-
Notifications
You must be signed in to change notification settings - Fork 45
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge remote-tracking branch 'origin/main' into update_llvm_49af6502c…
…6dcb4a7f7520178bd14df396f78240c
- Loading branch information
Showing
13 changed files
with
550 additions
and
138 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Large diffs are not rendered by default.
Oops, something went wrong.
File renamed without changes.
File renamed without changes.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,59 @@ | ||
| // RUN: imex-opt -imex-convert-gpu-to-spirv %s | FileCheck %s | ||
| // RUN: IMEX_NOT_PREFER_RAWSEND=1 imex-opt -imex-convert-gpu-to-spirv %s | FileCheck %s --check-prefix=LSC | ||
| module @gemm attributes {gpu.container_module} { | ||
| memref.global "private" constant @__constant_8x16xf16 : memref<8x16xf16> = dense<5.000000e-01> | ||
| memref.global "private" constant @__constant_16x16xf16 : memref<16x16xf16> = dense<1.099610e+00> | ||
| func.func @test(%arg0: memref<8x16xf16>, %arg1: memref<16x16xf16>) -> memref<8x16xf32> attributes {llvm.emit_c_interface} { | ||
| %c1 = arith.constant 1 : index | ||
| %memref = gpu.alloc host_shared () : memref<8x16xf16> | ||
| memref.copy %arg0, %memref : memref<8x16xf16> to memref<8x16xf16> | ||
| %memref_0 = gpu.alloc host_shared () : memref<16x16xf16> | ||
| memref.copy %arg1, %memref_0 : memref<16x16xf16> to memref<16x16xf16> | ||
| %memref_1 = gpu.alloc host_shared () : memref<8x16xf32> | ||
| gpu.launch_func @test_kernel::@test_kernel blocks in (%c1, %c1, %c1) threads in (%c1, %c1, %c1) args(%memref : memref<8x16xf16>, %memref_0 : memref<16x16xf16>, %memref_1 : memref<8x16xf32>) | ||
| gpu.dealloc %memref : memref<8x16xf16> | ||
| gpu.dealloc %memref_0 : memref<16x16xf16> | ||
| return %memref_1 : memref<8x16xf32> | ||
| } | ||
| gpu.module @test_kernel attributes {spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Addresses, Float16Buffer, Int64, Int16, Int8, Kernel, Linkage, Vector16, GenericPointer, Groups, Float16, Float64, AtomicFloat32AddEXT, ExpectAssumeKHR, SubgroupDispatch, VectorComputeINTEL, VectorAnyINTEL], [SPV_EXT_shader_atomic_float_add, SPV_KHR_expect_assume, SPV_INTEL_vector_compute]>, api=OpenCL, #spirv.resource_limits<>>} { | ||
| gpu.func @test_kernel(%arg0: memref<8x16xf16>, %arg1: memref<16x16xf16>, %arg2: memref<8x16xf32>) kernel attributes {VectorComputeFunctionINTEL, spirv.entry_point_abi = #spirv.entry_point_abi<>} { | ||
| // LSC: spirv.FunctionCall @llvm_genx_lsc_prefetch2d_stateless_i1_i64 | ||
| // LSC: spirv.FunctionCall @llvm_genx_lsc_prefetch2d_stateless_i1_i64 | ||
| // LSC: spirv.FunctionCall @llvm_genx_lsc_load2d_stateless_v64i32_i1_i64 | ||
| // LSC: spirv.FunctionCall @llvm_genx_lsc_load2d_stateless_v128i32_i1_i64 | ||
| // LSC: spirv.FunctionCall @llvm_genx_dpas_nosrc0_v128f32_v128i32_v64i32 | ||
| // LSC: spirv.FunctionCall @llvm_genx_lsc_store2d_stateless_i1_i64_v128f32 | ||
| // CHECK: %[[BASE:.*]] = spirv.ConvertPtrToU %arg0 : !spirv.ptr<!spirv.array<128 x f16>, CrossWorkgroup> to i64 | ||
| // CHECK: %[[BASE1:.*]] = spirv.VectorInsertDynamic %[[BASE]] | ||
| // CHECK: %[[BASE2:.*]] = spirv.Bitcast %[[BASE1]] | ||
| // CHECK: spirv.VectorInsertDynamic | ||
