问题
I have used the following command to profile my Python code :
python2.7 -m cProfile -o X2_non_flat_multiprocessing_dummy.prof X2_non_flat.py
Then, I can visualize globally the repartition of different greedy functions :
As you can see, a lot of time is spent into Pobs_C and interpolate routine which corresponds to the following code snippet :
def Pobs_C(z, zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, C_IAp, A_IAp, n_IAp, B_IAp, E_T, R_T, DG_T_fid, DG_T, WGT_T, WT_T, WIAT_T, cl, P_dd_spec, RT500):
cc = 0
P_dd_ok = np.zeros(len(z_pk))
while cc < len(z_pk):
if ((cl+0.5)/RT500[cc] < 35 and (cl+0.5)/RT500[cc] > 0.0005):
P_dd_ok[cc] = P_dd_spec[cc]((cl+0.5)/RT500[cc])
cc=cc+1
P_dd_ok = CubicSpline(z_pk, P_dd_ok)
if paramo == 8:
P_dd_ok = P_dd_ok(z)*(DG_T(z)/DG_T_fid(z))**2
else:
P_dd_ok = P_dd_ok(z)
if paramo != 9 or paramo != 10 or paramo != 11:
C_gg = c/(100.*h_p)*0.5*delta_zpm*np.sum((F_dd_GG(z[1:], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, E_T(z[1:]), R_T(z[1:]), WGT_T[aa][1:], WGT_T[bb][1:], DG_T(z[1:]), P_dd_ok[1:]) + F_dd_GG(z[:-1], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, E_T(z[:-1]), R_T(z[:-1]), WGT_T[aa][:-1], WGT_T[bb][:-1], DG_T(z[:-1]), P_dd_ok[:-1]))) + P_shot_GC(zi, zj)
else:
C_gg = 0.
if paramo < 12:
C_ee = c/(100.*h_p)*0.5*delta_zpm*(np.sum(F_dd_LL(z[1:], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, E_T(z[1:]), R_T(z[1:]), WT_T[aa][1:], WT_T[bb][1:], DG_T(z[1:]), P_dd_ok[1:]) + F_dd_LL(z[:-1], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, E_T(z[:-1]), R_T(z[:-1]), WT_T[aa][:-1], WT_T[bb][:-1], DG_T(z[:-1]), P_dd_ok[:-1])) + np.sum(F_IA_d(z[1:], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, C_IAp, A_IAp, n_IAp, B_IAp, E_T(z[1:]), R_T(z[1:]), DG_T(z[1:]), WT_T[aa][1:], WT_T[bb][1:], WIAT_T[aa][1:], WIAT_T[bb][1:], P_dd_ok[1:]) + F_IA_d(z[:-1], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, C_IAp, A_IAp, n_IAp, B_IAp, E_T(z[:-1]), R_T(z[:-1]), DG_T(z[:-1]), WT_T[aa][:-1], WT_T[bb][:-1], WIAT_T[aa][:-1], WIAT_T[bb][:-1], P_dd_ok[:-1])) + np.sum(F_IAIA(z[1:], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, C_IAp, A_IAp, n_IAp, B_IAp, E_T(z[1:]), R_T(z[1:]), DG_T(z[1:]), WIAT_T[aa][1:], WIAT_T[bb][1:], P_dd_ok[1:]) + F_IAIA(z[:-1], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, C_IAp, A_IAp, n_IAp, B_IAp, E_T(z[:-1]), R_T(z[:-1]), DG_T(z[:-1]), WIAT_T[aa][:-1], WIAT_T[bb][:-1], P_dd_ok[:-1]))) + P_shot_WL(zi, zj)
else:
C_ee = 0.
C_gl = c/(100.*h_p)*0.5*delta_zpm*np.sum((F_dd_GL(z[1:], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, C_IAp, A_IAp, n_IAp, B_IAp, E_T(z[1:]), R_T(z[1:]), DG_T(z[1:]), WGT_T[aa][1:], WT_T[bb][1:], WIAT_T[bb][1:], P_dd_ok[1:]) + F_dd_GL(z[:-1], zi, zj, h_p, wm_p, wDE_p, w0_p, wa_p, C_IAp, A_IAp, n_IAp, B_IAp, E_T(z[:-1]), R_T(z[:-1]), DG_T(z[:-1]), WGT_T[aa][:-1], WT_T[bb][:-1], WIAT_T[bb][:-1], P_dd_ok[:-1])))
return C_gg, C_ee, C_gl
1) MAIN QUESTION : Is there a way to implement a GPU/OpenCL layer in this routine, especially for CubicSpline or the whole Pobs_C function.
What are the alternatives that would allow me to reduce the time passed into Pobs_C and its inner function CubicSpline ?
I have few notions with OpenCL (not PyOpenCL) like for example the map-reduce method or solving Heat 2D equation with classical kernel.
2) PREVIOUS FEEDBACK :
I know that we can't have optimization by thinking naively that calling an extern function inside a kernel brings a higher speedup since GPU can achieve a lot of calls. Instead, I should rather put all the content of the different functions allow to get optimization : Do you agree with that and confirm it ? So, can I declare inside the kernel code a call to an extern function (I mean a function not inside kernel, i.e the classical part code (called Host code ?) ?
3) OPTIONAL QUESTION : Maybe can I declare this extern function inside the kernel : is it possible by doing explicitely this declaration inside ? Indeed, that could avoid to copy all the content of all functions potentially GPU-parallilzed.
Any idea/remark/advice would be great, Regards.
PS: Sorry if this is a general topic but it will allow me to see clearer about the available ways to include GPU/OpenCL in my code above and then optimize it.
回答1:
- Is there a way to implement a GPU/OpenCL layer in this routine, especially for CubicSpline or the whole Pobs_C function
In all probability, no. The majority of time in the profiling seems to be in 12 million polynomial evaluations, and each evaluation call is only taking 6 microseconds on the CPU. It is unclear whether there would be significant embarrassing parallelism to expose in that operation. And GPUs are only useful for performing embarrassingly parallel tasks.
- So, can I declare inside the kernel code a call to an extern function (I mean a function not inside kernel, i.e the classical part code (called Host code ?) ?
No. That is impossible. And it is difficult to fathom what benefit that could possibly impart given that the Python code has to run on the host CPU anyway.
- Maybe can I declare this extern function inside the kernel : is it possible by doing explicitely[sic] this declaration inside ?
No.
来源：https://stackoverflow.com/questions/59024244/different-ways-to-optimize-with-gpu-pyopencl-a-python-code-extern-function-ins