Skip to content

Commit 38c6284

Browse files
antiagainstantiagainst
authored
Revert "[AMD][Pipeliner] Improve clustering and add prefetch (#4881)" (#5157)
This reverts commit cc25374 due to perf regressions.
1 parent d5e06fe commit 38c6284

File tree

10 files changed

+648
-548
lines changed

10 files changed

+648
-548
lines changed

‎test/TritonGPU/amd/amd-reorder-instructions.mlir‎

Lines changed: 345 additions & 0 deletions
Large diffs are not rendered by default.

‎test/TritonGPU/amd/amd-sched-2nd-load.mlir‎

Lines changed: 16 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -35,11 +35,11 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
3535
%c1 = arith.constant 1 : i32
3636
%cst = arith.constant dense<0.000000e+00> : tensor<256x256xf32, #mma>
3737
%0:1 = scf.for %arg0 = %c0 to %c1 step %c1 iter_args(%arg1 = %cst) -> (tensor<256x256xf32, #mma>) : i32 {
38-
%4 = tt.load %A_ptr : tensor<256x128x!tt.ptr<f16>, #blocked>
3938
%1 = triton_gpu.local_load %A_LDS : !tt.memdesc<256x128xf16, #shared, #triton_gpu.shared_memory, mutable> -> tensor<256x128xf16, #dotOp0>
40-
%5 = tt.load %B_ptr : tensor<128x256x!tt.ptr<f16>, #blocked1>
4139
%2 = triton_gpu.local_load %B_LDS : !tt.memdesc<128x256xf16, #shared1, #triton_gpu.shared_memory, mutable> -> tensor<128x256xf16, #dotOp1>
4240
%3 = tt.dot %1, %2, %arg1 : tensor<256x128xf16, #dotOp0> * tensor<128x256xf16, #dotOp1> -> tensor<256x256xf32, #mma>
41+
%4 = tt.load %A_ptr : tensor<256x128x!tt.ptr<f16>, #blocked>
42+
%5 = tt.load %B_ptr : tensor<128x256x!tt.ptr<f16>, #blocked1>
4343
triton_gpu.local_store %4, %A_LDS : tensor<256x128xf16, #blocked> -> !tt.memdesc<256x128xf16, #shared, #triton_gpu.shared_memory, mutable>
4444
triton_gpu.local_store %5, %B_LDS : tensor<128x256xf16, #blocked1> -> !tt.memdesc<128x256xf16, #shared1, #triton_gpu.shared_memory, mutable>
4545
scf.yield %3 : tensor<256x256xf32, #mma>
@@ -74,11 +74,11 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
7474
%c1 = arith.constant 1 : i32
7575
%cst = arith.constant dense<0.000000e+00> : tensor<256x256xf32, #mma>
7676
%0:1 = scf.for %arg0 = %c0 to %c1 step %c1 iter_args(%arg1 = %cst) -> (tensor<256x256xf32, #mma>) : i32 {
77-
%4 = tt.load %A_ptr : tensor<256x64x!tt.ptr<f16>, #blocked>
7877
%1 = triton_gpu.local_load %A_LDS : !tt.memdesc<256x64xf16, #shared, #triton_gpu.shared_memory, mutable> -> tensor<256x64xf16, #dotOp0>
79-
%5 = tt.load %B_ptr : tensor<64x256x!tt.ptr<f16>, #blocked1>
8078
%2 = triton_gpu.local_load %B_LDS : !tt.memdesc<64x256xf16, #shared1, #triton_gpu.shared_memory, mutable> -> tensor<64x256xf16, #dotOp1>
8179
%3 = tt.dot %1, %2, %arg1 : tensor<256x64xf16, #dotOp0> * tensor<64x256xf16, #dotOp1> -> tensor<256x256xf32, #mma>
80+
%4 = tt.load %A_ptr : tensor<256x64x!tt.ptr<f16>, #blocked>
81+
%5 = tt.load %B_ptr : tensor<64x256x!tt.ptr<f16>, #blocked1>
8282
triton_gpu.local_store %4, %A_LDS : tensor<256x64xf16, #blocked> -> !tt.memdesc<256x64xf16, #shared, #triton_gpu.shared_memory, mutable>
8383
triton_gpu.local_store %5, %B_LDS : tensor<64x256xf16, #blocked1> -> !tt.memdesc<64x256xf16, #shared1, #triton_gpu.shared_memory, mutable>
8484
scf.yield %3 : tensor<256x256xf32, #mma>
@@ -101,9 +101,9 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
101101
// Should NOT apply: tile size 256x64x128 with single dot
102102
// CHECK-LABEL: sink_2nd_load_256x64x128
103103
// CHECK: %[[tileA:.*]] = tt.load
104-
// CHECK-NEXT: local_load
105104
// CHECK-NEXT: %[[tileB:.*]] = tt.load
106105
// CHECK-NEXT: local_load
106+
// CHECK-NEXT: local_load
107107
// CHECK-NEXT: tt.dot
108108
// CHECK-NEXT: triton_gpu.local_store %[[tileA]]
109109
// CHECK-NEXT: triton_gpu.local_store %[[tileB]]
@@ -113,11 +113,11 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
113113
%c1 = arith.constant 1 : i32
114114
%cst = arith.constant dense<0.000000e+00> : tensor<256x64xf32, #mma>
115115
%0:1 = scf.for %arg0 = %c0 to %c1 step %c1 iter_args(%arg1 = %cst) -> (tensor<256x64xf32, #mma>) : i32 {
116-
%4 = tt.load %A_ptr : tensor<256x128x!tt.ptr<f16>, #blocked>
117116
%1 = triton_gpu.local_load %A_LDS : !tt.memdesc<256x128xf16, #shared, #triton_gpu.shared_memory, mutable> -> tensor<256x128xf16, #dotOp0>
118-
%5 = tt.load %B_ptr : tensor<128x64x!tt.ptr<f16>, #blocked1>
119117
%2 = triton_gpu.local_load %B_LDS : !tt.memdesc<128x64xf16, #shared1, #triton_gpu.shared_memory, mutable> -> tensor<128x64xf16, #dotOp1>
120118
%3 = tt.dot %1, %2, %arg1 : tensor<256x128xf16, #dotOp0> * tensor<128x64xf16, #dotOp1> -> tensor<256x64xf32, #mma>
119+
%4 = tt.load %A_ptr : tensor<256x128x!tt.ptr<f16>, #blocked>
120+
%5 = tt.load %B_ptr : tensor<128x64x!tt.ptr<f16>, #blocked1>
