[test_mxfp8_mxfp4_matmul] The kernels compilation time is too long

[AndreyPavlenko]

Describe the issue

It hakes hours in my env to run the test, but after the first run, when the kernels are cached, the test runs much faster.

Environment details

main

[AndreyPavlenko]

The ?root cause? reproducer:

./build/cmake.linux-x86_64-cpython-3.12/bin/triton-opt -convert-triton-intel-gpu-to-llvm module_dump.txt > module.llir

This command produces a huge IR with 233087 lines of code and size 1.4G.

module_dump.txt is produced by the test test_matmul.py::test_mxfp8_mxfp4_matmul[0-True-True-float4-float4-True-True-True-3-128-128-64-1024-512-512]

For the tests with block size 256-128-128 it's much bigger and takes much more time.

[AndreyPavlenko]

Another reproducer created from the test python/test/unit/intel/test_mxfp_matmul.py::test_mxfp_matmul[True-True-float4-float4-True-True-1-128-128-128-1024-512-512]
module_dump2.txt

The command

./build/cmake.linux-x86_64-cpython-3.12/bin/triton-opt -convert-triton-intel-gpu-to-llvm module_dump2.txt > module2.llir

is running about 3 minutes and produces IR with 458075 lines of code and the size 5.6G.

[AndreyPavlenko]

Another issue - the binary built from this IR is loaded s few minutes and then fails with:

    def load_binary(self, *args):
...
>       return self.shared_library.load_binary(args)
E       RuntimeError: Triton Error [ZE]: 0x78000011

[AndreyPavlenko]

One more issue to be addressed is the ttg.fp4_to_fp pass. The following results in a huge IR, when lowered to llvm:

#blocked1 = #ttg.blocked<{sizePerThread = [1, 16], threadsPerWarp = [16, 2], warpsPerCTA = [4, 1], order = [1, 0]}>
#blocked2 = #ttg.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 32], warpsPerCTA = [4, 1], order = [1, 0]}>
#blocked3 = #ttg.blocked<{sizePerThread = [1, 2], threadsPerWarp = [1, 32], warpsPerCTA = [4, 1], order = [1, 0]}>
module attributes {triton_intel_gpu.min_sg_size = 16 : i32, triton_intel_gpu.support_bf16_conversion, triton_intel_gpu.support_dpas, triton_intel_gpu.support_sg_2d_block, triton_intel_gpu.target_arch = "spir64", ttg.global_scratch_memory_alignment = 1 : i32, ttg.global_scratch_memory_size = 0 : i32, "ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32, ttg.shared = 16384 : i32, ttg.target = "xpu", "ttg.threads-per-warp" = 32 : i32} {
  tt.func public @test(%arg0: tensor<128x32x!tt.ptr<i8>, #blocked1>) -> tensor<128x64xbf16, #blocked3> {
    %0 = tt.load %arg0 : tensor<128x32x!tt.ptr<i8>, #blocked1>
    %1 = ttg.convert_layout %0 {allocation.offset = 0 : i32} : tensor<128x32xi8, #blocked1> -> tensor<128x32xi8, #blocked2>
    %2 = ttg.fp4_to_fp %1 {axis = 1 : i32} : tensor<128x32xi8, #blocked2> -> tensor<128x64xbf16, #blocked3>
    tt.return %2: tensor<128x64xbf16, #blocked3>
  }
}

$ ./build/cmake.linux-x86_64-cpython-3.12/bin/triton-opt -convert-triton-intel-gpu-to-llvm test_ir.txt | wc -l
3253