| // CHECK: spirv.VectorInsertDynamic | ||
| // CHECK: spirv.FunctionCall @llvm_genx_raw_send2_noresult_i1_v8i32 | ||
| // CHECK: spirv.FunctionCall @llvm_genx_raw_send2_noresult_i1_v8i32 | ||
| // CHECK: spirv.FunctionCall @llvm_genx_raw_send2_v64i32_i1_v8i32 | ||
| // CHECK: spirv.FunctionCall @llvm_genx_raw_send2_v128i32_i1_v8i32 | ||
| // CHECK: spirv.FunctionCall @llvm_genx_dpas_nosrc0_v128f32_v128i32_v64i32 | ||
| // CHECK: spirv.FunctionCall @llvm_genx_raw_sends2_noresult_i1_v8i32_v128f32 | ||
| %0 = xegpu.create_nd_tdesc %arg0[0, 0] {mode = vc} : memref<8x16xf16> -> !xegpu.tensor_desc<8x16xf16> | ||
| %1 = xegpu.create_nd_tdesc %arg1[0, 0] {mode = vc} : memref<16x16xf16> -> !xegpu.tensor_desc<16x16xf16> | ||
| %2 = xegpu.create_nd_tdesc %arg2[0, 0] {mode = vc} : memref<8x16xf32> -> !xegpu.tensor_desc<8x16xf32> | ||
| xegpu.prefetch_nd %0 {mode = vc} : !xegpu.tensor_desc<8x16xf16> | ||
| xegpu.prefetch_nd %1 {mode = vc} : !xegpu.tensor_desc<16x16xf16> | ||
|
|
||
| %3 = xegpu.load_nd %0 {mode = vc, vnni_axis = 1} : !xegpu.tensor_desc<8x16xf16> -> vector<8x8x2xf16> | ||
| %4 = xegpu.load_nd %1 {mode = vc, vnni_axis = 0} : !xegpu.tensor_desc<16x16xf16> -> vector<8x16x2xf16> | ||
| %5 = xegpu.dpas %3, %4 {mode = vc} : vector<8x8x2xf16>, vector<8x16x2xf16> -> vector<8x16xf32> | ||
| xegpu.store_nd %5, %2 {mode = vc} : vector<8x16xf32>, !xegpu.tensor_desc<8x16xf32> | ||
| gpu.return | ||
| } | ||
| } | ||
| func.func @main() attributes {llvm.emit_c_interface} { | ||
| %0 = memref.get_global @__constant_8x16xf16 : memref<8x16xf16> | ||
| %1 = memref.get_global @__constant_16x16xf16 : memref<16x16xf16> | ||
| %2 = call @test(%0, %1) : (memref<8x16xf16>, memref<16x16xf16>) -> memref<8x16xf32> | ||
| %cast = memref.cast %2 : memref<8x16xf32> to memref<*xf32> | ||
| //call @printMemrefF32(%cast) : (memref<*xf32>) -> () | ||
| return | ||
| } | ||
| func.func private @printMemrefF32(memref<*xf32>) attributes {llvm.emit_c_interface} | ||
| } |
File renamed without changes.
File renamed without changes.
File renamed without changes.
111 changes: 111 additions & 0 deletions
111
test/Integration/Dialect/XeGPU/gemm_1024x1024xf16.using.updateoffset.mlir
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,111 @@ | ||
| // RUN: %python_executable %imex_runner --requires=l0-runtime -i %s --pass-pipeline-file=%p/xegpu-to-llvm.pp \ | ||
| // RUN: --runner imex-cpu-runner -e main \ | ||
| // RUN: --entry-point-result=void \ | ||
| // RUN: --shared-libs=%irunner_utils,%mlir_runner_utils,%mlir_c_runner_utils,%levelzero_runtime --filecheck | ||
| // RUN: %python_executable %imex_runner --requires=sycl-runtime -i %s --pass-pipeline-file=%p/xegpu-to-llvm.pp \ | ||
| // RUN: --runner imex-cpu-runner -e main \ | ||
| // RUN: --entry-point-result=void \ | ||
| // RUN: --shared-libs=%irunner_utils,%mlir_runner_utils,%mlir_c_runner_utils,%sycl_runtime --filecheck | ||
| module @gemm attributes {gpu.container_module} { | ||
| memref.global "private" @__constant_1024x1024xf16 : memref<1024x1024xf16> = dense<0.0> | ||
| memref.global "private" @__constant_1024x1024xf16_ : memref<1024x1024xf16> = dense<0.0> | ||
| memref.global "private" @__constant_1024x1024xf32 : memref<1024x1024xf32> = dense<0.0> | ||
