1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948
|
# mypy: allow-untyped-defs
import dataclasses
import sys
from enum import Enum
from typing import Callable, Dict, List, Optional, Type
import sympy
import torch
from .. import cpp_builder, ir
from ..cpu_vec_isa import pick_vec_isa, VecAMX, VecAVX2, VecAVX512, VecISA
from ..utils import IndentedBuffer, parallel_num_threads
from ..virtualized import V
from .common import KernelTemplate
from .cpp_template_kernel import CppTemplateKernel
from .cpp_utils import DTYPE_TO_CPP, GemmBlocking, value_to_cpp
class LayoutType(Enum):
NORMAL = 0
VNNI2 = 1
VNNI4 = 2
_IS_WINDOWS = sys.platform == "win32"
def get_restrict_keyword() -> str:
if _IS_WINDOWS:
# https://learn.microsoft.com/en-us/cpp/cpp/extension-restrict?view=msvc-170
return "__restrict"
else:
return "__restrict__"
class CppMicroGemm:
"""
A class that codegens a kernel that computes small-sized matrix multiplication.
A micro GEMM kernel is responsible for register blocking, instruction selection,
and other CPU architecture-specific optimizations.
The subclasses need to override `codegen_define` to define the kernel function
that is called by the code generated by `codegen_call`.
"""
# TODO(jgong5): support constant shapes and lds as template args.
DECLARE_KERNEL = r"""
template <bool accum>
inline void {{kernel_name}}(
{%- if kernel_extra_args_declare %}
{{kernel_extra_args_declare}}
{%- endif %}
const {{input_t}}* {{restrict_keyword}} A,
const {{input2_t}}* {{restrict_keyword}} B,
{{output_t}}* {{restrict_keyword}} C,
int64_t M,
int64_t N,
int64_t K,
int64_t lda,
int64_t ldb,
int64_t ldc
)
"""
def __init__(
self,
name,
input_dtype,
input2_dtype,
output_dtype,
compute_dtype,
register_blocking,
alpha=1,
) -> None:
self.name = name
self.input_dtype = input_dtype
assert input2_dtype is not None
self.input2_dtype = input2_dtype
self.output_dtype = output_dtype
self.compute_dtype = compute_dtype
self.register_blocking = register_blocking
self.alpha = alpha
def get_common_options(self):
if self.input_dtype == torch.uint8:
assert self.compute_dtype == torch.int32
assert self.output_dtype == torch.int32
assert self.input2_dtype == torch.int8
return {
"torch": torch,
"kernel_name": self.name,
"input_dtype": self.input_dtype,
"input2_dtype": self.input2_dtype,
"output_dtype": self.output_dtype,
"compute_dtype": self.compute_dtype,
"input_t": DTYPE_TO_CPP[self.input_dtype],
"input2_t": DTYPE_TO_CPP[self.input2_dtype],
"output_t": DTYPE_TO_CPP[self.output_dtype],
"compute_t": DTYPE_TO_CPP[self.compute_dtype],
"alpha": self.alpha,
"kernel_extra_args_declare": self.get_kernel_extra_args_declare(),
"int8_gemm": self.input_dtype == torch.uint8,
"vnni_size": 4 if self.input_dtype == torch.uint8 else 2,
"restrict_keyword": get_restrict_keyword(),
"is_msvc_compiler": cpp_builder.is_msvc_cl(),
}
def get_kernel_declaration(self):
options = self.get_common_options()
return KernelTemplate._template_from_string(self.DECLARE_KERNEL).render(options)
def get_kernel_extra_args_declare(self) -> str:
return ""
def get_kernel_extra_args(self) -> str:
return ""
def codegen_define(self, kernel: CppTemplateKernel) -> str:
raise NotImplementedError
def codegen_call(
self,
kernel: CppTemplateKernel,
A: ir.Buffer,
B: ir.Buffer,
C: ir.Buffer,
accum: bool,
) -> str:
"""
Generate the code for calling the templated kernel that computes
`C += alpha * A @ B` if `accum` is True, or `C = alpha * A @ B` otherwise.