121121
triton_gpu.local_store %4, %A_LDS : tensor<256x128xf16, #blocked> -> !tt.memdesc<256x128xf16, #shared, #triton_gpu.shared_memory, mutable>
122122
triton_gpu.local_store %5, %B_LDS : tensor<128x64xf16, #blocked1> -> !tt.memdesc<128x64xf16, #shared1, #triton_gpu.shared_memory, mutable>
123123
scf.yield %3 : tensor<256x64xf32, #mma>
@@ -140,9 +140,9 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
140140
// Should NOT apply: tile size 256x256x32 with single dot
141141
// CHECK-LABEL: sink_2nd_load_256x256x32
142142
// CHECK: %[[tileA:.*]] = tt.load
143-
// CHECK-NEXT: local_load
144143
// CHECK-NEXT: %[[tileB:.*]] = tt.load
145144
// CHECK-NEXT: local_load
145+
// CHECK-NEXT: local_load
146146
// CHECK-NEXT: tt.dot
147147
// CHECK-NEXT: triton_gpu.local_store %[[tileA]]
148148
// CHECK-NEXT: triton_gpu.local_store %[[tileB]]
@@ -152,11 +152,11 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
152152
%c1 = arith.constant 1 : i32
153153
%cst = arith.constant dense<0.000000e+00> : tensor<256x256xf32, #mma>
154154
%0:1 = scf.for %arg0 = %c0 to %c1 step %c1 iter_args(%arg1 = %cst) -> (tensor<256x256xf32, #mma>) : i32 {
155-
%4 = tt.load %A_ptr : tensor<256x32x!tt.ptr<f16>, #blocked>
156155
%1 = triton_gpu.local_load %A_LDS : !tt.memdesc<256x32xf16, #shared, #triton_gpu.shared_memory, mutable> -> tensor<256x32xf16, #dotOp0>
157-
%5 = tt.load %B_ptr : tensor<32x256x!tt.ptr<f16>, #blocked1>
158156
%2 = triton_gpu.local_load %B_LDS : !tt.memdesc<32x256xf16, #shared1, #triton_gpu.shared_memory, mutable> -> tensor<32x256xf16, #dotOp1>
159157
%3 = tt.dot %1, %2, %arg1 : tensor<256x32xf16, #dotOp0> * tensor<32x256xf16, #dotOp1> -> tensor<256x256xf32, #mma>
158+
%4 = tt.load %A_ptr : tensor<256x32x!tt.ptr<f16>, #blocked>
159+
%5 = tt.load %B_ptr : tensor<32x256x!tt.ptr<f16>, #blocked1>
160160
triton_gpu.local_store %4, %A_LDS : tensor<256x32xf16, #blocked> -> !tt.memdesc<256x32xf16, #shared, #triton_gpu.shared_memory, mutable>
161161
triton_gpu.local_store %5, %B_LDS : tensor<32x256xf16, #blocked1> -> !tt.memdesc<32x256xf16, #shared1, #triton_gpu.shared_memory, mutable>
162162
scf.yield %3 : tensor<256x256xf32, #mma>
@@ -181,9 +181,9 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
181181
// Should NOT apply: the 2nd load has a user before the dot
182182
// CHECK-LABEL: sink_2nd_load_128x128x128_user_before_dot
183183
// CHECK: %[[tileA:.*]] = tt.load
184-
// CHECK-NEXT: local_load
185184
// CHECK-NEXT: %[[tileB:.*]] = tt.load
186185
// CHECK-NEXT: local_load
186+
// CHECK-NEXT: local_load
187187
// CHECK-NEXT: tt.store
188188
// CHECK-NEXT: tt.dot
189189
// CHECK-NEXT: triton_gpu.local_store %[[tileA]]
@@ -193,10 +193,10 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
193193
%c1 = arith.constant 1 : i32
194194
%cst = arith.constant dense<0.000000e+00> : tensor<128x128xf32, #mma>
195195
%0:1 = scf.for %arg0 = %c0 to %c1 step %c1 iter_args(%arg1 = %cst) -> (tensor<128x128xf32, #mma>) : i32 {
196-
%4 = tt.load %A_ptr : tensor<128x128x!tt.ptr<f16>, #blocked>
197196
%1 = triton_gpu.local_load %A_LDS : !tt.memdesc<128x128xf16, #shared, #triton_gpu.shared_memory, mutable> -> tensor<128x128xf16, #dotOp0>
198-
%5 = tt.load %B_ptr : tensor<128x128x!tt.ptr<i64>, #blocked>
199197
%2 = triton_gpu.local_load %B_LDS : !tt.memdesc<128x128xf16, #shared1, #triton_gpu.shared_memory, mutable> -> tensor<128x128xf16, #dotOp1>
198+
%4 = tt.load %A_ptr : tensor<128x128x!tt.ptr<f16>, #blocked>
199+
%5 = tt.load %B_ptr : tensor<128x128x!tt.ptr<i64>, #blocked>
200200
tt.store %B_ptr, %5 : tensor<128x128x!tt.ptr<i64>, #blocked>
201201
%3 = tt.dot %1, %2, %arg1 : tensor<128x128xf16, #dotOp0> * tensor<128x128xf16, #dotOp1> -> tensor<128x128xf32, #mma>
202202
triton_gpu.local_store %4, %A_LDS : tensor<128x128xf16, #blocked> -> !tt.memdesc<128x128xf16, #shared, #triton_gpu.shared_memory, mutable>
@@ -213,12 +213,12 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
213213
// Category 3: two dots in the for loop. Make sure the optimization is not applied
214214
// should NOT apply: two dots
215215
// CHECK-LABEL: sink_2nd_load_256x256x64_two_dot
216-
// CHECK: tt.load
217-
// CHECK-NEXT: tt.load
218-
// CHECK-NEXT: triton_gpu.local_load
216+
// CHECK: triton_gpu.local_load
219217
// CHECK-NEXT: triton_gpu.local_load
220218
// CHECK-NEXT: tt.dot
221219
// CHECK-NEXT: tt.dot
220+
// CHECK-NEXT: tt.load
221+
// CHECK-NEXT: tt.load
222222
// CHECK-NEXT: triton_gpu.local_store
223223
// CHECK-NEXT: triton_gpu.local_store
224224
#blocked = #triton_gpu.blocked<{sizePerThread = [1, 8], threadsPerWarp = [8, 8], warpsPerCTA = [1, 1], order = [1, 0]}>

