ExecuteNetwork profiling

[prabhakarlad]

Hello all,

I have been trying to run profiling with ExecuteNetwork (armnn version 21.02), when I run with GPU backend I do see the timer values for example:

            "NeonConvolution2dWorkload_Execute_#8": {
                "type": "Event",
                "NeonKernelTimer/0: NEIm2ColKernel_#8": {
                    "type": "Measurement",
                    "raw": [
                        1458.791000
                    ],
                    "unit": "us"
                },
                "NeonKernelTimer/1: NEGEMMAssemblyWrapperKernel/a64_smallK_hybrid_u8u32_dot_8x4_#8": {
                    "type": "Measurement",
                    "raw": [
                        7416.292000
                    ],
                    "unit": "us"
                },
                "Wall clock time_#8": {
                    "type": "Measurement",
                    "raw": [
                        9041.875000
                    ],
                    "unit": "us"
                },
                "Wall clock time (Start)_#8": {
                    "type": "Measurement",
                    "raw": [
                        134619045.564000
                    ],
                    "unit": "us"
                },
                "Wall clock time (Stop)_#8": {
                    "type": "Measurement",
                    "raw": [
                        134628087.439000
                    ],
                    "unit": "us"
                }
            },

For the GPU backend I always get the values as 00.0

            "ClConvolution2dWorkload_Execute_#11": {
                "type": "Event",
                "OpenClKernelTimer/0: gemmlowp_matrix_a_reduction GWS[112,112,1]_#11": {
                    "type": "Measurement",
                    "raw": [
                        0.000000
                    ],
                    "unit": "us"
                },
                "OpenClKernelTimer/1: gemmlowp_mm_reshaped_only_rhs_t_fused_output_stage_fixedpoint GWS[32,32,112]_#11": {
                    "type": "Measurement",
                    "raw": [
                        0.000000
                    ],
                    "unit": "us"
                },
                "Wall clock time_#11": {
                    "type": "Measurement",
                    "raw": [
                        411.333000
                    ],
                    "unit": "us"
                },
                "Wall clock time (Start)_#11": {
                    "type": "Measurement",
                    "raw": [
                        300657261.851000
                    ],
                    "unit": "us"
                },
                "Wall clock time (Stop)_#11": {
                    "type": "Measurement",
                    "raw": [
                        300657673.184000
                    ],
                    "unit": "us"
                }
            },

Is there additional flags required while building the DDK/armnn for the GPU backend ?

Cheers, Prabhakar

[matthewsloyanARM]

Hi @prabhakarlad,

Thanks for reaching out. The GPU profiling wasn't being enabled correctly in Arm NN version 21.02. A patch can be found here to fix this which is on the master branch, would you be able try this out and see if this fixes your issue? Thank you!

Kind regards,

Matthew

[prabhakarlad]

Thank you @matthewsloyanARM, I am now able get the timings. (patch doesn't apply cleanly on 21.02 release so I had to manually apply the changes)