Skip to content

Commit

Permalink
Optimize dynamic objectFifo (#1934)
Browse files Browse the repository at this point in the history
  • Loading branch information
@pvasireddy-amd
pvasireddy-amd authored Dec 9, 2024
1 parent fe1b7b1 commit c7e37e9
Show file tree
Hide file tree
Showing 3 changed files with 63 additions and 45 deletions.
6 changes: 5 additions & 1 deletion lib/Dialect/AIE/Transforms/AIEObjectFifoStatefulTransform.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1063,7 +1063,11 @@ struct AIEObjectFifoStatefulTransformPass
Value val = builder.create<arith::ConstantOp>(
oldCounter.getLoc(), builder.getI32IntegerAttr(relOp.getSize()));
Value sum = builder.create<arith::AddIOp>(val.getLoc(), oldCounter, val);
Value newCounter = builder.create<arith::RemSIOp>(sum.getLoc(), sum, size);
Value isGreaterEqual = builder.create<arith::CmpIOp>(
sum.getLoc(), arith::CmpIPredicate::sge, sum, size);
Value newCounter = builder.create<arith::SelectOp>(
sum.getLoc(), isGreaterEqual,
builder.create<arith::SubIOp>(sum.getLoc(), sum, size), sum);
builder.create<memref::StoreOp>(size.getLoc(), newCounter, globalNextIndex,
ValueRange(ArrayRef({index.getResult()})));
}
Expand Down
16 changes: 10 additions & 6 deletions test/objectFifo-stateful-transform/dynamic_lowering_flag_test.mlir
View file
Original file line number Diff line number Diff line change
Expand Up @@ -53,14 +53,18 @@
// CHECK: %6 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %c1_i32 = arith.constant 1 : i32
// CHECK: %7 = arith.addi %6, %c1_i32 : i32
// CHECK: %8 = arith.remsi %7, %c2_i32_0 : i32
// CHECK: memref.store %8, %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %8 = arith.cmpi sge, %7, %c2_i32_0 : i32
// CHECK: %9 = arith.subi %7, %c2_i32_0 : i32
// CHECK: %10 = arith.select %8, %9, %7 : i32
// CHECK: memref.store %10, %buffer_0_2[%c1] : memref<2xi32>
// CHECK: aie.use_lock(%output_fifo_cons_lock, Release, 1)
// CHECK: %9 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %11 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %c1_i32_3 = arith.constant 1 : i32
// CHECK: %10 = arith.addi %9, %c1_i32_3 : i32
// CHECK: %11 = arith.remsi %10, %c2_i32 : i32
// CHECK: memref.store %11, %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %12 = arith.addi %11, %c1_i32_3 : i32
// CHECK: %13 = arith.cmpi sge, %12, %c2_i32 : i32
// CHECK: %14 = arith.subi %12, %c2_i32 : i32
// CHECK: %15 = arith.select %13, %14, %12 : i32
// CHECK: memref.store %15, %buffer_0_2[%c0] : memref<2xi32>
// CHECK: }
// CHECK: aie.end
// CHECK: } {dynamic_objfifo_lowering = true}
Expand Down
86 changes: 48 additions & 38 deletions test/objectFifo-stateful-transform/dynamic_lowering_test.mlir
View file
Original file line number Diff line number Diff line change
Expand Up @@ -81,13 +81,15 @@
// CHECK: %6 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %c1_i32 = arith.constant 1 : i32
// CHECK: %7 = arith.addi %6, %c1_i32 : i32
// CHECK: %8 = arith.remsi %7, %c2_i32 : i32
// CHECK: memref.store %8, %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %8 = arith.cmpi sge, %7, %c2_i32 : i32
// CHECK: %9 = arith.subi %7, %c2_i32 : i32
// CHECK: %10 = arith.select %8, %9, %7 : i32
// CHECK: memref.store %10, %buffer_0_2[%c0] : memref<2xi32>
// CHECK: scf.for %arg0 = %c0_0 to %c9 step %c1_1 {