| func.func @test(%arg0: memref<1024x1024xf16>, %arg1: memref<1024x1024xf16>) -> memref<1024x1024xf32> attributes {llvm.emit_c_interface} { | ||
| %c64 = arith.constant 64 : index | ||
| %c128 = arith.constant 128 : index | ||
| %c1 = arith.constant 1 : index | ||
| %memref = gpu.alloc host_shared () : memref<1024x1024xf16> | ||
| memref.copy %arg0, %memref : memref<1024x1024xf16> to memref<1024x1024xf16> | ||
| %memref_0 = gpu.alloc host_shared () : memref<1024x1024xf16> | ||
| memref.copy %arg1, %memref_0 : memref<1024x1024xf16> to memref<1024x1024xf16> | ||
| %memref_1 = gpu.alloc host_shared () : memref<1024x1024xf32> | ||
| gpu.launch_func @test_kernel::@test_kernel blocks in (%c128, %c64, %c1) threads in (%c1, %c1, %c1) args(%memref : memref<1024x1024xf16>, %memref_0 : memref<1024x1024xf16>, %memref_1 : memref<1024x1024xf32>) | ||
| gpu.dealloc %memref : memref<1024x1024xf16> | ||
| gpu.dealloc %memref_0 : memref<1024x1024xf16> | ||
| return %memref_1 : memref<1024x1024xf32> | ||
| } | ||
| gpu.module @test_kernel attributes {spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Addresses, Float16Buffer, Int64, Int16, Int8, Kernel, Linkage, Vector16, GenericPointer, Groups, Float16, Float64, AtomicFloat32AddEXT, ExpectAssumeKHR, SubgroupDispatch, VectorComputeINTEL, VectorAnyINTEL], [SPV_EXT_shader_atomic_float_add, SPV_KHR_expect_assume, SPV_INTEL_vector_compute]>, api=OpenCL, #spirv.resource_limits<>>} { | ||
| gpu.func @test_kernel(%arg0: memref<1024x1024xf16>, %arg1: memref<1024x1024xf16>, %arg2: memref<1024x1024xf32>) kernel attributes {VectorComputeFunctionINTEL, gpu.known_block_size = array<i32: 1, 1, 1>, gpu.known_grid_size = array<i32: 128, 64, 1>, spirv.entry_point_abi = #spirv.entry_point_abi<>} { | ||
| %c0 = arith.constant 0 : index | ||
| %c16 = arith.constant 16 : index | ||
| %c8 = arith.constant 8 : index | ||
| %c1024 = arith.constant 1024 : index | ||
| %0 = gpu.block_id x | ||
| %1 = gpu.block_id y | ||
| %2 = arith.muli %0, %c8 : index | ||
| %3 = arith.muli %1, %c16 : index | ||
| %4 = xegpu.create_nd_tdesc %arg2[%2, %3] {mode = vc} : memref<1024x1024xf32> -> !xegpu.tensor_desc<8x16xf32> | ||
| %5 = xegpu.load_nd %4 {mode = vc} : !xegpu.tensor_desc<8x16xf32> -> vector<8x16xf32> | ||
| // each work-group has 1 subgroup. the subgroup caculates a [8x16 = 8x1024 * 1024x16] block | ||
| %7 = xegpu.create_nd_tdesc %arg0[%2, %c0] {mode=vc}: memref<1024x1024xf16> -> !xegpu.tensor_desc<8x16xf16> | ||
| %8 = xegpu.create_nd_tdesc %arg1[%c0, %3] {mode=vc}: memref<1024x1024xf16> -> !xegpu.tensor_desc<16x16xf16> | ||
| %6:3 = scf.for %arg3 = %c0 to %c1024 step %c16 iter_args(%arg4 = %5, %subA = %7, %subB = %8) -> (vector<8x16xf32>, !xegpu.tensor_desc<8x16xf16>, !xegpu.tensor_desc<16x16xf16>) { | ||
| %9 = xegpu.load_nd %subA {mode=vc, vnni_axis = 1}: !xegpu.tensor_desc<8x16xf16> -> vector<8x8x2xf16> | ||
| %10 = xegpu.load_nd %subB {mode=vc, vnni_axis = 0} : !xegpu.tensor_desc<16x16xf16> -> vector<8x16x2xf16> | ||
| %11 = xegpu.dpas %9, %10, %arg4 {mode=vc}: vector<8x8x2xf16>, vector<8x16x2xf16>, vector<8x16xf32> -> vector<8x16xf32> | ||
| %12 = xegpu.update_nd_offset %subA, [%c0, %c16] {mode=vc}: !xegpu.tensor_desc<8x16xf16> -> !xegpu.tensor_desc<8x16xf16> | ||
