[LoadStoreOpToLLVM] Refactor block load lowering of tt.load with tensor pointer.

Signed-off-by: Lu,Chengjun <[email protected]>

Author: chengjunlu

python/test/unit/intel/test_block_store.py

@@ -133,27 +133,35 @@ def test_block_store(M, N, dtype_str, layout, block_ptr, device, tmp_path: pathl
     support_block_io = torch.xpu.get_device_capability()['has_subgroup_2d_block_io']
 
     if block_ptr:
+        load_ops = f"""
+            %src_ptr = tt.make_tensor_ptr %src, [%M_i64, %N_i64], [%N_i64, %c1_i64], [%c0_i32, %c0_i32] {{order = array<i32: 1, 0>}} : <tensor<{M}x{N}x{ty}, #layout>>
+            %store_val = tt.load %src_ptr {{boundaryCheck = array<i32: 0, 1>, padding = 1 : i32}} : !tt.ptr<tensor<{M}x{N}x{ty}, #layout>>
+            """
         store_ops = f"""
-            %M_i64 = arith.constant {M} : i64
-            %N_i64 = arith.constant {N} : i64
-            %c1_i64 = arith.constant 1 : i64
-            %c0_i32 = arith.constant 0 : i32
-
-            %blk_ptr = tt.make_tensor_ptr %dst, [%M_i64, %N_i64], [%N_i64, %c1_i64], [%c0_i32, %c0_i32] {{order = array<i32: 1, 0>}} : <tensor<{M}x{N}x{ty}, #layout>>
-            tt.store %blk_ptr, %store_val {{ttig.block_io = "row_major", boundaryCheck = array<i32: 0, 1>}} : !tt.ptr<tensor<{M}x{N}x{ty}, #layout>>
+            %dst_ptr = tt.make_tensor_ptr %dst, [%M_i64, %N_i64], [%N_i64, %c1_i64], [%c0_i32, %c0_i32] {{order = array<i32: 1, 0>}} : <tensor<{M}x{N}x{ty}, #layout>>
+            tt.store %dst_ptr, %store_val {{ttig.block_io = "row_major", boundaryCheck = array<i32: 0, 1>}} : !tt.ptr<tensor<{M}x{N}x{ty}, #layout>>
             """
     else:
+        load_ops = f"""
+            %src_base = tt.splat %src : !tt.ptr<{ty}> -> tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>
+            %src_ptr = tt.addptr %src_base, %8 : tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>, tensor<{M}x{N}xi32, #layout>
+            %store_val = tt.load %src_ptr {{ttig.block_io = "row_major"}} : tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>
+            """
         store_ops = f"""
-            %12 = tt.splat %dst : !tt.ptr<{ty}> -> tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>
-            %13 = tt.addptr %12, %8 : tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>, tensor<{M}x{N}xi32, #layout>
-            tt.store %13, %store_val {{ttig.block_io = "row_major"}} : tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>
+            %dst_base = tt.splat %dst : !tt.ptr<{ty}> -> tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>
+            %dst_ptr = tt.addptr %dst_base, %8 : tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>, tensor<{M}x{N}xi32, #layout>
+            tt.store %dst_ptr, %store_val {{ttig.block_io = "row_major"}} : tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>
             """
 
     ir = f"""
     #layout = {layout}
     module attributes {{{"ttig.support_sg_2d_block," if support_block_io else ""} "ttg.num-ctas" = 1 : i32, "ttg.num-warps" = {num_warps} : i32, ttg.target = "xpu", "ttg.threads-per-warp" = {threads_per_warp} : i32}} {{
         tt.func public @block_store(%src: !tt.ptr<{ty}> {{tt.divisibility = 16 : i32}}, %dst: !tt.ptr<{ty}> {{tt.divisibility = 16 : i32}}) {{
 
+            %M_i64 = arith.constant {M} : i64
+            %N_i64 = arith.constant {N} : i64
+            %c1_i64 = arith.constant 1 : i64
+            %c0_i32 = arith.constant 0 : i32
             %stride = arith.constant dense<{N}> : tensor<{M}x1xi32, #layout>
             %1 = tt.make_range {{end = {M} : i32, start = 0 : i32}} : tensor<{M}xi32, #ttg.slice<{{dim = 1, parent = #layout}}>>
             %2 = tt.expand_dims %1 {{axis = 1 : i32}} : tensor<{M}xi32, #ttg.slice<{{dim = 1, parent = #layout}}>> -> tensor<{M}x1xi32, #layout>
@@ -163,9 +171,7 @@ def test_block_store(M, N, dtype_str, layout, block_ptr, device, tmp_path: pathl
             %6 = tt.broadcast %3 : tensor<{M}x1xi32, #layout> -> tensor<{M}x{N}xi32, #layout>
             %7 = tt.broadcast %5 : tensor<1x{N}xi32, #layout> -> tensor<{M}x{N}xi32, #layout>
             %8 = arith.addi %6, %7 : tensor<{M}x{N}xi32, #layout>
-            %9 = tt.splat %src : !tt.ptr<{ty}> -> tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>
-            %10 = tt.addptr %9, %8 : tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>, tensor<{M}x{N}xi32, #layout>
-            %store_val = tt.load %10 : tensor<{M}x{N}x!tt.ptr<{ty}>, #layout>
+            {load_ops}
 
             {store_ops}
 
@@ -191,3 +197,5 @@ def test_block_store(M, N, dtype_str, layout, block_ptr, device, tmp_path: pathl
 
     if support_block_io:
         assert 'spirv_Subgroup2DBlockStoreINTEL' in kernel.asm['llir'] or 'GenISA.LSC2DBlockWrite' in kernel.asm['llir']
+        if not block_ptr:
+            assert 'spirv_Subgroup2DBlockLoad' in kernel.asm['llir'] or 'GenISA.LSC2DBlockRead' in kernel.asm['llir']