File size: 8,320 Bytes
20347e1 | 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 | // Copyright © 2024 Apple Inc.
#pragma once
#include "gemm/loader.h"
#include "gemm/mma.h"
#include "gemm/params.h"
#include "gemm/transforms.h"
#include "gemm/utils.h"
using namespace metal;
///////////////////////////////////////////////////////////////////////////////
// GEMM kernel class
///////////////////////////////////////////////////////////////////////////////
namespace mlx {
namespace steel {
template <bool M_aligned, bool N_aligned, bool K_aligned>
struct LoopAlignment {};
template <
typename T,
typename U,
int BM,
int BN,
int BK,
int WM,
int WN,
bool transpose_a,
bool transpose_b,
bool MN_aligned,
bool K_aligned,
typename AccumType = typename AccumHelper<T>::accum_type,
typename Epilogue = TransformNone<U, AccumType>>
struct GEMMKernel {
STEEL_CONST short tgp_padding_a = 16 / sizeof(T);
STEEL_CONST short tgp_padding_b = 16 / sizeof(T);
STEEL_CONST short tgp_mem_size_a =
transpose_a ? BK * (BM + tgp_padding_a) : BM * (BK + tgp_padding_a);
STEEL_CONST short tgp_mem_size_b =
transpose_b ? BN * (BK + tgp_padding_b) : BK * (BN + tgp_padding_b);
STEEL_CONST short tgp_mem_size = tgp_mem_size_a + tgp_mem_size_b;
STEEL_CONST short tgp_size = WM * WN * 32;
using loader_a_t = BlockLoader<
T,
transpose_a ? BK : BM,
transpose_a ? BM : BK,
transpose_a ? BM + tgp_padding_a : BK + tgp_padding_a,
!transpose_a,
tgp_size>;
using loader_b_t = BlockLoader<
T,
transpose_b ? BN : BK,
transpose_b ? BK : BN,
transpose_b ? BK + tgp_padding_b : BN + tgp_padding_b,
transpose_b,
tgp_size>;
using mma_t = BlockMMA<
T,
U,
BM,
BN,
BK,
WM,
WN,
transpose_a,
transpose_b,
transpose_a ? BM + tgp_padding_a : BK + tgp_padding_a,
transpose_b ? BK + tgp_padding_b : BN + tgp_padding_b,
AccumType,
Epilogue>;
/* Main kernel function */
template <bool M_aligned, bool N_aligned, bool K_aligned_>
static METAL_FUNC void gemm_loop(
threadgroup T* As [[threadgroup(0)]],
threadgroup T* Bs [[threadgroup(1)]],
const int gemm_k_iterations,
thread loader_a_t& loader_a,
thread loader_b_t& loader_b,
thread mma_t& mma_op,
thread const short& tgp_bm,
thread const short& tgp_bn,
thread const short& lbk,
LoopAlignment<M_aligned, N_aligned, K_aligned_> l = {}) {
// Appease the compiler
(void)l;
short2 tile_dims_A = transpose_a ? short2(tgp_bm, BK) : short2(BK, tgp_bm);
short2 tile_dims_B = transpose_b ? short2(BK, tgp_bn) : short2(tgp_bn, BK);
for (int k = 0; k < gemm_k_iterations; k++) {
threadgroup_barrier(mem_flags::mem_threadgroup);
// Load elements into threadgroup
if (M_aligned) {
loader_a.load_unsafe();
} else {
loader_a.load_safe(tile_dims_A);
}
if (N_aligned) {
loader_b.load_unsafe();
} else {
loader_b.load_safe(tile_dims_B);
}
threadgroup_barrier(mem_flags::mem_threadgroup);
// Multiply and accumulate threadgroup elements
mma_op.mma(As, Bs);
// Prepare for next iteration
loader_a.next();
loader_b.next();
}
if (!K_aligned_) {
threadgroup_barrier(mem_flags::mem_threadgroup);
short2 tile_dims_A_last =
transpose_a ? short2(tgp_bm, lbk) : short2(lbk, tgp_bm);
short2 tile_dims_B_last =
transpose_b ? short2(lbk, tgp_bn) : short2(tgp_bn, lbk);
loader_a.load_safe(tile_dims_A_last);
loader_b.load_safe(tile_dims_B_last);
threadgroup_barrier(mem_flags::mem_threadgroup);
mma_op.mma(As, Bs);
}
}
/* Main kernel function */
static METAL_FUNC void run(
const device T* A [[buffer(0)]],
const device T* B [[buffer(1)]],
device U* D [[buffer(2)]],