"""
A_ptr = f"&({kernel.index(A, [0, 0])})"
B_ptr = f"&({kernel.index(B, [0, 0])})"
C_ptr = f"&({kernel.index(C, [0, 0])})"
M = kernel.size(C, 0)
N = kernel.size(C, 1)
K = kernel.size(A, 1)
lda = kernel.stride(A, 0)
ldb = kernel.stride(B, 0)
ldc = kernel.stride(C, 0)
res = IndentedBuffer()
res.writeline(f"{self.name}<{value_to_cpp(accum, 'bool')}>(")
with res.indent():
extra_args = self.get_kernel_extra_args()
if extra_args:
res.writeline(extra_args)
res.writeline(f"{A_ptr},")
res.writeline(f"{B_ptr},")
res.writeline(f"{C_ptr},")
res.writeline(f"{M},")
res.writeline(f"{N},")
res.writeline(f"{K},")
res.writeline(f"{lda},")
res.writeline(f"{ldb},")
res.writeline(f"{ldc}")
res.writeline(");")
return res.getvalue()
def codegen_init(
self,
kernel: CppTemplateKernel,
) -> str:
return ""
def codegen_finalize(
self,
kernel: CppTemplateKernel,
) -> str:
return ""
def get_b_layout(self) -> LayoutType:
return LayoutType.NORMAL
@dataclasses.dataclass
class CppMicroGemmConfig:
input_dtype: torch.dtype
input2_dtype: torch.dtype
output_dtype: torch.dtype
compute_dtype: torch.dtype
vec_isa_cls: Type[VecISA]
register_blocking: GemmBlocking
extra_check: Optional[Callable[..., bool]] = None
micro_gemm_configs: Dict[Type[CppMicroGemm], List[CppMicroGemmConfig]] = {}
def register_micro_gemm(*configs):
def inner(cls):
assert (
cls not in micro_gemm_configs
), f"Duplicate micro_gemm registration for {cls}"
assert len(configs) > 0, f"No micro_gemm configs provided for {cls}"
micro_gemm_configs[cls] = list(configs)
return cls
return inner
def generate_gemm_config(
vec_isa_cls,
register_blockings,
input_dtype=torch.float,
input2_dtype=None,
output_dtype=None,
compute_dtype=None,
extra_check=None,
):
if output_dtype is None:
output_dtype = input_dtype
if compute_dtype is None:
compute_dtype = output_dtype
if input2_dtype is None:
input2_dtype = input_dtype
return [
CppMicroGemmConfig(
input_dtype,
input2_dtype,
output_dtype,
compute_dtype,
vec_isa_cls,
GemmBlocking(*blocking),
extra_check,
)
for blocking in register_blockings
]
class CppMicroGemmRef(CppMicroGemm):
"""
A reference implementation of the CppMicroGemm class with naive C++ code.
It is used for correctness debugging.
