Merge triton-mlir branch - Complete rewrite of the backend from scratch (#1004)
This PR merges the `triton-mlir` branch, in which we have been quietly rewriting the Triton backend from scratch to increase maintainability, stability and ultimately performance. Changes to the runtime are minimal, and this new version aims to remain backward-compatible with the previous commit. The legacy backend is now officially deprecated, but can still be accessed via the `legacy-backend` tag. Co-authored-by: Keren Zhou <kerenzhou@openai.com> Co-authored-by: Yan Chunwei <yanchunwei@outlook.com> Co-authored-by: goostavz <109190422+goostavz@users.noreply.github.com> Co-authored-by: Shintaro Iwasaki <siwasaki@fb.com> Co-authored-by: Yan Da <dyanab@connect.ust.hk> Co-authored-by: Jun Yang <yangjunpro@gmail.com> Co-authored-by: Ian Bearman <ianb@microsoft.com> Co-authored-by: Jason Ansel <jansel@jansel.net> Co-authored-by: Qingyi Liu <qingyil@nvidia.com> Co-authored-by: ben-zhang-609 <110140741+ben-zhang-609@users.noreply.github.com> Co-authored-by: Chenggang Zhao <lyricz@yeah.net> Co-authored-by: ben-zhang-609 <benzh609@gmail.com> Co-authored-by: dongdongl <dongdongl@nvidia.com>
This commit is contained in:
Philippe Tillet
GitHub
146
test/Triton/combine.mlir
Normal file
146
test/Triton/combine.mlir
Normal file
@@ -0,0 +1,146 @@
|
||||
// RUN: triton-opt %s -split-input-file -canonicalize -triton-combine
|
||||
// RUN: triton-opt %s -split-input-file -canonicalize -triton-combine | FileCheck %s
|
||||
|
||||
// CHECK-LABEL: @test_combine_dot_add_pattern
|
||||
func @test_combine_dot_add_pattern() -> (tensor<128x128xf32>, tensor<128x128xf32>) {
|
||||
// CHECK: %[[d:.*]] = arith.constant dense<3.000000e+00> : tensor<128x128xf32>
|
||||
// CHECK: %[[b:.*]] = arith.constant dense<2.000000e+00> : tensor<128x128xf32>
|
||||
// CHECK: %[[a:.*]] = arith.constant dense<1.000000e+00> : tensor<128x128xf32>
|
||||
%a = arith.constant dense<1.0> : tensor<128x128xf32>
|
||||
%b = arith.constant dense<2.0> : tensor<128x128xf32>
|
||||
%zero = arith.constant dense<0.0> : tensor<128x128xf32>
|
||||
%d = arith.constant dense<3.0> : tensor<128x128xf32>
|
||||
|
||||
%dot_out = tt.dot %a, %b, %zero {allowTF32 = true, transA = false, transB = false} : tensor<128x128xf32> * tensor<128x128xf32> -> tensor<128x128xf32>
|
||||
|
||||
// CHECK-NEXT: %[[res0:.*]] = tt.dot %[[a]], %[[b]], %[[d]] {allowTF32 = true} : tensor<128x128xf32> * tensor<128x128xf32> -> tensor<128x128xf32>
|
||||
%res0 = arith.addf %dot_out, %d : tensor<128x128xf32>
|
||||
|
||||
// CHECK-NEXT: %[[res1:.*]] = tt.dot %[[a]], %[[b]], %[[d]] {allowTF32 = true} : tensor<128x128xf32> * tensor<128x128xf32> -> tensor<128x128xf32>
|
||||
%res1 = arith.addf %d, %dot_out : tensor<128x128xf32>
|
||||
|
||||