const constant GEMMParams* params [[buffer(3)]],
threadgroup T* As [[threadgroup(0)]],
threadgroup T* Bs [[threadgroup(1)]],
uint simd_lane_id [[thread_index_in_simdgroup]],
uint simd_group_id [[simdgroup_index_in_threadgroup]],
uint3 tid [[threadgroup_position_in_grid]],
uint3 lid [[thread_position_in_threadgroup]]) {
// Pacifying compiler
(void)lid;
const int tid_y = ((tid.y) << params->swizzle_log) +
((tid.x) & ((1 << params->swizzle_log) - 1));
const int tid_x = (tid.x) >> params->swizzle_log;
if (params->tiles_n <= tid_x || params->tiles_m <= tid_y) {
return;
}
threadgroup_barrier(mem_flags::mem_none);
// Find block in A, B, C
const int c_row = tid_y * BM;
const int c_col = tid_x * BN;
const size_t c_row_long = size_t(c_row);
const size_t c_col_long = size_t(c_col);
A += transpose_a ? c_row_long : c_row_long * params->lda;
B += transpose_b ? c_col_long * params->ldb : c_col_long;
D += c_row_long * params->ldd + c_col_long;
// Prepare threadgroup loading operations
thread loader_a_t loader_a(A, params->lda, As, simd_group_id, simd_lane_id);
thread loader_b_t loader_b(B, params->ldb, Bs, simd_group_id, simd_lane_id);
// Prepare threadgroup mma operation
thread mma_t mma_op(simd_group_id, simd_lane_id);
int gemm_k_iterations = params->gemm_k_iterations_aligned;
///////////////////////////////////////////////////////////////////////////////
// MNK aligned loop
if (MN_aligned) {
for (int k = 0; k < gemm_k_iterations; k++) {
threadgroup_barrier(mem_flags::mem_threadgroup);
// Load elements into threadgroup
loader_a.load_unsafe();
loader_b.load_unsafe();
threadgroup_barrier(mem_flags::mem_threadgroup);
// Multiply and accumulate threadgroup elements
mma_op.mma(As, Bs);
// Prepare for next iteration
loader_a.next();
loader_b.next();
}
threadgroup_barrier(mem_flags::mem_none);
// Loop tail
if (!K_aligned) {
int lbk = params->K - params->gemm_k_iterations_aligned * BK;
short2 tile_dims_A = transpose_a ? short2(BM, lbk) : short2(lbk, BM);
short2 tile_dims_B = transpose_b ? short2(lbk, BN) : short2(BN, lbk);
loader_a.load_safe(tile_dims_A);
loader_b.load_safe(tile_dims_B);
threadgroup_barrier(mem_flags::mem_threadgroup);
mma_op.mma(As, Bs);
}
// Store results to device memory
mma_op.store_result(D, params->ldd);
return;
}
///////////////////////////////////////////////////////////////////////////////
// MN unaligned loop
else { // Loop over K - unaligned case
short tgp_bm = min(BM, params->M - c_row);
short tgp_bn = min(BN, params->N - c_col);
short leftover_bk = params->K - params->gemm_k_iterations_aligned * BK;
if (tgp_bm == BM && tgp_bn == BN) {
gemm_loop<true, true, K_aligned>(
As,
Bs,
gemm_k_iterations,
loader_a,
loader_b,
mma_op,
tgp_bm,
tgp_bn,
leftover_bk);
mma_op.store_result(D, params->ldd);
return;
} else if (tgp_bn == BN) {
gemm_loop<false, true, K_aligned>(
As,
Bs,
gemm_k_iterations,
loader_a,
loader_b,
mma_op,
tgp_bm,
tgp_bn,
leftover_bk);
mma_op.store_result_safe(D, params->ldd, short2(tgp_bn, tgp_bm));
return;
} else if (tgp_bm == BM) {
gemm_loop<true, false, K_aligned>(
As,
Bs,
gemm_k_iterations,
loader_a,
loader_b,
mma_op,
tgp_bm,
tgp_bn,
leftover_bk);
mma_op.store_result_safe(D, params->ldd, short2(tgp_bn, tgp_bm));
return;
} else {
gemm_loop<false, false, K_aligned>(
As,
Bs,
gemm_k_iterations,
loader_a,
loader_b,
mma_op,
tgp_bm,
tgp_bn,
leftover_bk);
mma_op.store_result_safe(D, params->ldd, short2(tgp_bn, tgp_bm));
return;
}
}
}
};
} // namespace steel
} // namespace mlx |