[Mlir-commits] [mlir] Add E2E test cases to test fastmath attribute. (PR #165423)
Md Abdullah Shahneous Bari
llvmlistbot at llvm.org
Thu Oct 30 08:35:01 PDT 2025
https://github.com/mshahneo updated https://github.com/llvm/llvm-project/pull/165423
>From 14dbc7519da688f2b8d75eb0c6260bd8988c6d3f Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Tue, 28 Oct 2025 15:56:37 +0000
Subject: [PATCH 1/2] Add E2E test cases to test fastmath attribute.
Ops tested:
- math.exp
- arith.maximumf
---
.../Dialect/XeGPU/WG/arith_ops.mlir | 116 ++++++++++++++++++
.../Dialect/XeGPU/WG/math_exp.mlir | 109 ++++++++++++++++
2 files changed, 225 insertions(+)
create mode 100644 mlir/test/Integration/Dialect/XeGPU/WG/arith_ops.mlir
create mode 100644 mlir/test/Integration/Dialect/XeGPU/WG/math_exp.mlir
diff --git a/mlir/test/Integration/Dialect/XeGPU/WG/arith_ops.mlir b/mlir/test/Integration/Dialect/XeGPU/WG/arith_ops.mlir
new file mode 100644
index 0000000000000..fc1fa0104b6ec
--- /dev/null
+++ b/mlir/test/Integration/Dialect/XeGPU/WG/arith_ops.mlir
@@ -0,0 +1,116 @@
+// RUN: mlir-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=workgroup" \
+// RUN: | mlir-runner \
+// RUN: --shared-libs=%mlir_levelzero_runtime \
+// RUN: --shared-libs=%mlir_runner_utils \
+// RUN: --entry-point-result=void \
+// RUN: | FileCheck %s
+
+#map = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], inst_data = [8, 16]>
+module @gemm attributes {gpu.container_module} {
+ func.func @test_fast_math(%input1: memref<256x256xf32>, %input2: memref<256x256xf32>) -> (memref<256x256xf32>, memref<256x256xf32>) attributes {llvm.emit_c_interface} {
+ %c1 = arith.constant 1 : index
+ %c4 = arith.constant 4 : index
+ %c8 = arith.constant 8 : index
+ %c16 = arith.constant 16 : index
+ %c32 = arith.constant 32 : index
+ %c64 = arith.constant 64 : index
+ %c128 = arith.constant 128 : index
+ %c512 = arith.constant 512 : index
+ %input1_gpu = gpu.alloc () : memref<256x256xf32>
+ gpu.memcpy %input1_gpu, %input2 : memref<256x256xf32>, memref<256x256xf32>
+ %input2_gpu = gpu.alloc () : memref<256x256xf32>
+ gpu.memcpy %input2_gpu, %input2 : memref<256x256xf32>, memref<256x256xf32>
+ %result_gpu = gpu.alloc () : memref<256x256xf32>
+ %result_gpu_with_fastmath = gpu.alloc () : memref<256x256xf32>
+ // NOTE: Here we can't use [8, 64] wi threads following
+ // the SG thread layout of [8, 4]. Because runtime will linearize
+ // the x dimension first (we need y dimension to be linearized first).
+ // So just use linearized thread layout of [512, 1] wi threads.