// CHECK: aie.use_lock(%output_fifo_prod_lock, AcquireGreaterEqual, 1)
// CHECK: %24 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %25 = arith.index_cast %24 : i32 to index
// CHECK: %26 = scf.index_switch %25 -> memref<10xi32>
// CHECK: %30 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %31 = arith.index_cast %30 : i32 to index
// CHECK: %32 = scf.index_switch %31 -> memref<10xi32>
// CHECK: case 0 {
// CHECK: scf.yield %output_fifo_buff_0 : memref<10xi32>
// CHECK: }
Expand All @@ -98,9 +100,9 @@
// CHECK: scf.yield %output_fifo_buff_0 : memref<10xi32>
// CHECK: }
// CHECK: aie.use_lock(%input_fifo_cons_cons_lock, AcquireGreaterEqual, 1)
// CHECK: %27 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %28 = arith.index_cast %27 : i32 to index
// CHECK: %29 = scf.index_switch %28 -> memref<10xi32>
// CHECK: %33 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %34 = arith.index_cast %33 : i32 to index
// CHECK: %35 = scf.index_switch %34 -> memref<10xi32>
// CHECK: case 0 {
// CHECK: scf.yield %input_fifo_cons_buff_0 : memref<10xi32>
// CHECK: }
Expand All @@ -113,9 +115,9 @@
// CHECK: default {
// CHECK: scf.yield %input_fifo_cons_buff_0 : memref<10xi32>
// CHECK: }
// CHECK: %30 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %31 = arith.index_cast %30 : i32 to index
// CHECK: %32 = scf.index_switch %31 -> memref<10xi32>
// CHECK: %36 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %37 = arith.index_cast %36 : i32 to index
// CHECK: %38 = scf.index_switch %37 -> memref<10xi32>
// CHECK: case 0 {
// CHECK: scf.yield %input_fifo_cons_buff_1 : memref<10xi32>
// CHECK: }
Expand All @@ -128,24 +130,28 @@
// CHECK: default {
// CHECK: scf.yield %input_fifo_cons_buff_1 : memref<10xi32>
// CHECK: }
// CHECK: func.call @add_10_i32(%29, %32, %26) : (memref<10xi32>, memref<10xi32>, memref<10xi32>) -> ()
// CHECK: func.call @add_10_i32(%35, %38, %32) : (memref<10xi32>, memref<10xi32>, memref<10xi32>) -> ()
// CHECK: aie.use_lock(%input_fifo_cons_prod_lock, Release, 1)
// CHECK: %33 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %39 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %c1_i32_4 = arith.constant 1 : i32
// CHECK: %34 = arith.addi %33, %c1_i32_4 : i32
// CHECK: %35 = arith.remsi %34, %c3_i32 : i32
// CHECK: memref.store %35, %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %40 = arith.addi %39, %c1_i32_4 : i32
// CHECK: %41 = arith.cmpi sge, %40, %c3_i32 : i32
// CHECK: %42 = arith.subi %40, %c3_i32 : i32
// CHECK: %43 = arith.select %41, %42, %40 : i32
// CHECK: memref.store %43, %buffer_0_2[%c1] : memref<2xi32>
// CHECK: aie.use_lock(%output_fifo_cons_lock, Release, 1)
// CHECK: %36 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %44 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %c1_i32_5 = arith.constant 1 : i32
// CHECK: %37 = arith.addi %36, %c1_i32_5 : i32
// CHECK: %38 = arith.remsi %37, %c2_i32 : i32
// CHECK: memref.store %38, %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %45 = arith.addi %44, %c1_i32_5 : i32
// CHECK: %46 = arith.cmpi sge, %45, %c2_i32 : i32
// CHECK: %47 = arith.subi %45, %c2_i32 : i32
// CHECK: %48 = arith.select %46, %47, %45 : i32
// CHECK: memref.store %48, %buffer_0_2[%c0] : memref<2xi32>
// CHECK: }
// CHECK: aie.use_lock(%output_fifo_prod_lock, AcquireGreaterEqual, 1)