| %13 = xegpu.update_nd_offset %subB, [%c16, %c0] {mode=vc}: !xegpu.tensor_desc<16x16xf16> -> !xegpu.tensor_desc<16x16xf16> | ||
| scf.yield %11, %12, %13: vector<8x16xf32>, !xegpu.tensor_desc<8x16xf16>, !xegpu.tensor_desc<16x16xf16> | ||
| } | ||
| xegpu.store_nd %6#0, %4 {mode = vc}: vector<8x16xf32>, !xegpu.tensor_desc<8x16xf32> | ||
| gpu.return | ||
| } | ||
| } | ||
| func.func @main() attributes {llvm.emit_c_interface} { | ||
| %0 = memref.get_global @__constant_1024x1024xf16 : memref<1024x1024xf16> | ||
| %1 = memref.get_global @__constant_1024x1024xf16_ : memref<1024x1024xf16> | ||
| %ref = memref.get_global @__constant_1024x1024xf32 : memref<1024x1024xf32> | ||
| %init = arith.constant 0.0 : f16 | ||
| %c0 = arith.constant 0 : index | ||
| %c1 = arith.constant 1 : index | ||
| %c8 = arith.constant 8 : index | ||
| %c16 = arith.constant 16 : index | ||
| %c128 = arith.constant 128 : index | ||
| %c1024 = arith.constant 1024 : index | ||
| // fill the top-left block 128x128 | ||
| // A matrix: row-major, start from 0.0, increase 0.01 per element | ||
| // B matrix: A matrix + 1.0 | ||
| scf.for %arg0 = %c0 to %c128 step %c1 { | ||
| scf.for %arg1 = %c0 to %c128 step %c1 { | ||
| %int0 = arith.index_cast %arg0 : index to i16 | ||
| %int1 = arith.index_cast %arg1 : index to i16 | ||
| %c128_i16 = arith.constant 128 : i16 | ||
| %idx0 = arith.muli %int0, %c128_i16 : i16 | ||
| %idx1 = arith.addi %int1, %idx0 : i16 | ||
| %fp = arith.uitofp %idx1 : i16 to f16 | ||
| %cst100 = arith.constant 100.0 : f16 | ||
| %val0 = arith.divf %fp, %cst100 : f16 | ||
| %cst1 = arith.constant 1.0 : f16 | ||
| %val1 = arith.addf %val0, %cst1 : f16 | ||
| memref.store %val0, %0[%arg0, %arg1] : memref<1024x1024xf16> | ||
| memref.store %val1, %1[%arg0, %arg1] : memref<1024x1024xf16> | ||
| } | ||
| } | ||
| // caculate the result C matrix | ||
| scf.for %arg0 = %c0 to %c1024 step %c1 { | ||
| scf.for %arg1 = %c0 to %c1024 step %c1 { | ||
| %acc = memref.load %ref[%arg0, %arg1] : memref<1024x1024xf32> | ||
| %res = scf.for %arg2 = %c0 to %c1024 step %c1 iter_args(%arg3 = %acc) -> f32 { | ||
| %a = memref.load %0[%arg0, %arg2] : memref<1024x1024xf16> | ||
| %b = memref.load %1[%arg2, %arg1] : memref<1024x1024xf16> | ||
| %c = arith.mulf %a, %b : f16 | ||
| %cc = arith.extf %c : f16 to f32 | ||
| %ccc = arith.addf %cc, %arg3 : f32 | ||
| scf.yield %ccc : f32 | ||
| } | ||
| memref.store %res, %ref[%arg0, %arg1] : memref<1024x1024xf32> | ||
| } | ||
| } | ||
|
|
||
| %2 = call @test(%0, %1) : (memref<1024x1024xf16>, memref<1024x1024xf16>) -> memref<1024x1024xf32> | ||
| %cast = memref.cast %2 : memref<1024x1024xf32> to memref<*xf32> | ||
| //call @printMemrefF32(%cast) : (memref<*xf32>) -> () | ||
| %cast_ref = memref.cast %ref : memref<1024x1024xf32> to memref<*xf32> | ||
| //call @printMemrefF32(%cast_ref) : (memref<*xf32>) -> () | ||
| // CHECK: [ALLCLOSE: TRUE] | ||
| call @printAllcloseF32(%cast, %cast_ref) : (memref<*xf32>, memref<*xf32>) -> () | ||
| return | ||
| } | ||
| func.func private @printMemrefF32(memref<*xf32>) attributes {llvm.emit_c_interface} | ||
| func.func private @printAllcloseF32(memref<*xf32>, memref<*xf32>) attributes {llvm.emit_c_interface} | ||
| } |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters