Skip to content

debug.md

Latest commit

 

History

History
83 lines (65 loc) · 2.61 KB

debug.md

File metadata and controls

83 lines (65 loc) · 2.61 KB
Raw

cutf Debug Functions Reference

fp namespace

cutf::debug::fp::*
  • bitstring_t : This structure contains same size integer type of a given floating point type.

Example

using bitstring_t = typename cutf::debug::fp::bitstring_t<half>::type;
// bitstring_t = uint16_t

print namespace

cutf::debug::print::*

floating point (debug/fp.hpp)

  • print_bin : This function prints bitstring of an input variable
  • print_hex : This function prints hex-code of an input variable

Supported types are below.

  • Integer : uint8_t, uint16_t, uint32_t, uint64_t
  • Floating point : half, float, double

matrix (debug/matrix.hpp)

  • print_matrix : This function prints a matrix.
  • print_matrix_from_host : This function prints a matrix on the device memory from host code.
  • print_matrix_debug : This function prints a matrix in hex.
  • print_matrix_hex_from_host : This function prints a matrix on the device memory from host code in hex.
  • print_numpy_matrix : This function prints a matrix as a numpy matrix format.
  • print_numpy_matrix_from_host : This function prints a matrix as a numpy matrix format on the device memory from host code.

Example

cutf::debug::print::print_matrix(mat_ptr, M, N, ldm, "mat_a");
cutf::debug::print::print_numpy_matrix(mat_ptr, M, N, ldm, "mat_a");

fragment (debug/fragment.hpp)

  • print_fragment : This function prints each element of a fragment.

Example

nvcuda::wmma::fragment<nvcuda::wmma::matrix_a, N, N, N, half, nvcuda::wmma::col_major> fragment;
nvcuda::wmma::load_matrix_sync(fragment, mat, N);
cutf::debug::print::print_fragment(fragment, "frag");

time_breakdown

Sample code

cutf::debug::time_breakdown::profiler profiler(cuda_stream);

// put some operations between start and stop functions
profiler.start_timer_sync("malloc_a");
cudaMalloc(...);
profiler.stop_timer_sync("malloc_a");

// give some operations to measure function as a lambda function
profiler.measure("malloc_b", [&](){cudaMalloc(...);});

// output time breakdown
profiler.print_result();

Sample result

# cutf time breakdown result (Total:      3.396 [ms])
        Name    Total [ms]                    N   Avg [ms]   Min [ms]   Max [ms]
  cudaMemcpy         2.225 ( 65.52%)          1      2.225      2.225      2.225
add_1_kernel         1.070 ( 31.52%)        100      0.011      0.010      0.015
  cudaMalloc         0.081 (  2.39%)          1      0.081      0.081      0.081
 init_kernel         0.019 (  0.57%)          1      0.019      0.019      0.019

clock_breakdown

See clock_breakdown.