// CHECK: %9 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %10 = arith.index_cast %9 : i32 to index
// CHECK: %11 = scf.index_switch %10 -> memref<10xi32>
// CHECK: %11 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %12 = arith.index_cast %11 : i32 to index
// CHECK: %13 = scf.index_switch %12 -> memref<10xi32>
// CHECK: case 0 {
// CHECK: scf.yield %output_fifo_buff_0 : memref<10xi32>
// CHECK: }
Expand All @@ -156,9 +162,9 @@
// CHECK: scf.yield %output_fifo_buff_0 : memref<10xi32>
// CHECK: }
// CHECK: aie.use_lock(%input_fifo_cons_cons_lock, AcquireGreaterEqual, 1)
// CHECK: %12 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %13 = arith.index_cast %12 : i32 to index
// CHECK: %14 = scf.index_switch %13 -> memref<10xi32>
// CHECK: %14 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %15 = arith.index_cast %14 : i32 to index
// CHECK: %16 = scf.index_switch %15 -> memref<10xi32>
// CHECK: case 0 {
// CHECK: scf.yield %input_fifo_cons_buff_0 : memref<10xi32>
// CHECK: }
Expand All @@ -171,9 +177,9 @@
// CHECK: default {
// CHECK: scf.yield %input_fifo_cons_buff_0 : memref<10xi32>
// CHECK: }
// CHECK: %15 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %16 = arith.index_cast %15 : i32 to index
// CHECK: %17 = scf.index_switch %16 -> memref<10xi32>
// CHECK: %17 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %18 = arith.index_cast %17 : i32 to index
// CHECK: %19 = scf.index_switch %18 -> memref<10xi32>
// CHECK: case 0 {
// CHECK: scf.yield %input_fifo_cons_buff_1 : memref<10xi32>
// CHECK: }
Expand All @@ -186,19 +192,23 @@
// CHECK: default {
// CHECK: scf.yield %input_fifo_cons_buff_1 : memref<10xi32>
// CHECK: }
// CHECK: func.call @add_10_i32(%14, %17, %11) : (memref<10xi32>, memref<10xi32>, memref<10xi32>) -> ()
// CHECK: func.call @add_10_i32(%16, %19, %13) : (memref<10xi32>, memref<10xi32>, memref<10xi32>) -> ()
// CHECK: aie.use_lock(%input_fifo_cons_prod_lock, Release, 2)
// CHECK: %18 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %20 = memref.load %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %c2_i32_2 = arith.constant 2 : i32
// CHECK: %19 = arith.addi %18, %c2_i32_2 : i32
// CHECK: %20 = arith.remsi %19, %c3_i32 : i32
// CHECK: memref.store %20, %buffer_0_2[%c1] : memref<2xi32>
// CHECK: %21 = arith.addi %20, %c2_i32_2 : i32
// CHECK: %22 = arith.cmpi sge, %21, %c3_i32 : i32
// CHECK: %23 = arith.subi %21, %c3_i32 : i32
// CHECK: %24 = arith.select %22, %23, %21 : i32
// CHECK: memref.store %24, %buffer_0_2[%c1] : memref<2xi32>
// CHECK: aie.use_lock(%output_fifo_cons_lock, Release, 1)
// CHECK: %21 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %25 = memref.load %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %c1_i32_3 = arith.constant 1 : i32
// CHECK: %22 = arith.addi %21, %c1_i32_3 : i32
// CHECK: %23 = arith.remsi %22, %c2_i32 : i32
// CHECK: memref.store %23, %buffer_0_2[%c0] : memref<2xi32>
// CHECK: %26 = arith.addi %25, %c1_i32_3 : i32
// CHECK: %27 = arith.cmpi sge, %26, %c2_i32 : i32
// CHECK: %28 = arith.subi %26, %c2_i32 : i32
// CHECK: %29 = arith.select %27, %28, %26 : i32
// CHECK: memref.store %29, %buffer_0_2[%c0] : memref<2xi32>
// CHECK: aie.end
// CHECK: }
// CHECK: aie.shim_dma_allocation @input_fifo(MM2S, 0, 0)
Expand Down

0 comments on commit c7e37e9

Please sign in to comment.