return %res0, %res1 : tensor<128x128xf32>, tensor<128x128xf32>
|
||||
}
|
||||
|
||||
|
||||
// COM: CHECK-LABEL: @test_combine_addptr_pattern
|
||||
func @test_combine_addptr_pattern(%base: !tt.ptr<f32>) -> tensor<8x!tt.ptr<f32>> {
|
||||
%off0 = arith.constant 10 : i32
|
||||
%off1 = arith.constant 15 : i32
|
||||
|
||||
// 10 + 15 = 25
|
||||
// COM: CHECK-NEXT: %[[cst:.*]] = arith.constant dense<25> : tensor<8xi32>
|
||||
|
||||
%base_ = tt.broadcast %base : (!tt.ptr<f32>) -> tensor<8x!tt.ptr<f32>>
|
||||
|
||||
// COM: CHECK-NEXT: %[[tmp0:.*]] = tt.broadcast %{{.*}} : (!tt.ptr<f32>) -> tensor<8x!tt.ptr<f32>>
|
||||
|
||||
%idx0 = tt.broadcast %off0 : (i32) -> tensor<8xi32>
|
||||
%idx1 = tt.broadcast %off1 : (i32) -> tensor<8xi32>
|
||||
|
||||
// COM: CHECK-NEXT: %1 = tt.addptr %[[tmp0]], %[[cst]] : tensor<8x!tt.ptr<f32>>, tensor<8xi32>
|
||||
%ptr0 = tt.addptr %base_, %idx0 : tensor<8x!tt.ptr<f32>>, tensor<8xi32>
|
||||
%ptr1 = tt.addptr %ptr0, %idx1 : tensor<8x!tt.ptr<f32>>, tensor<8xi32>
|
||||
|
||||
return %ptr1 : tensor<8x!tt.ptr<f32>>
|
||||
}
|
||||
|
||||
|
||||
// CHECK-LABEL: @test_combine_select_masked_load_pattern
|
||||
func @test_combine_select_masked_load_pattern(%ptr: tensor<8x!tt.ptr<f32>>, %cond: i1) -> (tensor<8xf32>, tensor<8xf32>) {
|
||||
%mask = tt.broadcast %cond : (i1) -> tensor<8xi1>
|
||||
%false_val = arith.constant dense<0.0> : tensor<8xf32>
|
||||
|
||||
// CHECK: %[[res1:.*]] = tt.load %{{.*}}, %{{.*}}, %{{.*}} {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
%x = tt.load %ptr, %mask, %false_val {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
%0 = select %cond, %x, %false_val : tensor<8xf32>
|
||||
|
||||
// CHECK: %[[res2:.*]] = tt.load %{{.*}}, %{{.*}}, %{{.*}} {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
%y = tt.load %ptr, %mask, %false_val {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
%1 = select %cond, %y, %false_val : tensor<8xf32>
|
||||
|
||||
// CHECK: return %[[res1]], %[[res2]] : tensor<8xf32>, tensor<8xf32>
|
||||
return %0, %1 : tensor<8xf32>, tensor<8xf32>
|
||||
}
|
||||
|
||||
// CHECK-LABEL: @test_combine_select_masked_load_fail_pattern
|
||||
func @test_combine_select_masked_load_fail_pattern(%ptr: tensor<8x!tt.ptr<f32>>, %dummy_load: tensor<8xf32>, %dummy_broadcast: tensor<8xi1>, %cond: i1) -> (tensor<8xf32>, tensor<8xf32>) {
|
||||
%false_val = arith.constant dense<0.0> : tensor<8xf32>
|
||||
|
||||
// Case 1: value at the "load" position is not an "op". Select should not be canonicalized.
|
||||
// CHECK: %{{.*}} = select %{{.*}}, %{{.*}}, %{{.*}} : tensor<8xf32>
|
||||
%0 = select %cond, %dummy_load, %false_val : tensor<8xf32>
|
||||
|
||||
// Case 2: value at the "broadcast" position is not an "op". Select should not be canonicalized.
