idx([warp_size])
 *
 */
TVM_DLL const Op &get_warp_idx();

/*!
 * \brief Return the canonical warp group index for converged threads.
 *
 * get_warp_group_idx([warp_size, warps_per_group])
 *
 */
TVM_DLL const Op &get_warp_group_idx();

418
419
420
/*!
 * \brief Wait the previous wgmma to finish
 *
421
 * wait_wgmma(num_mma)
422
423
 *
 */
424
TVM_DLL const Op &wait_wgmma();
425

426
427
428
429
430
431
/*!
 * \brief Synchronize all threads in a grid
 *
 * sync_grid()
 *
 */
432
TVM_DLL const Op &sync_grid();
433
434
435
436
437
438
439

/*!
 * \brief tvm intrinsic for loop continue
 *
 * loop_break()
 *
 */
440
TVM_DLL const Op &loop_break();
441

442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
/*!
 * \brief tvm intrinsic for amd matrix core mfma instructions.
 *
 *  void tvm_mfma(StringImm shape, StringImm A_layout, StringImm B_layout,
 *               StringImm A_dtype, StringImm B_dtype, StringImm C_dtype,
 *               Var multiplicand_a, Expr a_index,
 *               Var multiplicand_b, Expr b_index,
 *               Var accumulator, Expr c_index);
 */
TVM_DLL const Op &tvm_mfma();

/*!
 * \brief tvm intrinsic for storing the result of AMD MFMA into a destination
 * pointer.
 *
 *        There is no real instruction that does that, but we want to hide
 * details of complex index manipulation behind this intrinsic to simplify TIR
 * lowering passes (e.g. LowerWarpMemory) like cuda ptx backend does.
 *
 * void tvm_mfma_store(IntImm m, IntImm n, Var dst_ptr, Var src_ptr, Expr
 * src_offset, Var dst_stride);
 */
TVM_DLL const Op &tvm_mfma_store();

/*!
 * \brief tvm intrinsic for amd rdna matrix core instructions.
 *
 *  void tvm_rdna_wmma(StringImm shape, StringImm A_layout, StringImm B_layout,
 *               StringImm A_dtype, StringImm B_dtype, StringImm C_dtype,
 *               Var multiplicand_a, Expr a_index,
 *               Var multiplicand_b, Expr b_index,
 *               Var accumulator, Expr c_index);
 */
TVM_DLL const Op &tvm_rdna_wmma();

/*!
 * \brief tvm intrinsic for storing the result of AMD RDNA WMMA into a
 * destination pointer.
 *
 *        There is no real instruction that does that, but we want to hide
 * details of complex index manipulation behind this intrinsic to simplify TIR
 * lowering passes (e.g. LowerWarpMemory) like cuda ptx backend does.
 *
 * void tvm_rdna_wmma_store(IntImm m, IntImm n, Var dst_ptr, Var src_ptr, Expr
 * src_offset, Var dst_stride);
 */
TVM_DLL const Op &tvm_rdna_wmma_store();

490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
/*!
 * \brief tilelang intrinsic for general matrix multiplication (GEMM).
 *
 *  This op is used to represent a generic GEMM operation in tilelang.
 */
TVM_DLL const Op &tl_gemm();

/*!
 * \brief tilelang intrinsic for sparse matrix multiplication (GEMM with
 * sparsity).
 *
 *  This op is used to represent a sparse GEMM operation in tilelang.
 */
TVM_DLL const Op &tl_gemm_sp();

505
506
507
508
509
510
511
/*!
 * \brief tilelang intrinsic for shuffle elect.
 *
 *  This op is used to represent a shuffle elect operation in tilelang.
 */
TVM_DLL const Op &tl_shuffle_elect();

512
513
514
515
516
517
518
/*!
 * \brief tilelang intrinsic for initializing a descriptor buffer for
 * wgmma/utcmma.
 *
 *  This op is used to represent a descriptor initialization operation in
 * tilelang.
 */
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
TVM_DLL const Op &initialize_wgmma_descriptor();

/*!
 * \brief tilelang intrinsic for initializing a descriptor buffer for
 * tcgen05 mma.
 */
TVM_DLL const Op &initialize_tcgen05_descriptor();

/*!
 * \brief tilelang intrinsic for committing UMMA (TCGEN05) barrier arrive.
 *
 *  This op wraps the device-side arrive used to signal completion of MMA work
 *  to a shared-memory mbarrier. It mirrors CUTLASS's umma_arrive.
 */
TVM_DLL const Op &tcgen05_mma_arrive();
534
535
536
537
538
539
540
541

/*!
 * \brief tilelang intrinsic for setting the start address of a descriptor
 * buffer for wgmma/utcmma.
 *
 *  This op is used to represent a descriptor start address setting operation in
 * tilelang.
 */
542

543
TVM_DLL const Op &increase_descriptor_offset();
544

545
546
547
548
549
550
551
/*!
 * \brief tilelang intrinsic for element-wise atomic addition.
 *
 *  This op is used to represent an element-wise atomic add operation in
 * tilelang.
 */
TVM_DLL const Op &atomicadd_elem_op();
552

553
554
555
556
557
558
559
560
561
562
563
564
565
566
/*!
 * \brief tilelang intrinsic for assert on device.
 *
 *  This op is used to represent an assert on device
 */
TVM_DLL const Op &device_assert();

/*!
 * \brief tilelang intrinsic for assert on device with additional message.
 *
 *  This op is used to represent an assert on device with additional message.
 */
TVM_DLL const Op &device_assert_with_msg();

Tong WU's avatar
Tong WU committed
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
/*!
 * \brief tilelang intrinsic for warp reduction sum.
 */
TVM_DLL const Op &warp_reduce_sum();

/*!
 * \brief tilelang intrinsic for warp reduction max.
 */
TVM_DLL const Op &warp_reduce_max();

/*!
 * \brief tilelang intrinsic for warp reduction min.
 */
TVM_DLL const Op &warp_reduce_min();

/*!
 * \brief tilelang intrinsic for warp reduction bitand.
 */
TVM_DLL const Op &warp_reduce_bitand();

/*!
 * \brief tilelang intrinsic for warp reduction bitor.
 */
TVM_DLL const Op &warp_reduce_bitor();

592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
/*!
 * \brief tilelang intrinsic for CUDA read-only cache load (__ldg).
 *
 *  This op allows users to explicitly request a non-coherent cached load
 *  from global memory on CUDA by emitting `__ldg(&ptr[idx])` for 32-bit
 *  element types on supported architectures. It provides a direct way to
 *  leverage the read-only data cache for performance-sensitive loads when
 *  the compiler cannot infer `const __restrict__` automatically.
 *
 *  Usage from TVMScript:
 *    y[i] = T.__ldg(x[i])
 *
 *  The op takes one argument preferred as a BufferLoad identifying the
 *  source element; alternatively, backends may support passing a Buffer and
 *  index expression.
 */
TVM_DLL const Op &__ldg();

610
611
} // namespace tl
} // namespace tvm
612

613
#endif //  TVM_TL_OP_BUILTIN_H_