File size: 10,937 Bytes
36457f1 | 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 | """
Kernel descriptions: Pre-built dependence graphs for common GPU kernels.
These correspond to the kernels evaluated in Section 6 of the paper.
"""
import numpy as np
from twill.graph import (
DependenceGraph, Instruction, DependenceEdge, MachineDescription,
hopper_machine, blackwell_machine, make_rrt,
)
def flash_attention_forward_simplified(
machine: MachineDescription = None,
gemm_cycles: int = 1,
exp_cycles: int = 1,
) -> DependenceGraph:
"""Simplified Flash Attention forward pass (Figure 1 of the paper).
The simplified loop body:
S = gemm(Q, K[i]) # uses TC
P = exp(S) # uses EXP
O += gemm(P, V[i]) # uses TC
This is the running example from Section 3.
On Hopper, GEMM and EXP on a tile have roughly the same cost.
Args:
machine: Target machine (default: Hopper)
gemm_cycles: Normalized cycle count for GEMM
exp_cycles: Normalized cycle count for EXP
Returns:
DependenceGraph ready for Twill
"""
if machine is None:
machine = hopper_machine()
graph = DependenceGraph(machine)
num_fus = machine.num_functional_units
tc_idx = machine.fu_index("TC")
exp_idx = machine.fu_index("EXP")
# S = gemm(Q, K[i]) - uses TC for gemm_cycles
rrt_S = make_rrt(gemm_cycles, {tc_idx: [1] * gemm_cycles}, num_fus)
graph.add_instruction(Instruction("S", rrt_S))
# P = exp(S) - uses EXP for exp_cycles
rrt_P = make_rrt(exp_cycles, {exp_idx: [1] * exp_cycles}, num_fus)
graph.add_instruction(Instruction("P", rrt_P))
# O = gemm(P, V[i]) - uses TC for gemm_cycles
rrt_O = make_rrt(gemm_cycles, {tc_idx: [1] * gemm_cycles}, num_fus)
graph.add_instruction(Instruction("O", rrt_O))
# Dependencies (Figure 1c):
# S -> P (d = gemm_cycles, δ = 0) - P must wait for GEMM to finish
graph.add_edge(DependenceEdge("S", "P", delay=gemm_cycles, iteration_delay=0))
# P -> O (d = exp_cycles, δ = 0) - O must wait for EXP to finish
graph.add_edge(DependenceEdge("P", "O", delay=exp_cycles, iteration_delay=0))
# O -> O (d = gemm_cycles, δ = 1) - loop-carried: O accumulates across iterations
graph.add_edge(DependenceEdge("O", "O", delay=gemm_cycles, iteration_delay=1))
return graph
def flash_attention_forward_hopper(
tma_cycles: int = 1,
gemm_cycles: int = 2,
exp_cycles: int = 2,
) -> DependenceGraph:
"""Full FMHA forward pass on Hopper (Section 6.2.1).
Includes TMA loads for K and V tiles:
K_load = tma_load(K[i]) # uses TMA (variable latency)
V_load = tma_load(V[i]) # uses TMA (variable latency)
S = wgmma(Q, K_load) # uses TC
P = exp(S) # uses EXP
O += wgmma(P, V_load) # uses TC
FA3 discovered: SWP extracts S=gemm(Q,K[0]) into prologue,
and ping-pong scheduling alternates EXP and TC across warp groups.
