Backward weight v4r4r2 with xdlops

[ltqin]

This PR is for backward weight with xdlops, data layout is nchw

[asroy]

@ltqin There is not atomic-add yet. Do you want to add it in this PR, or in future PR?

[ltqin]

@ltqin There is not atomic-add yet. Do you want to add it in this PR, or in future PR?

will be another PR

[zjing14]

@ltqin did you test both fp32 and fp16?

[zjing14]

@ltqin You may merge the latest develop branch and test, since there are couples of changes that affect hacks.

[ltqin]

@ltqin There is not atomic-add yet. Do you want to add it in this PR, or in future PR?

I want add it in future PR

[ltqin]

@ltqin did you test both fp32 and fp16?

yes,using follow code:
#if 0
using in_data_t = float;
using acc_data_t = float;
using out_data_t = float;
#elif 1
using in_data_t = half_t;
using acc_data_t = float;
using out_data_t = half_t;
#elif 1
using in_data_t = int8_t;
using acc_data_t = int32_t;
using out_data_t = int8_t;
#endif

[ltqin]

@ltqin You may merge the latest develop branch and test, since there are couples of changes that affect hacks.

done

[ltqin]

@ltqin You may merge the latest develop branch and test, since there are couples of changes that affect hacks.

done

[asroy]

I tested this PR and found failure, using commit b2ea9aa

./host/driver_offline/conv_wrw_driver_offline  0 0 1 4 0 1   128  256  128 3 3  14   14     1 1       1 1      1 1       1 1

layout: 0
in: dim 4, lengths {128, 128, 14, 14}, strides {25088, 196, 14, 1}
wei: dim 4, lengths {256, 128, 3, 3}, strides {1152, 9, 3, 1}
out: dim 4, lengths {128, 256, 14, 14}, strides {50176, 196, 14, 1}
InLeftPads size 2, {1, 1, }
InRightPads size 2, {1, 1, }
ConvStrides size 2, {1, 1, }
ConvDilations size 2, {1, 1, }
device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw
a_k0_m_k1_grid_desc{3136, 256, 8}
b_k0_n_k1_grid_desc{3136, 1152, 8}
c_m_n_grid_desc{ 256, 1152}
launch_and_time_kernel: grid_dim {18, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Average time : 3.41418 ms, 4.33413 TFlop/s
a_k0_m_k1_grid_desc{3136, 256, 8}
b_k0_n_k1_grid_desc{3136, 1152, 8}
c_m_n_grid_desc{ 256, 1152}
Average time : 3.39738 ms, 4.35556 TFlop/s
error: 2.8372e+07
max_diff: 843, 7786, 6943

@ltqin

[asroy]

@ltqin I fixed some review comments (adding comment, using Number<> instead of index_t, etc)

[ltqin]

I tested this PR and found failure, using commit b2ea9aa

./host/driver_offline/conv_wrw_driver_offline  0 0 1 4 0 1   128  256  128 3 3  14   14     1 1       1 1      1 1       1 1
...
@ltqin

@asroy change vector load = 4, so config's ho*wo must be a multiple of 4