+ gpu.launch_func @math_kernels::@gpu_maximumf blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>)
+ gpu.launch_func @math_kernels::@gpu_maximumf_with_fastmath blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>)
+
+ %result_host = memref.alloc() : memref<256x256xf32>
+ %result_host_with_fastmath = memref.alloc() : memref<256x256xf32>
+ gpu.memcpy %result_host, %result_gpu : memref<256x256xf32>, memref<256x256xf32>
+ gpu.dealloc %input_gpu : memref<256x256xf32>
+ gpu.dealloc %result_gpu : memref<256x256xf32>
+ return %result_host, %result_host_with_fastmath : memref<256x256xf32>, memref<256x256xf32>
+ }
+
+ gpu.module @math_kernels {
+ gpu.func @gpu_maximumf(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>) kernel {
+ %c256 = arith.constant 256 : index
+ %block_id_x = gpu.block_id x
+ %block_id_y = gpu.block_id y
+ %m = arith.muli %block_id_x, %c256 : index
+ %n = arith.muli %block_id_y, %c256 : index
+ %input_tdesc_1 = xegpu.create_nd_tdesc %input1_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ %input_val_1 = xegpu.load_nd %input_tdesc_1[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
+ %input_tdesc_2 = xegpu.create_nd_tdesc %input2_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ %input_val_2 = xegpu.load_nd %input_tdesc_2[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
+ %result_val = arith.maximumf %input_val_1, %input_val_2 : vector<256x256xf32>
+ %result_tdesc = xegpu.create_nd_tdesc %result_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ xegpu.store_nd %result_val, %result_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #map>
+ gpu.return
+ }
+
+ // Kernel with fastmath attribute
+ gpu.func @gpu_maximumf_with_fastmath(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>) kernel {
+ %c256 = arith.constant 256 : index
+ %block_id_x = gpu.block_id x
+ %block_id_y = gpu.block_id y
+ %m = arith.muli %block_id_x, %c256 : index
+ %n = arith.muli %block_id_y, %c256 : index
+ %input_tdesc_1 = xegpu.create_nd_tdesc %input1_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ %input_val_1 = xegpu.load_nd %input_tdesc_1[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
+ %input_tdesc_2 = xegpu.create_nd_tdesc %input2_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ %input_val_2 = xegpu.load_nd %input_tdesc_2[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
+ %result_val = arith.maximumf %input_val_1, %input_val_2 fastmath<fast> : vector<256x256xf32>
+ %result_tdesc = xegpu.create_nd_tdesc %result_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ xegpu.store_nd %result_val, %result_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #map>
+ gpu.return
+ }
+ }
+
+ func.func @main() attributes {llvm.emit_c_interface} {
+ %c0 = arith.constant 0 : index
+ %c1 = arith.constant 1 : index
+ %c2_f32 = arith.constant 2.2 : f32
+ %c256 = arith.constant 256 : index
+ %input_1 = memref.alloc() : memref<256x256xf32>
+ %input_2 = memref.alloc() : memref<256x256xf32>
+ %cpu_ref_result = memref.alloc() : memref<256x256xf32>
+
+ scf.for %arg0 = %c0 to %c256 step %c1 {
+ scf.for %arg1 = %c0 to %c256 step %c1 {
+ memref.store %c2_f32, %input_1[%arg0, %arg1] : memref<256x256xf32>
+ memref.store %c2_f32, %input_2[%arg0, %arg1] : memref<256x256xf32>
+ }
+ }
+
+ // Run CPU version
+ scf.for %arg0 = %c0 to %c256 step %c1 {
+ scf.for %arg1 = %c0 to %c256 step %c1 {
+ %val_1 = memref.load %input_1[%arg0, %arg1] : memref<256x256xf32>
+ %val_2 = memref.load %input_2[%arg0, %arg1] : memref<256x256xf32>
+ %res_val = arith.maximumf %val_1, %val_2 : f32
+ memref.store %res_val, %cpu_ref_result[%arg0, %arg1] : memref<256x256xf32>
+ }
+ }
+
+ // Run GPU version.