|
||||
%real_load = tt.load %ptr, %dummy_broadcast, %false_val {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
// CHECK: %{{.*}} = select %{{.*}}, %{{.*}}, %{{.*}} : tensor<8xf32>
|
||||
%1 = select %cond, %real_load, %false_val : tensor<8xf32>
|
||||
|
||||
return %0, %1 : tensor<8xf32>, tensor<8xf32>
|
||||
}
|
||||
|
||||
// CHECK-LABEL: @test_combine_broadcast_constant_pattern
|
||||
func @test_combine_broadcast_constant_pattern(%cst : f32) -> tensor<8x2xf32> {
|
||||
// CHECK: %[[cst:.*]] = arith.constant dense<1.000000e+00> : tensor<8x2xf32>
|
||||
%const = arith.constant dense<1.0> : tensor<8xf32>
|
||||
%bst_out = tt.broadcast %const : (tensor<8xf32>) -> tensor<8x2xf32>
|
||||
|
||||
// CHECK-NEXT: return %[[cst]] : tensor<8x2xf32>
|
||||
return %bst_out : tensor<8x2xf32>
|
||||
}
|
||||
|
||||
// CHECK-LABEL: @test_canonicalize_masked_load_pattern
|
||||
func @test_canonicalize_masked_load_pattern(%ptr: tensor<8x!tt.ptr<f32>>) -> (tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) {
|
||||
%true_mask = arith.constant dense<true> : tensor<8xi1>
|
||||
%false_mask = arith.constant dense<false> : tensor<8xi1>
|
||||
%other_val = arith.constant dense<0.0> : tensor<8xf32>
|
||||
|
||||
// true_mask with other
|
||||
// CHECK: %[[res1:.*]] = tt.load %{{.*}} {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
%x = tt.load %ptr, %true_mask {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
|
||||
// true_mask without other
|
||||
// CHECK: %[[res2:.*]] = tt.load %{{.*}} {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
%y = tt.load %ptr, %true_mask, %other_val {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
|
||||
// false_mask with other. It should become "other" (i.e., %y)
|
||||
%z = tt.load %ptr, %false_mask, %y {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
|
||||
// CHECK: return %[[res1]], %[[res2]], %[[res2]] : tensor<8xf32>, tensor<8xf32>, tensor<8xf32>
|
||||
return %x, %y, %z: tensor<8xf32>, tensor<8xf32>, tensor<8xf32>
|
||||
}
|
||||
|
||||
// CHECK-LABEL: @test_canonicalize_masked_load_fail_pattern
|
||||
func @test_canonicalize_masked_load_fail_pattern(%ptr: tensor<8x!tt.ptr<f32>>, %mask: tensor<8xi1>) -> (tensor<8xf32>, tensor<8xf32>) {
|
||||
%other_val = arith.constant dense<0.0> : tensor<8xf32>
|
||||
|
||||
// Case: value at the "mask" position is not an "op". Load should not be canonicalized.
|
||||
// CHECK: %[[res1:.*]] = tt.load %{{.*}}, %{{.*}} {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
%x = tt.load %ptr, %mask {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
// CHECK: %[[res1:.*]] = tt.load %{{.*}}, %{{.*}}, %{{.*}} {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
%y = tt.load %ptr, %mask, %other_val {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<8xf32>
|
||||
|
||||
return %x, %y: tensor<8xf32>, tensor<8xf32>
|
||||
}
|
||||
|
||||
// CHECK-LABEL: @test_canonicalize_masked_store_pattern
|
||||
func @test_canonicalize_masked_store_pattern(%ptr: tensor<8x!tt.ptr<f32>>, %val: tensor<8xf32>) {
|
||||
%true_mask = arith.constant dense<true> : tensor<8xi1>
|
||||
%false_mask = arith.constant dense<false> : tensor<8xi1>
|
||||
|
||||
// CHECK: tt.store %{{.*}}, %{{.*}} : tensor<8xf32>
|
||||
tt.store %ptr, %val, %true_mask : tensor<8xf32>
|
||||
|
||||
// The following store should disappear.
|
||||
// CHECK-NEXT: return
|
||||
tt.store %ptr, %val, %false_mask : tensor<8xf32>
|
||||
return
|
||||
}
|
||||
|
||||
// CHECK-LABEL: @test_canonicalize_masked_store_fail_pattern
|
||||
func @test_canonicalize_masked_store_fail_pattern(%ptr: tensor<8x!tt.ptr<f32>>, %val: tensor<8xf32>, %mask: tensor<8xi1>) {
|
||||
// Case: value at the "mask" position is not an "op". Store should not be canonicalized.
|
||||
// CHECK: tt.store %{{.*}}, %{{.*}}, %{{.*}} : tensor<8xf32>
|
||||
tt.store %ptr, %val, %mask : tensor<8xf32>
|
||||
return
|
||||
}
|
||||
130
test/Triton/vecadd.mlir
Normal file
130
test/Triton/vecadd.mlir
Normal file
@@ -0,0 +1,130 @@
|
||||
// RUN: triton-opt %s -verify-diagnostics
|
||||
|
||||
module {
|
||||
func @add_kernel__Pfp32_Pfp32_Pfp32_i32_i32_i32__(%arg0: !tt.ptr<f32>, %arg1: !tt.ptr<f32>, %arg2: !tt.ptr<f32>, %arg3: i32, %arg4: i32, %arg5: i32) {
|
||||
%0 = tt.get_program_id {axis = 0 : i32} : i32
|
||||
%c256_i32 = arith.constant 256 : i32
|
||||
%1 = arith.muli %0, %c256_i32 : i32
|
||||
%2 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32>
|
||||
%3 = tt.broadcast %1 : (i32) -> tensor<256xi32>
|
||||
%4 = arith.addi %3, %2 : tensor<256xi32>
|
||||
%5 = tt.broadcast %arg3 : (i32) -> tensor<256xi32>
|
||||
%6 = arith.cmpi slt, %4, %5 : tensor<256xi32>
|
||||
%7 = tt.broadcast %arg0 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>>
|
||||
%8 = tt.addptr %7, %4 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
|
||||
%9 = tt.broadcast %arg1 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>>
|
||||
%10 = tt.addptr %9, %4 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
|
||||
%cst = arith.constant 0.000000e+00 : f32
|
||||
%11 = tt.broadcast %cst : (f32) -> tensor<256xf32>
|
||||
%c0_i32 = arith.constant 0 : i32
|
||||
%c32_i32 = arith.constant 32 : i32
|
||||
%12 = arith.index_cast %c0_i32 : i32 to index
|
||||
%13 = arith.index_cast %arg4 : i32 to index
|
||||
%14 = arith.index_cast %c32_i32 : i32 to index
|
||||
%15:3 = scf.for %arg6 = %12 to %13 step %14 iter_args(%arg7 = %11, %arg8 = %8, %arg9 = %10) -> (tensor<256xf32>, tensor<256x!tt.ptr<f32>>, tensor<256x!tt.ptr<f32>>) {
|
||||
%cst_0 = arith.constant 0.000000e+00 : f32
|
||||
%18 = tt.broadcast %cst_0 : (f32) -> tensor<256xf32>
|
||||
%19 = tt.load %arg8, %6, %18 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256xf32>
|
||||
%cst_1 = arith.constant 0.000000e+00 : f32
|
||||
%20 = tt.broadcast %cst_1 : (f32) -> tensor<256xf32>
|
||||
%21 = tt.load %arg9, %6, %20 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256xf32>
|
||||
%22 = arith.addf %19, %21 : tensor<256xf32>
|
||||
%23 = arith.addf %arg7, %22 : tensor<256xf32>
|
||||
%24 = tt.broadcast %arg5 : (i32) -> tensor<256xi32>
|
||||
%25 = tt.addptr %arg8, %24 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
|
||||
%26 = tt.broadcast %arg5 : (i32) -> tensor<256xi32>
|
||||
%27 = tt.addptr %arg9, %26 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
|
||||
scf.yield %23, %25, %27 : tensor<256xf32>, tensor<256x!tt.ptr<f32>>, tensor<256x!tt.ptr<f32>>
|
||||
}
|
||||
%16 = tt.broadcast %arg2 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>>
|
||||
%17 = tt.addptr %16, %4 : tensor<256x!tt.ptr<f32>>, tensor<256xi32>
|
||||
tt.store %17, %15#0, %6 : tensor<256xf32>
|
||||
return
|
||||
}
|
||||
}
|
||||
// module {
|
||||
// func @add_kernel__Pfp32_Pfp32_Pfp32_i32_i32_i32__(%arg0: !tt.ptr<f32>, %arg1: !tt.ptr<f32>, %arg2: !tt.ptr<f32>, %arg3: i32, %arg4: i32, %arg5: i32) {
|
||||
// %c64 = arith.constant 64 : index
|
||||
// %c32 = arith.constant 32 : index
|
||||
// %c0 = arith.constant 0 : index
|
||||
// %cst = arith.constant 0.000000e+00 : f32
|
||||
// %c256_i32 = arith.constant 256 : i32
|
||||
// %0 = tt.get_program_id {axis = 0 : i32} : i32
|
||||
// %1 = arith.muli %0, %c256_i32 : i32
|
||||
// %2 = tt.make_range {end = 256 : i32, start = 0 : i32} : tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %3 = tt.broadcast %1 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %4 = arith.addi %3, %2 : tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %5 = tt.broadcast %arg3 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %6 = "triton_gpu.cmpi"(%4, %5) {predicate = 2 : i64} : (tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %7 = tt.broadcast %arg0 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %8 = tt.addptr %7, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %9 = tt.broadcast %arg1 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %10 = tt.addptr %9, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %11 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %12 = arith.index_cast %arg4 : i32 to index
|
||||
// %13 = arith.cmpi slt, %c0, %12 : index
|
||||
// %14 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %15 = tt.broadcast %13 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %16 = arith.andi %6, %15 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %17 = triton_gpu.copy_async %8, %16, %14 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %18 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %19 = tt.broadcast %13 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %20 = arith.andi %6, %19 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %21 = triton_gpu.copy_async %10, %20, %18 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %22 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %23 = tt.addptr %8, %22, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %24 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %25 = tt.addptr %10, %24, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %26 = arith.cmpi slt, %c32, %12 : index
|
||||
// %27 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %28 = tt.broadcast %26 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %29 = arith.andi %6, %28 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %30 = triton_gpu.copy_async %23, %29, %27 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %31 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %32 = tt.broadcast %26 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %33 = arith.andi %6, %32 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %34 = triton_gpu.copy_async %25, %33, %31 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %35 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %36 = tt.addptr %23, %35, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %37 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %38 = tt.addptr %25, %37, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %39 = arith.cmpi slt, %c64, %12 : index
|
||||
// %40 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %41 = tt.broadcast %39 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %42 = arith.andi %6, %41 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %43 = triton_gpu.copy_async %36, %42, %40 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %44 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %45 = tt.broadcast %39 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %46 = arith.andi %6, %45 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %47 = triton_gpu.copy_async %38, %46, %44 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %48 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %49 = tt.addptr %36, %48, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %50 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %51 = tt.addptr %38, %50, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %52:12 = scf.for %arg6 = %c0 to %12 step %c32 iter_args(%arg7 = %11, %arg8 = %8, %arg9 = %10, %arg10 = %17, %arg11 = %30, %arg12 = %43, %arg13 = %21, %arg14 = %34, %arg15 = %47, %arg16 = %51, %arg17 = %49, %arg18 = %c64) -> (tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, index) {
|
||||
// %55 = arith.addf %arg10, %arg13 : tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %56 = arith.addf %arg7, %55 : tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %57 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %58 = tt.addptr %arg8, %57, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %59 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %60 = tt.addptr %arg9, %59, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %61 = arith.addi %arg18, %c32 : index
|
||||
// %62 = arith.cmpi slt, %61, %12 : index
|
||||
// %63 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %64 = tt.broadcast %62 : (i1) -> tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %65 = arith.andi %64, %6 : tensor<256xi1, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %66 = triton_gpu.copy_async %arg17, %65, %63 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %67 = tt.broadcast %cst : (f32) -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %68 = triton_gpu.copy_async %arg16, %65, %67 {cache = 1 : i32, evict = 1 : i32, isVolatile = false} : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">> -> tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %69 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %70 = tt.addptr %arg17, %69, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// %71 = tt.broadcast %arg5 : (i32) -> tensor<256xi32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %72 = tt.addptr %arg16, %71, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// scf.yield %56, %58, %60, %arg11, %arg12, %66, %arg14, %arg15, %68, %72, %70, %61 : tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, index
|
||||
// }
|
||||
// %53 = tt.broadcast %arg2 : (!tt.ptr<f32>) -> tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// %54 = tt.addptr %53, %4, : tensor<256x!tt.ptr<f32>, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>, tensor<256xi32>
|
||||
// tt.store %54, %52#0, %6 : tensor<256xf32, #triton_gpu<"coalesced encoding<threadTileSize = 1, blockTileSize = 32, order = 0>">>
|
||||
// return
|
||||
// }
|
||||
// }
|
||||
Reference in New Issue
Block a user