"""
machine = hopper_machine()
graph = DependenceGraph(machine)
num_fus = machine.num_functional_units
tc_idx = machine.fu_index("TC")
exp_idx = machine.fu_index("EXP")
tma_idx = machine.fu_index("TMA")
# TMA loads (variable latency, streaming)
rrt_K_load = make_rrt(tma_cycles, {tma_idx: [1] * tma_cycles}, num_fus)
graph.add_instruction(Instruction(
"K_load", rrt_K_load, variable_latency=True, streaming=True,
memory_footprint={"SMEM": 128 * 128 * 2} # FP16 tile
))
rrt_V_load = make_rrt(tma_cycles, {tma_idx: [1] * tma_cycles}, num_fus)
graph.add_instruction(Instruction(
"V_load", rrt_V_load, variable_latency=True, streaming=True,
memory_footprint={"SMEM": 128 * 128 * 2}
))
# S = wgmma(Q, K_load)
rrt_S = make_rrt(gemm_cycles, {tc_idx: [1] * gemm_cycles}, num_fus)
graph.add_instruction(Instruction("S", rrt_S))
# P = exp(S) - softmax rescaling
rrt_P = make_rrt(exp_cycles, {exp_idx: [1] * exp_cycles}, num_fus)
graph.add_instruction(Instruction("P", rrt_P))
# O += wgmma(P, V_load) - accumulate
rrt_O = make_rrt(gemm_cycles, {tc_idx: [1] * gemm_cycles}, num_fus)
graph.add_instruction(Instruction("O", rrt_O))
# Dependencies:
# K_load -> S (TMA must complete before GEMM can consume K)
graph.add_edge(DependenceEdge("K_load", "S", delay=tma_cycles, iteration_delay=0))
# V_load -> O (TMA must complete before GEMM can consume V)
graph.add_edge(DependenceEdge("V_load", "O", delay=tma_cycles, iteration_delay=0))
# S -> P (GEMM result needed for softmax)
graph.add_edge(DependenceEdge("S", "P", delay=gemm_cycles, iteration_delay=0))
# P -> O (softmax result needed for second GEMM)
graph.add_edge(DependenceEdge("P", "O", delay=exp_cycles, iteration_delay=0))
# O -> O (loop carried: accumulation)
graph.add_edge(DependenceEdge("O", "O", delay=gemm_cycles, iteration_delay=1))
return graph
def flash_attention_forward_blackwell(
tma_cycles: int = 1,
gemm_cycles: int = 1, # TC 2x faster on Blackwell
exp_cycles: int = 2, # EXP unchanged
tmem_cycles: int = 1,
) -> DependenceGraph:
"""FMHA forward pass on Blackwell (Section 6.2.2).
Blackwell differences:
- TC throughput 2x Hopper -> GEMM takes fewer relative cycles
- Tensor Memory (TMEM) tier for TC accumulators
- Explicit register <-> TMEM transfers needed
- EXP/softmax becomes the bottleneck (doesn't scale as fast)
FA4 strategy (rediscovered by Twill):
- TMA loads on variable-latency warp (producer)
- TC GEMMs on compute warps
- Specific cross-warp communication for TMEM
"""
machine = blackwell_machine()
graph = DependenceGraph(machine)
num_fus = machine.num_functional_units
tc_idx = machine.fu_index("TC")
exp_idx = machine.fu_index("EXP")
tma_idx = machine.fu_index("TMA")
tmem_idx = machine.fu_index("TMEM")
# TMA loads (variable latency, streaming)
rrt_K_load = make_rrt(tma_cycles, {tma_idx: [1] * tma_cycles}, num_fus)
graph.add_instruction(Instruction(
"K_load", rrt_K_load, variable_latency=True, streaming=True,
memory_footprint={"SMEM": 128 * 128 * 2}
))
rrt_V_load = make_rrt(tma_cycles, {tma_idx: [1] * tma_cycles}, num_fus)
graph.add_instruction(Instruction(
"V_load", rrt_V_load, variable_latency=True, streaming=True,
memory_footprint={"SMEM": 128 * 128 * 2}
))
# S = wgmma(Q, K_load) - output goes to TMEM
rrt_S = make_rrt(gemm_cycles, {tc_idx: [1] * gemm_cycles}, num_fus)
graph.add_instruction(Instruction("S", rrt_S, memory_footprint={"TMEM": 128 * 128 * 4}))
# S_read: TMEM -> register transfer
rrt_S_read = make_rrt(tmem_cycles, {tmem_idx: [1] * tmem_cycles}, num_fus)
graph.add_instruction(Instruction("S_read", rrt_S_read))