+ %gpu_result, %gpu_result_fastmath = call @test_fast_math(%input_1, %input_2) : (memref<256x256xf32>, memref<256x256xf32>) -> (memref<256x256xf32>, memref<256x256xf32>)
+ %gpu_result_cast = memref.cast %gpu_result : memref<256x256xf32> to memref<*xf32>
+ // CHECK: Unranked Memref base@ = 0x{{[0-9a-f]+}}
+ // CHECK-COUNT-256: [0, 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]
+ call @printMemrefF32(%gpu_result_cast) : (memref<*xf32>) -> ()
+
+ memref.dealloc %input_1 : memref<256x256xf32>
+ memref.dealloc %input_2 : memref<256x256xf32>
+ memref.dealloc %cpu_ref_result : memref<256x256xf32>
+ memref.dealloc %gpu_result : memref<256x256xf32>
+ memref.dealloc %gpu_result_fastmath : memref<256x256xf32>
+ return
+ }
+ func.func private @printMemrefF32(memref<*xf32>) attributes {llvm.emit_c_interface}
+}
diff --git a/mlir/test/Integration/Dialect/XeGPU/WG/math_exp.mlir b/mlir/test/Integration/Dialect/XeGPU/WG/math_exp.mlir
new file mode 100644
index 0000000000000..abbd4bd1422f6
--- /dev/null
+++ b/mlir/test/Integration/Dialect/XeGPU/WG/math_exp.mlir
@@ -0,0 +1,109 @@
+// RUN: mlir-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=workgroup" \
+// RUN: | mlir-runner \
+// RUN: --shared-libs=%mlir_levelzero_runtime \
+// RUN: --shared-libs=%mlir_runner_utils \
+// RUN: --entry-point-result=void \
+// RUN: | FileCheck %s
+
+#map = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], inst_data = [8, 16]>
+module @gemm attributes {gpu.container_module} {
+ func.func @test_fast_math(%input: memref<256x256xf32>) -> (memref<256x256xf32>, memref<256x256xf32>) attributes {llvm.emit_c_interface} {
+ %c1 = arith.constant 1 : index
+ %c4 = arith.constant 4 : index
+ %c8 = arith.constant 8 : index
+ %c16 = arith.constant 16 : index
+ %c32 = arith.constant 32 : index
+ %c64 = arith.constant 64 : index
+ %c128 = arith.constant 128 : index
+ %c512 = arith.constant 512 : index
+ %input_gpu = gpu.alloc () : memref<256x256xf32>
+ gpu.memcpy %input_gpu, %input : memref<256x256xf32>, memref<256x256xf32>
+ %result_gpu = gpu.alloc () : memref<256x256xf32>
+ %result_gpu_with_fastmath = gpu.alloc () : memref<256x256xf32>
+ // NOTE: Here we can't use [8, 64] wi threads following
+ // the SG thread layout of [8, 4]. Because runtime will linearize
+ // the x dimension first (we need y dimension to be linearized first).
+ // So just use linearized thread layout of [512, 1] wi threads.
+ gpu.launch_func @math_kernels::@gpu_exp blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>)
+ gpu.launch_func @math_kernels::@gpu_exp_with_fastmath blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input_gpu : memref<256x256xf32>, %result_gpu_with_fastmath : memref<256x256xf32>)
+
+ %result_host = memref.alloc() : memref<256x256xf32>
+ %result_host_with_fastmath = memref.alloc() : memref<256x256xf32>
+ gpu.memcpy %result_host, %result_gpu : memref<256x256xf32>, memref<256x256xf32>
+ gpu.dealloc %input_gpu : memref<256x256xf32>
+ gpu.dealloc %result_gpu : memref<256x256xf32>
+ return %result_host, %result_host_with_fastmath : memref<256x256xf32>, memref<256x256xf32>
+ }
+
+ gpu.module @math_kernels {
+ gpu.func @gpu_exp(%input_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>) kernel {
+ %c256 = arith.constant 256 : index
+ %block_id_x = gpu.block_id x
+ %block_id_y = gpu.block_id y
+ %m = arith.muli %block_id_x, %c256 : index
+ %n = arith.muli %block_id_y, %c256 : index
+ %input_tdesc = xegpu.create_nd_tdesc %input_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ %input_val = xegpu.load_nd %input_tdesc[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
+ %result_val = math.exp %input_val : vector<256x256xf32>
+ %result_tdesc = xegpu.create_nd_tdesc %result_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ xegpu.store_nd %result_val, %result_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #map>
+ gpu.return
+ }
+
+ // Kernel with fastmath attribute
+ gpu.func @gpu_exp_with_fastmath(%input_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>) kernel {
+ %c256 = arith.constant 256 : index
+ %block_id_x = gpu.block_id x
+ %block_id_y = gpu.block_id y
+ %m = arith.muli %block_id_x, %c256 : index
+ %n = arith.muli %block_id_y, %c256 : index
+ %input_tdesc = xegpu.create_nd_tdesc %input_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ %input_val = xegpu.load_nd %input_tdesc[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
+ %result_val = math.exp %input_val fastmath<fast> : vector<256x256xf32>
+ %result_tdesc = xegpu.create_nd_tdesc %result_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ xegpu.store_nd %result_val, %result_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #map>
+ gpu.return
+ }
+ }
+
+ func.func @main() attributes {llvm.emit_c_interface} {
+ %c0 = arith.constant 0 : index
+ %c1 = arith.constant 1 : index
+ %c2_f32 = arith.constant 2.2 : f32
+ %c256 = arith.constant 256 : index
+ %input = memref.alloc() : memref<256x256xf32>
+ %input_ref = memref.alloc() : memref<256x256xf32>
+ %cpu_ref_result = memref.alloc() : memref<256x256xf32>
+
+ scf.for %arg0 = %c0 to %c256 step %c1 {
+ scf.for %arg1 = %c0 to %c256 step %c1 {
+ memref.store %c2_f32, %input[%arg0, %arg1] : memref<256x256xf32>
+ memref.store %c2_f32, %input_ref[%arg0, %arg1] : memref<256x256xf32>
+ }
+ }
+
+ // Run CPU version
+ scf.for %arg0 = %c0 to %c256 step %c1 {
+ scf.for %arg1 = %c0 to %c256 step %c1 {
+ %val = memref.load %input_ref[%arg0, %arg1] : memref<256x256xf32>
+ %res_val = math.exp %val : f32
+ memref.store %res_val, %cpu_ref_result[%arg0, %arg1] : memref<256x256xf32>
+ }
+ }
+
+ // Run GPU version.
+ %gpu_result, %gpu_result_fastmath = call @test_fast_math(%input) : (memref<256x256xf32>) -> (memref<256x256xf32>, memref<256x256xf32>)
+ %gpu_result_cast = memref.cast %gpu_result : memref<256x256xf32> to memref<*xf32>
+ // CHECK: Unranked Memref base@ = 0x{{[0-9a-f]+}}
+ // CHECK-COUNT-256: [0, 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]
+ call @printMemrefF32(%gpu_result_cast) : (memref<*xf32>) -> ()
+
+ memref.dealloc %input : memref<256x256xf32>
+ memref.dealloc %input_ref : memref<256x256xf32>
+ memref.dealloc %cpu_ref_result : memref<256x256xf32>
+ memref.dealloc %gpu_result : memref<256x256xf32>
+ memref.dealloc %gpu_result_fastmath : memref<256x256xf32>
+ return
+ }
+ func.func private @printMemrefF32(memref<*xf32>) attributes {llvm.emit_c_interface}
+}
>From 36f193212846750f7d3a9b1935fc05c10987cc62 Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Thu, 30 Oct 2025 15:34:13 +0000
Subject: [PATCH 2/2] Fix a issue.
Currently, only one kernel per-module is supported.
---
.../Dialect/XeGPU/WG/arith_ops.mlir | 16 +++++++++-----
.../Dialect/XeGPU/WG/math_exp.mlir | 22 ++++++++++++-------
2 files changed, 24 insertions(+), 14 deletions(-)
diff --git a/mlir/test/Integration/Dialect/XeGPU/WG/arith_ops.mlir b/mlir/test/Integration/Dialect/XeGPU/WG/arith_ops.mlir
index fc1fa0104b6ec..c3d9c69d26b4b 100644
--- a/mlir/test/Integration/Dialect/XeGPU/WG/arith_ops.mlir
+++ b/mlir/test/Integration/Dialect/XeGPU/WG/arith_ops.mlir
@@ -26,18 +26,20 @@ module @gemm attributes {gpu.container_module} {
// the SG thread layout of [8, 4]. Because runtime will linearize
// the x dimension first (we need y dimension to be linearized first).