‎test/TritonGPU/loop-pipeline-hip.mlir‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -35,9 +35,9 @@ module attributes {"triton_gpu.target" = "hip:gfx942", "triton_gpu.num-ctas" = 1
3535
%16 = tt.addptr %14, %15 : tensor<64x16x!tt.ptr<f16>, #blocked>, tensor<64x16xi32, #blocked>
3636
// CHECK: triton_gpu.local_store
3737
// CHECK: scf.for
38-
// CHECK: tt.load
3938
// CHECK: tt.dot
4039
// CHECK: tt.dot
40+
// CHECK: tt.load
4141
// CHECK: triton_gpu.local_store
4242
// CHECK: scf.yield
4343
%17:2 = scf.for %arg2 = %c0_i32 to %c8_i32 step %c1_i32 iter_args(%arg3 = %cst_1, %arg4 = %cst_2) -> (tensor<128x16xf32, #mma>, tensor<128x64xf32, #mma>) : i32 {
@@ -165,9 +165,9 @@ module attributes {"triton_gpu.target" = "hip:gfx942", "triton_gpu.num-ctas" = 1
165165
// CHECK-LABEL: tt.func public @add_barrier_kernel
166166
// CHECK: tt.load
167167
// CHECK: scf.for
168-
// CHECK: tt.load
169168
// CHECK: gpu.barrier
170169
// CHECK: tt.store
170+
// CHECK: tt.load
171171
// CHECK: scf.yield
172172
// CHECK: gpu.barrier
173173
// CHECK: tt.store

0 commit comments

Comments
 (0)