# P = exp(S_read) - softmax in registers
rrt_P = make_rrt(exp_cycles, {exp_idx: [1] * exp_cycles}, num_fus)
graph.add_instruction(Instruction("P", rrt_P))
# P_write: register -> TMEM for second GEMM
rrt_P_write = make_rrt(tmem_cycles, {tmem_idx: [1] * tmem_cycles}, num_fus)
graph.add_instruction(Instruction("P_write", rrt_P_write))
# O += wgmma(P_write, V_load) - accumulate in TMEM
rrt_O = make_rrt(gemm_cycles, {tc_idx: [1] * gemm_cycles}, num_fus)
graph.add_instruction(Instruction("O", rrt_O))
# Dependencies:
graph.add_edge(DependenceEdge("K_load", "S", delay=tma_cycles, iteration_delay=0))
graph.add_edge(DependenceEdge("S", "S_read", delay=gemm_cycles, iteration_delay=0))
graph.add_edge(DependenceEdge("S_read", "P", delay=tmem_cycles, iteration_delay=0))
graph.add_edge(DependenceEdge("P", "P_write", delay=exp_cycles, iteration_delay=0))
graph.add_edge(DependenceEdge("P_write", "O", delay=tmem_cycles, iteration_delay=0))
graph.add_edge(DependenceEdge("V_load", "O", delay=tma_cycles, iteration_delay=0))
graph.add_edge(DependenceEdge("O", "O", delay=gemm_cycles, iteration_delay=1))
return graph
def simple_gemm_pipeline(
machine: MachineDescription = None,
load_cycles: int = 1,
compute_cycles: int = 2,
) -> DependenceGraph:
"""Simple GEMM with load-compute overlap.
The simplest pipelining case:
A_load = tma_load(A[i])
B_load = tma_load(B[i])
C += gemm(A_load, B_load)
"""
if machine is None:
machine = hopper_machine()
graph = DependenceGraph(machine)
num_fus = machine.num_functional_units
tc_idx = machine.fu_index("TC")
tma_idx = machine.fu_index("TMA")
rrt_A = make_rrt(load_cycles, {tma_idx: [1] * load_cycles}, num_fus)
graph.add_instruction(Instruction("A_load", rrt_A, variable_latency=True, streaming=True))
rrt_B = make_rrt(load_cycles, {tma_idx: [1] * load_cycles}, num_fus)
graph.add_instruction(Instruction("B_load", rrt_B, variable_latency=True, streaming=True))
rrt_C = make_rrt(compute_cycles, {tc_idx: [1] * compute_cycles}, num_fus)
graph.add_instruction(Instruction("C", rrt_C))
graph.add_edge(DependenceEdge("A_load", "C", delay=load_cycles, iteration_delay=0))
graph.add_edge(DependenceEdge("B_load", "C", delay=load_cycles, iteration_delay=0))
graph.add_edge(DependenceEdge("C", "C", delay=compute_cycles, iteration_delay=1))
return graph
def custom_kernel(
machine: MachineDescription,
instructions: list,
edges: list,
) -> DependenceGraph:
"""Build a custom kernel dependence graph.
Args:
machine: Target machine description
instructions: List of dicts with keys:
- name: str
- cycles: int
- fu: str (functional unit name)
- variable_latency: bool (optional)
- streaming: bool (optional)
- memory: dict (optional, memory space -> bytes)
edges: List of dicts with keys:
- src: str
- dst: str
- delay: int
- delta: int (iteration delay, default 0)
Returns:
DependenceGraph
"""
graph = DependenceGraph(machine)
num_fus = machine.num_functional_units
for instr_desc in instructions:
fu_name = instr_desc["fu"]
fu_idx = machine.fu_index(fu_name)
cycles = instr_desc["cycles"]
rrt = make_rrt(cycles, {fu_idx: [1] * cycles}, num_fus)
graph.add_instruction(Instruction(
name=instr_desc["name"],
rrt=rrt,
variable_latency=instr_desc.get("variable_latency", False),
streaming=instr_desc.get("streaming", False),
memory_footprint=instr_desc.get("memory", {}),
))
for edge_desc in edges:
graph.add_edge(DependenceEdge(
src=edge_desc["src"],
dst=edge_desc["dst"],
delay=edge_desc["delay"],
iteration_delay=edge_desc.get("delta", 0),
))
return graph
|