Fix performance regression (#5559) (#5567)

Fixes issue #5553 and
#5518

---------


(cherry picked from commit 96b9f83)


Signed-off-by: Ettore Tiotto <[email protected]>
Co-authored-by: Ettore Tiotto <[email protected]>

Author: anmyachev

python/test/unit/language/test_matmul.py

@@ -1224,10 +1224,8 @@ def test_mxfp8_mxfp4_matmul(M, N, K, BLOCK_M, BLOCK_N, BLOCK_K, NUM_STAGES, B_TR
             pytest.xfail("None scale has not been tested on XPU backend")
         if not (A_DATA_TYPE == "float8e5" and B_DATA_TYPE == "float4"):
             pytest.xfail(f"(A: {A_DATA_TYPE}, B: {B_DATA_TYPE}) has not been tested on XPU backend")
-        if (BLOCK_M, BLOCK_N,
-                BLOCK_K) == (128, 256,
-                             256) and CONST_SCALE and triton.runtime.driver.active.utils.get_device_properties(
-                                 triton.runtime.driver.active.get_current_device())["max_shared_mem"] < 196608:
+        if ((BLOCK_M, BLOCK_N, BLOCK_K) == (128, 256, 256) and triton.runtime.driver.active.utils.get_device_properties(
+                triton.runtime.driver.active.get_current_device())["max_shared_mem"] < 196608):
             pytest.xfail("XPU: Not enough shared memory")
     if not PACK_B_ALONG_K and B_DATA_TYPE != "float4":
         pytest.xfail("Pack along K can only be False for float4")

third_party/intel/lib/TritonIntelGPUTransforms/RemoveLayoutConversions.cpp

@@ -1646,7 +1646,11 @@ void LayoutRematerialization::hoistConvertDotOperand() {
 void LayoutRematerialization::hoistConvertDotOperand(
     ConvertLayoutOp convertOp) {
   auto targetType = convertOp.getType();
-  // The pass is targeted to MMA dot operands
+
+  // The pass is targeted to NVidia.
+  auto dotEnc = dyn_cast<DotOperandEncodingAttr>(targetType.getEncoding());
+  if (!(dotEnc && isa<NvidiaMmaEncodingAttr>(dotEnc.getParent())))
+    return;
 
   auto canBePipelined = [&](ConvertLayoutOp convertOp) {
     // FIXME: Check that the parent is a for loop