// So just use linearized thread layout of [512, 1] wi threads.
- gpu.launch_func @math_kernels::@gpu_maximumf blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>)
- gpu.launch_func @math_kernels::@gpu_maximumf_with_fastmath blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>)
+ gpu.launch_func @arith_maximumf_module::@gpu_maximumf blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>)
+ gpu.launch_func @arith_maximumf_fastmath_module::@gpu_maximumf_with_fastmath blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu_with_fastmath : memref<256x256xf32>)
%result_host = memref.alloc() : memref<256x256xf32>
%result_host_with_fastmath = memref.alloc() : memref<256x256xf32>
gpu.memcpy %result_host, %result_gpu : memref<256x256xf32>, memref<256x256xf32>
- gpu.dealloc %input_gpu : memref<256x256xf32>
+ gpu.memcpy %result_host_with_fastmath, %result_gpu_with_fastmath : memref<256x256xf32>, memref<256x256xf32>
+ gpu.dealloc %input1_gpu : memref<256x256xf32>
+ gpu.dealloc %input2_gpu : memref<256x256xf32>
gpu.dealloc %result_gpu : memref<256x256xf32>
return %result_host, %result_host_with_fastmath : memref<256x256xf32>, memref<256x256xf32>
}
- gpu.module @math_kernels {
+ gpu.module @arith_maximumf_module {
gpu.func @gpu_maximumf(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>) kernel {
%c256 = arith.constant 256 : index
%block_id_x = gpu.block_id x
@@ -48,12 +50,14 @@ module @gemm attributes {gpu.container_module} {
%input_val_1 = xegpu.load_nd %input_tdesc_1[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
%input_tdesc_2 = xegpu.create_nd_tdesc %input2_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
%input_val_2 = xegpu.load_nd %input_tdesc_2[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
- %result_val = arith.maximumf %input_val_1, %input_val_2 : vector<256x256xf32>
+ %result_val = arith.maximumf %input_val_1, %input_val_2 {layout_result_0 = #map} : vector<256x256xf32>
%result_tdesc = xegpu.create_nd_tdesc %result_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
xegpu.store_nd %result_val, %result_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #map>
gpu.return
}
+ }
+ gpu.module @arith_maximumf_fastmath_module {
// Kernel with fastmath attribute
gpu.func @gpu_maximumf_with_fastmath(%input1_gpu : memref<256x256xf32>, %input2_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>) kernel {
%c256 = arith.constant 256 : index
@@ -65,7 +69,7 @@ module @gemm attributes {gpu.container_module} {
%input_val_1 = xegpu.load_nd %input_tdesc_1[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
%input_tdesc_2 = xegpu.create_nd_tdesc %input2_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
%input_val_2 = xegpu.load_nd %input_tdesc_2[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
- %result_val = arith.maximumf %input_val_1, %input_val_2 fastmath<fast> : vector<256x256xf32>
+ %result_val = arith.maximumf %input_val_1, %input_val_2 fastmath<fast> {layout_result_0 = #map} : vector<256x256xf32>
%result_tdesc = xegpu.create_nd_tdesc %result_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
xegpu.store_nd %result_val, %result_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #map>
gpu.return
diff --git a/mlir/test/Integration/Dialect/XeGPU/WG/math_exp.mlir b/mlir/test/Integration/Dialect/XeGPU/WG/math_exp.mlir
index abbd4bd1422f6..91a058c8ffe8f 100644
--- a/mlir/test/Integration/Dialect/XeGPU/WG/math_exp.mlir
+++ b/mlir/test/Integration/Dialect/XeGPU/WG/math_exp.mlir
@@ -24,18 +24,20 @@ module @gemm attributes {gpu.container_module} {
// the SG thread layout of [8, 4]. Because runtime will linearize
// the x dimension first (we need y dimension to be linearized first).
// So just use linearized thread layout of [512, 1] wi threads.
- gpu.launch_func @math_kernels::@gpu_exp blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>)
- gpu.launch_func @math_kernels::@gpu_exp_with_fastmath blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input_gpu : memref<256x256xf32>, %result_gpu_with_fastmath : memref<256x256xf32>)
+ gpu.launch_func @math_exp_module::@gpu_exp blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>)
+ gpu.launch_func @math_exp_fastmath_module::@gpu_exp_with_fastmath blocks in (%c1, %c1, %c1) threads in (%c512, %c1, %c1) args(%input_gpu : memref<256x256xf32>, %result_gpu_with_fastmath : memref<256x256xf32>)
%result_host = memref.alloc() : memref<256x256xf32>
%result_host_with_fastmath = memref.alloc() : memref<256x256xf32>
gpu.memcpy %result_host, %result_gpu : memref<256x256xf32>, memref<256x256xf32>
+ gpu.memcpy %result_host_with_fastmath, %result_gpu_with_fastmath : memref<256x256xf32>, memref<256x256xf32>
gpu.dealloc %input_gpu : memref<256x256xf32>
gpu.dealloc %result_gpu : memref<256x256xf32>
+ gpu.dealloc %result_gpu_with_fastmath : memref<256x256xf32>
return %result_host, %result_host_with_fastmath : memref<256x256xf32>, memref<256x256xf32>
}
- gpu.module @math_kernels {
+ gpu.module @math_exp_module {
gpu.func @gpu_exp(%input_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>) kernel {
%c256 = arith.constant 256 : index
%block_id_x = gpu.block_id x
@@ -44,23 +46,27 @@ module @gemm attributes {gpu.container_module} {
%n = arith.muli %block_id_y, %c256 : index
%input_tdesc = xegpu.create_nd_tdesc %input_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
%input_val = xegpu.load_nd %input_tdesc[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
- %result_val = math.exp %input_val : vector<256x256xf32>
+ %result_val = math.exp %input_val {layout_result_0 = #map} : vector<256x256xf32>
%result_tdesc = xegpu.create_nd_tdesc %result_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
xegpu.store_nd %result_val, %result_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #map>
gpu.return
}
+
+ }
+
+ gpu.module @math_exp_fastmath_module {
// Kernel with fastmath attribute
- gpu.func @gpu_exp_with_fastmath(%input_gpu : memref<256x256xf32>, %result_gpu : memref<256x256xf32>) kernel {
+ gpu.func @gpu_exp_with_fastmath(%input_gpu_with_fast_math : memref<256x256xf32>, %result_gpu_with_fastmath : memref<256x256xf32>) kernel {
%c256 = arith.constant 256 : index
%block_id_x = gpu.block_id x
%block_id_y = gpu.block_id y
%m = arith.muli %block_id_x, %c256 : index
%n = arith.muli %block_id_y, %c256 : index
- %input_tdesc = xegpu.create_nd_tdesc %input_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ %input_tdesc = xegpu.create_nd_tdesc %input_gpu_with_fast_math : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
%input_val = xegpu.load_nd %input_tdesc[%m, %n] : !xegpu.tensor_desc<256x256xf32, #map> -> vector<256x256xf32>
- %result_val = math.exp %input_val fastmath<fast> : vector<256x256xf32>
- %result_tdesc = xegpu.create_nd_tdesc %result_gpu : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
+ %result_val = math.exp %input_val fastmath<fast> {layout_result_0 = #map} : vector<256x256xf32>
+ %result_tdesc = xegpu.create_nd_tdesc %result_gpu_with_fastmath : memref<256x256xf32> -> !xegpu.tensor_desc<256x256xf32, #map>
xegpu.store_nd %result_val, %result_tdesc[%m, %n] : vector<256x256xf32>, !xegpu.tensor_desc<256x256xf32, #map>
gpu.return
}
More information about the Mlir-commits
mailing list