"""
TEMPLATE_ENTRY = r"""
{{declare_kernel}} {
for (int64_t m = 0; m < M; ++m) {
for (int64_t n = 0; n < N; ++n) {
{{compute_t}} result = accum ? C[m * ldc + n] : 0;
for (int64_t k = 0; k < K; ++k) {
result += ({{compute_t}})A[m * lda + k] * ({{compute_t}})B[k * ldb + n] * {{alpha}};
}
C[m * ldc + n] = result;
}
}
}
"""
def __init__(
self, name, input_dtype, input2_dtype, output_dtype, compute_dtype, alpha
) -> None:
super().__init__(
name,
input_dtype,
input2_dtype,
output_dtype,
compute_dtype,
GemmBlocking(1, 1, 1),
alpha,
)
def codegen_define(self, kernel: CppTemplateKernel) -> str:
options = {
"declare_kernel": self.get_kernel_declaration(),
**self.get_common_options(),
}
return KernelTemplate._template_from_string(self.TEMPLATE_ENTRY).render(options)
@register_micro_gemm(
*generate_gemm_config(
VecAVX512,
[(8, 48, 1), (8, 32, 1), (16, 16, 1)],
input_dtype=torch.float,
),
*generate_gemm_config(
VecAVX512,
[(8, 48, 1), (8, 32, 1), (16, 16, 1)],
input_dtype=torch.bfloat16,
output_dtype=torch.float,
),
*generate_gemm_config(
VecAVX512,
[(8, 48, 1), (8, 32, 1), (16, 16, 1)],
input_dtype=torch.half,
output_dtype=torch.float,
),
*generate_gemm_config(
VecAVX512,
[(8, 48, 1), (8, 32, 1), (16, 16, 1)],
input_dtype=torch.bfloat16,
input2_dtype=torch.int8,
output_dtype=torch.float,
compute_dtype=torch.float,
),
*generate_gemm_config(
VecAVX2,
[(4, 24, 1), (4, 16, 1), (8, 8, 1)],
input_dtype=torch.float,
),
*generate_gemm_config(
VecAVX2,
[(4, 24, 1), (4, 16, 1), (8, 8, 1)],
input_dtype=torch.bfloat16,
output_dtype=torch.float,
),
*generate_gemm_config(
VecAVX2,
[(4, 24, 1), (4, 16, 1), (8, 8, 1)],
input_dtype=torch.half,
output_dtype=torch.float,
),
*generate_gemm_config(
VecAVX2,
[(4, 24, 1), (4, 16, 1), (8, 8, 1)],
input_dtype=torch.bfloat16,
input2_dtype=torch.int8,
output_dtype=torch.float,
compute_dtype=torch.float,
),
)
class CppMicroGemmFP32Vec(CppMicroGemm):
"""
This class generates the code for micro gemm using fp32 vec instructions for compute.
It supports input types of torch.float, torch.bfloat16, and torch.half with fp32 output.
The output of the microkernel is in FP32, but it would be converted to BF16/FP16 in the template,
if the desired output is BF16/FP16.
"""
TEMPLATE_ENTRY = r"""
{{declare_kernel}} {
{{kernel.assert_function}}(N % {{block_n}} == 0, "N dimension must be multiple of {{block_n}}");
{{kernel.assert_function}}(K % {{block_k}} == 0, "K dimension must be multiple of {{block_k}}");
// TODO(jgong5): loop unroll for M and N
for (int64_t m = 0; m < M; m += {{block_m}}) {
int64_t block_m = std::min<int64_t>(M - m, {{block_m}});
for (int64_t n = 0; n < N; n += {{block_n}}) {
if (block_m == {{block_m}}) {
{{kernel_name}}_kernel<{{block_m}}, {{block_n}}, accum>(
A + m * lda,
B + n,
C + m * ldc + n,
K,
lda,
ldb,
ldc
);
} else {
switch (block_m) {
{%- for b in range(block_m - 1, 0, -1) %}
case {{b}}:
{{kernel_name}}_kernel<{{b}}, {{block_n}}, accum>(
A + m * lda,
B + n,
C + m * ldc + n,
K,
lda,
ldb,
ldc
);
break;
{%- endfor %}
default:
{{kernel.assert_function}}(false, "Unsupported block_m: {{block_m}}");
}
}
}
}
}
"""
TEMPLATE_KERNEL = r"""
template <int64_t BLOCK_M, int64_t BLOCK_N, bool accum>
inline void {{kernel_name}}_kernel(
const {{input_t}}* {{restrict_keyword}} A,
const {{input2_t}}* {{restrict_keyword}} B,
{{output_t}}* {{restrict_keyword}} C,
int64_t K,
int64_t lda,
int64_t ldb,
int64_t ldc
) {
using Vectorized = at::vec::Vectorized<{{compute_t}}>;
using VectorizedIn = at::vec::Vectorized<{{input_t}}>;
constexpr auto VLEN = Vectorized::size();
constexpr auto ROWS = BLOCK_M;
constexpr auto COLS = BLOCK_N / VLEN;
Vectorized va;
at::vec::VectorizedN<{{compute_t}}, COLS> vb;
at::vec::VectorizedN<{{compute_t}}, ROWS*COLS> vc;
auto loadc = [&](auto i) {
if constexpr (accum) {
constexpr int row = i / COLS;
constexpr int col = i % COLS;
vc[i] = Vectorized::loadu(C + row * ldc + col * VLEN);
} else {
vc[i] = Vectorized(0.0f);
}
};
c10::ForcedUnroll<ROWS * COLS>{}(loadc);
auto compute = [&, COLS](auto i, int k) {
constexpr int row = i / COLS;
constexpr int col = i % COLS;
if constexpr (col == 0) {
{%- if alpha != 1 %}
va = Vectorized(static_cast<{{compute_t}}>(A[row * lda + k]) * {{alpha}});
{%- else %}
va = Vectorized(static_cast<{{compute_t}}>(A[row * lda + k]));
{%- endif %}
}
if constexpr (row == 0) {
{%- if input2_dtype in [torch.bfloat16, torch.float16] %}
auto b = VectorizedIn::loadu(B + k * ldb + col * VLEN, VLEN);
vb[col] = at::vec::convert<{{compute_t}}>(b);
{%- elif input2_dtype == torch.int8 %}
// Convert VLEN int8 elements to int32, and then fp32
auto b32 = at::vec::convert_to_int32<int8_t>(B + k * ldb + col * VLEN);
vb[col] = at::vec::convert<float>(b32);
{%- else %}
vb[col] = Vectorized::loadu(B + k * ldb + col * VLEN);
{%- endif %}
}
constexpr int idx = row * COLS + col;
vc[idx] = at::vec::fmadd(va, vb[col], vc[idx]);
};
for (int k = 0; k < K; ++k) {
c10::ForcedUnroll<ROWS * COLS>{}(compute, k);
}
// store to C
auto storec = [&](auto i) {
constexpr int row = i / COLS;
constexpr int col = i % COLS;
vc[i].store(C + row * ldc + col * VLEN);
};
c10::ForcedUnroll<ROWS * COLS>{}(storec);
}
"""
def codegen_define(self, kernel: CppTemplateKernel) -> str:
options = {
"declare_kernel": self.get_kernel_declaration(),
"kernel": kernel,
"block_m": self.register_blocking.block_m,
"block_n": self.register_blocking.block_n,
"block_k": self.register_blocking.block_k,
"restrict_keyword": get_restrict_keyword(),
**self.get_common_options(),
}
result = KernelTemplate._template_from_string(self.TEMPLATE_KERNEL).render(
options
)
result += KernelTemplate._template_from_string(self.TEMPLATE_ENTRY).render(
options
)
return result
# extra check for CppMicroGemmAMX
def check_amx_extra(config, m, n, k, alpha, num_threads):
vnni_size = 4 if config.input_dtype == torch.uint8 else 2
return k % vnni_size == 0 and alpha == 1
@register_micro_gemm(
*generate_gemm_config(
VecAMX,
[(32, 32, 32), (48, 16, 32), (16, 48, 32)],
input_dtype=torch.bfloat16,
input2_dtype=torch.int8,
output_dtype=torch.float,
compute_dtype=torch.float,
extra_check=check_amx_extra,
),
*generate_gemm_config(
VecAMX,
[(32, 32, 32), (48, 16, 32), (16, 48, 32)],
input_dtype=torch.bfloat16,
output_dtype=torch.float,
extra_check=check_amx_extra,
),
*generate_gemm_config(
VecAMX,
[(32, 32, 64), (48, 16, 64)],
input_dtype=torch.uint8,
input2_dtype=torch.int8,
output_dtype=torch.int32,
compute_dtype=torch.int32,
extra_check=check_amx_extra,
),
)
class CppMicroGemmAMX(CppMicroGemm):
"""
This class generates the code for micro gemm using Advanced Matrix eXtention (AMX)
instructions available in 4th generation Intel Xeon for compute.
It supports input types of torch.bfloat16 with fp32 output.
"""
TEMPLATE_ENTRY = r"""
{{declare_kernel}} {
{{kernel.assert_function}}(N % {{block_n}} == 0, "N dimension must be multiple of {{block_n}}");
{{kernel.assert_function}}(K % 2 == 0, "K dimension must be multiple of 2");
{%- if use_cached_dequantized_B %}
// Create a stack-allocated buffer for tiles of B.
// Except maybe for the tail-case, an AMX tile of B has 16x32 BF16 elements.
// we cache K * {{block_n}} elements of dequantized B
const auto buf_size = K * {{block_n}};
{%- if is_msvc_compiler %}
// MSVC doesn't support stack-allocated dynamic-sized arrays, so using heap memory here.
std::unique_ptr<{{input_t}}[]> heap_deq_b_buf_ptr(new {{input_t}}[buf_size]);
{{input_t}}* dequantized_B_buf = heap_deq_b_buf_ptr.get();
{%- else %}
// It's safe to use a stack-allocated array since the blocking strategy would
// require us to allocate an array that's smaller than the size of L1D cache,
// and the default per thread max stack size on Linux is quite higher,
// so we need not worry about stack overflow.
alignas(4096) {{input_t}} dequantized_B_buf[buf_size];
{%- endif %}
auto load_dequantized_B = [&](int base_idx) {
// Load a tile of B & cache it in L1D.
{{input2_t}}* base_addr = const_cast<{{input2_t}}*>(B) + base_idx;
for (int idx_dq = 0, idx_q = 0; idx_dq < buf_size; idx_q += ldb, idx_dq += {{block_n}}) {
{%- for vec_idx in range(0, block_n - 1, 32) %}
auto b_int8 = at::vec::Vectorized<int8_t>::loadu(
base_addr + idx_q + {{vec_idx}} ,
static_cast<int64_t>(32)
);
auto b_bf16 = at::vec::convert<{{input_t}}>(b_int8);
b_bf16.store(dequantized_B_buf + idx_dq + {{vec_idx}});
{%- endfor %}
{%- if (block_n % 32) != 0 %}
auto b_int8_tail = at::vec::Vectorized<int8_t>::loadu(
base_addr + idx_q + {{block_n - (block_n % 32)}},
static_cast<int64_t>({{block_n % 32}})
);
auto b_bf16_tail = at::vec::convert<{{input_t}}>(b_int8_tail);
b_bf16_tail.store(
dequantized_B_buf + idx_dq + {{block_n - (block_n % 32)}},
static_cast<int64_t>({{block_n % 32}})
);
{%- endif %}
}
};
{%- endif %}
// The ldb would not be block_n if N != block_n
{%- if use_cached_dequantized_B %}
const int64_t updated_ldb = {{block_n}};
{%- else %}
const int64_t updated_ldb = ldb;
{%- endif %}
// TODO(jgong5): loop unroll for M and N
for (int64_t n = 0; n < N; n += {{block_n}}) {
{%- if use_cached_dequantized_B %}
// Dequantize K * block_n int8 B elements into BF16
load_dequantized_B(n);
{%- endif %}
for (int64_t m = 0; m < M; m += {{block_m}}) {
int64_t block_m = std::min<int64_t>(M - m, {{block_m}});
int64_t m_tail = m;
{%- for num_rows in range(block_m, 0, -16) %}
{%- if num_rows != block_m %}
else
{%- endif %}
if (block_m >= {{num_rows}}) {
{{kernel_name}}_amx_kernel_{{num_rows}}_{{num_columns}}<accum>(
amx_state,
A + m * lda,
{%- if use_cached_dequantized_B %}
dequantized_B_buf,
{%- else %}
B + n,
{%- endif %}
C + m * ldc + n,
K,
lda,
updated_ldb,
ldc,
16
);
block_m -= {{num_rows}};
m_tail += {{num_rows}};
}
{%- endfor %}
if (block_m > 0) {
{{kernel_name}}_amx_kernel_16_{{num_columns}}<accum>(
amx_state,
A + m_tail * lda,
{%- if use_cached_dequantized_B %}
dequantized_B_buf,
{%- else %}
B + n,
{%- endif %}
C + m_tail * ldc + n,
K,
lda,
updated_ldb,
ldc,
block_m
);
}
}
}
}
"""
TEMPLATE_KERNEL = r"""
template <bool accum>
inline void {{kernel_name}}_amx_kernel_{{num_rows}}_{{num_columns}}(
AMXState& amx_state,
const {{input_t}}* {{restrict_keyword}} A,
{%- if use_cached_dequantized_B %}
const {{input_t}}* {{restrict_keyword}} B,
{%- else %}
const {{input2_t}}* {{restrict_keyword}} B,
{%- endif %}
{{output_t}}* {{restrict_keyword}} C,
int64_t K,
int64_t lda,
int64_t ldb,
int64_t ldc,
uint8_t tilecfg_rows
) {
// TODO(jgong5): add prefetch hint for A, B, C
auto loadconfig = [](const amx_tilecfg& cfg) {
_tile_loadconfig(&cfg);
};
const auto last_k_offset = K / {{block_k}} * {{block_k}};
const auto tail_k_size = K - last_k_offset;
if C10_LIKELY (last_k_offset > 0) {
amx_state.configure(tilecfg_rows, 64, {{num_rows}} / 16, {{num_columns}}, loadconfig);
} else {
amx_state.configure(tilecfg_rows, tail_k_size * sizeof({{input_t}}), {{num_rows}} / 16, {{num_columns}}, loadconfig);
}
auto load_c = [&]() {
{%- for tile_row in range(num_rows // 16) %}
{%- for tile_col in range(num_columns) %}
{%- set tile_idx = tile_row * num_columns + tile_col %}
_tile_loadd({{tile_idx}}, C + {{tile_row * 16}} * ldc + {{tile_col * 16}}, ldc * sizeof({{output_t}}));
{%- endfor %}
{%- endfor %}
};
auto zero_c = [&]() {
{%- for tile_row in range(num_rows // 16) %}
{%- for tile_col in range(num_columns) %}
{%- set tile_idx = tile_row * num_columns + tile_col %}
_tile_zero({{tile_idx}});
{%- endfor %}
{%- endfor %}
};
if constexpr (accum) {
load_c();
} else {
zero_c();
}
auto compute = [&](int k) {
{%- set tile_offset_a = num_rows // 16 * num_columns %}
{%- set tile_offset_b = tile_offset_a + num_rows // 16 %}
{%- for tile_row in range(num_rows // 16) %}
{%- for tile_col in range(num_columns) %}
{%- set tile_idx_a = tile_offset_a + tile_row %}
{%- set tile_idx_b = tile_offset_b + tile_col %}
{%- set tile_idx_c = tile_row * num_columns + tile_col %}
{%- if tile_col == 0 %}
_tile_stream_loadd({{tile_idx_a}}, A + {{tile_row * 16}} * lda + k, lda * sizeof({{input_t}}));
{%- endif %}
{%- if tile_row == 0 %}
_tile_loadd({{tile_idx_b}}, B + k * ldb + {{tile_col * 16 * vnni_size}}, ldb * {{vnni_size}} * sizeof({{input_t}}));
{%- endif %}
{%- if int8_gemm %}
_tile_dpbusd({{tile_idx_c}}, {{tile_idx_a}}, {{tile_idx_b}});
{%- else %}
_tile_dpbf16ps({{tile_idx_c}}, {{tile_idx_a}}, {{tile_idx_b}});
{%- endif %}
{%- endfor %}
{%- endfor %}
};
{{kernel.unroll_pragma(4)}}
for (int k = 0; k < last_k_offset; k += {{block_k}}) {
compute(k);
}
auto store_c = [&]() {
// store to C
{%- for tile_row in range(num_rows // 16) %}
{%- for tile_col in range(num_columns) %}
{%- set tile_idx = tile_row * num_columns + tile_col %}
_tile_stored({{tile_idx}}, C + {{tile_row * 16}} * ldc + {{tile_col * 16}}, ldc * sizeof({{output_t}}));
{%- endfor %}
{%- endfor %}
};
// TODO(jgong5): move tail k computation to separate loopnest to save tile configuration overhead
if C10_UNLIKELY (tail_k_size > 0) {
if C10_LIKELY (last_k_offset > 0) {
store_c();
amx_state.configure(tilecfg_rows, tail_k_size * sizeof({{input_t}}), {{num_rows}} / 16, {{num_columns}}, loadconfig);
load_c();
}
compute(last_k_offset);
}
store_c();
}
"""
def codegen_define(self, kernel: CppTemplateKernel) -> str:
block_m, block_n, block_k = self.register_blocking
assert block_m % 16 == 0, "Only support block_m % 16 == 0 for AMX"
assert block_n % 16 == 0, "Only support block_n % 16 == 0 for AMX"
if self.input_dtype == torch.uint8:
assert block_k == 64, "Only support block_k = 64 for AMX INT8"
else:
assert block_k == 32, "Only support block_k = 32 for AMX Bfloat16/Float16"
num_columns = block_n // 16
options = {
"declare_kernel": self.get_kernel_declaration(),
"use_cached_dequantized_B": self.input_dtype == torch.bfloat16
and self.input2_dtype == torch.int8,
"kernel": kernel,
"block_m": block_m,
"block_n": block_n,
"block_k": block_k,
"num_columns": num_columns,
"restrict_keyword": get_restrict_keyword(),
**self.get_common_options(),
}
result = ""
for num_rows in range(block_m, 0, -16):
amx_kernel_options = {**options, "num_rows": num_rows}
result += KernelTemplate._template_from_string(self.TEMPLATE_KERNEL).render(
amx_kernel_options
)
result += KernelTemplate._template_from_string(self.TEMPLATE_ENTRY).render(
options
)
return result
def codegen_init(
self,
kernel: CppTemplateKernel,
) -> str:
return "AMXState amx_state;"
def codegen_finalize(
self,
kernel: CppTemplateKernel,
) -> str:
return "amx_state.release([]() { _tile_release(); });"
def get_kernel_extra_args_declare(self) -> str:
return "AMXState& amx_state,"
def get_kernel_extra_args(self) -> str:
return "amx_state,"
def get_b_layout(self):
if self.input_dtype == torch.uint8:
return LayoutType.VNNI4
else:
return LayoutType.VNNI2
# extra check for CppMicroBrgemm
def check_brgemm_extra(config, m, n, k, alpha, num_threads):
assert config.input_dtype == torch.half and config.output_dtype == torch.float
vnni_size = 2
# use brgemm for Half when amx_fp16 is supported
return torch.cpu._is_amx_fp16_supported() and k % vnni_size == 0 and alpha == 1
@register_micro_gemm(
*generate_gemm_config(
VecAMX,
[(32, 32, 32), (48, 16, 32), (16, 48, 32)],
input_dtype=torch.half,
output_dtype=torch.float,
extra_check=check_brgemm_extra,
),
)
class CppMicroBrgemm(CppMicroGemm):
"""
This class generates the code for micro gemm using oneDNN brgemm.
It supports input types of torch.half.
