Change DPAS from OCL to SPV builtin (#3732)

SPEC:
[SPV_INTEL_subgroup_matrix_multiply_accumulate](https://github.khronos.org/SPIRV-Registry/extensions/INTEL/SPV_INTEL_subgroup_matrix_multiply_accumulate.html)

Benchmark CI:
https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/15145720901
(no geomean regression)

Closes #3731

Signed-off-by: Whitney Tsang <[email protected]>

Author: whitneywhtsang

test/Conversion/intel/tritongpu_to_gen.mlir

@@ -814,7 +814,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 1 : i32, "ttg.thr
     %BB_DOT = ttg.local_load %BB : !ttg.memdesc<16x16xf16, #shared0, #smem> -> tensor<16x16xf16, #dot_operand_b>
     %cst0 = arith.constant dense<0.000000e+00> : tensor<16x16xf32, #dpas0>
 
-    // CHECK-COUNT-2: llvm.call spir_funccc @_Z38intel_sub_group_f16_f16_matrix_mad_k16Dv8_sDv8_iDv8_f(%{{.*}}, %{{.*}}, %{{.*}}) {{.*}} : (vector<8xi16>, vector<8xi32>, vector<8xf32>) -> vector<8xf32>
+    // CHECK-COUNT-2: llvm.call spir_funccc @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv8_sDv8_iDv8_fi(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}} : (i32, vector<8xi16>, vector<8xi32>, vector<8xf32>, i32) -> vector<8xf32>
     %D = tt.dot %AA_DOT, %BB_DOT, %cst0 : tensor<16x16xf16, #dot_operand_a> * tensor<16x16xf16, #dot_operand_b> -> tensor<16x16xf32, #dpas0>
 
     tt.return
@@ -964,7 +964,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32} {
     %a_mat = ttg.local_load %a : !ttg.memdesc<128x32xf16, #shared, #smem> -> tensor<128x32xf16, #dot_operand_a>
     %b_mat = ttg.local_load %b : !ttg.memdesc<32x256xf16, #shared, #smem> -> tensor<32x256xf16, #dot_operand_b>
 
-    // CHECK-COUNT-128: llvm.call spir_funccc @_Z38intel_sub_group_f16_f16_matrix_mad_k16Dv8_sDv8_iDv8_f(%{{.*}}, %{{.*}}, %{{.*}}) {{.*}} : (vector<8xi16>, vector<8xi32>, vector<8xf32>) -> vector<8xf32>
+    // CHECK-COUNT-128: llvm.call spir_funccc @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv8_sDv8_iDv8_fi(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}} : (i32, vector<8xi16>, vector<8xi32>, vector<8xf32>, i32) -> vector<8xf32>
     %28 = tt.dot %a_mat, %b_mat, %cst : tensor<128x32xf16, #dot_operand_a> * tensor<32x256xf16, #dot_operand_b> -> tensor<128x256xf32, #dpas>
     %38 = ttg.convert_layout %28 : tensor<128x256xf32, #dpas> -> tensor<128x256xf32, #blocked>
 
@@ -991,7 +991,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32} {
     %a_mat = ttg.local_load %a : !ttg.memdesc<32x64xf16, #shared0, #smem> -> tensor<32x64xf16, #dot_operand_a>
     %b_mat = ttg.local_load %b : !ttg.memdesc<64x64xf16, #shared1, #smem> -> tensor<64x64xf16, #dot_operand_b>
 
-    // CHECK-COUNT-16: llvm.call spir_funccc @_Z38intel_sub_group_f16_f16_matrix_mad_k16Dv8_sDv8_iDv8_f(%{{.*}}, %{{.*}}, %{{.*}}) {{.*}} : (vector<8xi16>, vector<8xi32>, vector<8xf32>) -> vector<8xf32>
+    // CHECK-COUNT-16: llvm.call spir_funccc @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv8_sDv8_iDv8_fi(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}} : (i32, vector<8xi16>, vector<8xi32>, vector<8xf32>, i32) -> vector<8xf32>
     %28 = tt.dot %a_mat, %b_mat, %cst : tensor<32x64xf16, #dot_operand_a> * tensor<64x64xf16, #dot_operand_b> -> tensor<32x64xf32, #dpas>
     %38 = ttg.convert_layout %28 : tensor<32x64xf32, #dpas> -> tensor<32x64xf32, #blocked>
     %30 = tt.splat %ptr : !tt.ptr<f32> -> tensor<32x1x!tt.ptr<f32>, #blocked>
@@ -1040,7 +1040,7 @@ module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32} {
     %a_mat = ttg.local_load %a : !ttg.memdesc<32x16xf32, #shared, #smem> -> tensor<32x16xf32, #dot_operand_a>
     %b_mat = ttg.local_load %b : !ttg.memdesc<16x32xf32, #shared, #smem> -> tensor<16x32xf32, #dot_operand_b>
 
-    // CHECK-COUNT-2: llvm.call spir_funccc @_Z39intel_sub_group_tf32_tf32_matrix_mad_k8Dv8_fS_S_(%{{.*}}, %{{.*}}, %{{.*}}) {{.*}} : (vector<8xf32>, vector<8xf32>, vector<8xf32>) -> vector<8xf32>
+    // CHECK-COUNT-2: llvm.call spir_funccc @_Z45__spirv_SubgroupMatrixMultiplyAccumulateINTELiDv8_fS_S_i(%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}) {{.*}} : (i32, vector<8xf32>, vector<8xf32>, vector<8xf32>, i32) -> vector<8xf32>
     %28 = tt.dot %a_mat, %b_mat, %cst, inputPrecision = tf32 : tensor<32x16xf32, #dot_operand_a> * tensor<16x32xf32, #dot_operand_b> -> tensor<32x32xf32, #dpas>
     %38 = ttg.convert_layout %28 : tensor<32x32xf32, #dpas> -> tensor<32x32xf32, #blocked>