"""
TEMPLATE_ENTRY = r"""
#include <ATen/native/CPUBlas.h>
{{declare_kernel}} {
at::native::cpublas::brgemm(
M, N, K,
lda, ldb, ldc,
accum,
A,
B,
C);
}
"""
def codegen_define(self, kernel: CppTemplateKernel) -> str:
options = {
"declare_kernel": self.get_kernel_declaration(),
"kernel": kernel,
"block_m": self.register_blocking.block_m,
"block_n": self.register_blocking.block_n,
"block_k": self.register_blocking.block_k,
"restrict_keyword": get_restrict_keyword(),
**self.get_common_options(),
}
result = ""
result += KernelTemplate._template_from_string(self.TEMPLATE_ENTRY).render(
options
)
return result
def codegen_finalize(
self,
kernel: CppTemplateKernel,
) -> str:
return "at::native::cpublas::brgemm_release();"
def get_b_layout(self):
assert self.input_dtype == torch.half and torch.cpu._is_amx_fp16_supported()
return LayoutType.VNNI2
def create_micro_gemm(
name,
m,
n,
k,
input_dtype,
input2_dtype,
output_dtype=None,
compute_dtype=None,
alpha=1,
num_threads=-1,
use_ref=True,
) -> Optional[CppMicroGemm]:
def create_from_config(cls, config: CppMicroGemmConfig):
return cls(
name,
config.input_dtype,
config.input2_dtype,
config.output_dtype,
config.compute_dtype,
config.register_blocking,
alpha,
)
assert isinstance(n, int) or n.is_number, n
assert isinstance(k, int) or k.is_number, k
m = V.graph.sizevars.size_hint(m, fallback=1) if isinstance(m, sympy.Expr) else m
assert isinstance(m, int), m
if output_dtype is None:
output_dtype = input_dtype
if compute_dtype is None:
compute_dtype = output_dtype
if num_threads < 0:
num_threads = parallel_num_threads()
vec_isa = pick_vec_isa()
matched_configs = []
for cls, configs in micro_gemm_configs.items():
for config in configs:
if not issubclass(vec_isa.__class__, config.vec_isa_cls):
continue
if (
config.input_dtype == input_dtype
and config.compute_dtype == compute_dtype
and config.input2_dtype == input2_dtype
and config.output_dtype == output_dtype
# The output_dtype here is the output dtype of the micro-kernel.
# In some cases, the actual output dtype of the op for which the micro-kernel
# is being created would be same as that of the activation, but the micro-kernels
# compute output in Float/int32, which is converted in the GEMM template. This is
# subject to change in the future.
):
if config.extra_check is not None and not config.extra_check(
config, m, n, k, alpha, num_threads
):
continue
block_m, block_n, block_k = config.register_blocking
if (
config.vec_isa_cls == VecAMX
and m < block_m
and input_dtype == torch.bfloat16
and input2_dtype == torch.int8
):
# For int8 WoQ GEMM, AMX micro-kernel may not perform well if m < block_m
continue
# Criteria on the ranking of configurations
# 1. ISA: AMX > VEC
# 2. Dividable by block sizes (block_m, block_n, block_k)
# 3. Number of mxn blocks is large enough to occupy all the threads
# 4. Register blocks are larger
isa_score = 0
if config.vec_isa_cls == VecAMX:
isa_score += 1
dividable_score = 0
if m % block_m == 0:
dividable_score += 1
if n % block_n == 0:
dividable_score += 1
if k % block_k == 0:
dividable_score += 1
occupancy_score = 0
n_blocks = (n + block_n - 1) // block_n
total_mxn_blocks = n_blocks * ((m + block_m - 1) // block_m)
if n_blocks >= num_threads:
occupancy_score += 1
if total_mxn_blocks >= num_threads:
occupancy_score += 1
register_bytes = (
block_m * block_n * config.compute_dtype.itemsize
+ (block_m * block_k + block_k * block_n)
* config.input_dtype.itemsize
)
matched_configs.append(
(
(isa_score, dividable_score, occupancy_score, register_bytes),
cls,
config,
)
)
if len(matched_configs) == 0:
if use_ref:
return CppMicroGemmRef(
name, input_dtype, input2_dtype, output_dtype, compute_dtype, alpha
)
else:
return None
# TODO(jgong5): allow autotuning on choices of configs
return create_from_config(*max(matched_configs, key=lambda x: x[0])[1:])
|