def eval_idct2d(x, expk0, expk1, expkM, expkN, runs): y_N = discrete_spectral_transform.idct2_N(x, expk0=expk0, expk1=expk1) torch.cuda.synchronize() tt = time.time() for i in range(runs): y_N = discrete_spectral_transform.idct2_N(x, expk0=expk0, expk1=expk1) torch.cuda.synchronize() print("PyTorch idct2_N takes %.7f ms" % ((time.time()-tt)/runs*1000)) idct2func = dct.IDCT2(expk0, expk1, algorithm='2N') y_N = idct2func.forward(x) torch.cuda.synchronize() tt = time.time() for i in range(runs): y_N = idct2func.forward(x) torch.cuda.synchronize() print("IDCT2_2N Function takes %.7f ms" % ((time.time()-tt)/runs*1000)) idct2func = dct.IDCT2(expk0, expk1, algorithm='N') y_N = idct2func.forward(x) torch.cuda.synchronize() tt = time.time() # with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = idct2func.forward(x)/x.size(0)/x.size(1)/4 torch.cuda.synchronize() # print(prof) print("IDCT2_N Function takes %.7f ms" % ((time.time()-tt)/runs*1000)) dct2func = dct2_fft2.IDCT2(expkM, expkN) y = dct2func.forward(x) torch.cuda.synchronize() tt = time.time() for i in range(runs): y_test = dct2func.forward(x) torch.cuda.synchronize() print("IDCT2_FFT2 Function takes %.7f ms" % ((time.time()-tt)/runs*1000)) print("")
def eval_runtime(): #x = torch.tensor([1, 2, 7, 9, 20, 31], dtype=torch.float64) #print(dct_N(x)) N = 512 runs = 10 x = torch.empty(10, N, N, dtype=torch.float64).uniform_(0, 10.0).cuda() perm = discrete_spectral_transform.get_perm(N, dtype=torch.int64, device=x.device) expk = discrete_spectral_transform.get_expk(N, dtype=x.dtype, device=x.device) #x_numpy = x.data.cpu().numpy() #tt = time.time() #for i in range(runs): # y = fftpack.dct(fftpack.dct(x_numpy[i%10].T, norm=None).T, norm=None) #print("scipy takes %.3f sec" % (time.time()-tt)) ## 9s for 200 iterations 1024x1024 on GTX 1080 #torch.cuda.synchronize() #tt = time.time() ##with torch.autograd.profiler.profile(use_cuda=True) as prof: #for i in range(runs): # y_2N = dct2_2N(x[0], expk0=expk, expk1=expk) #torch.cuda.synchronize() ##print(prof) #print("dct_2N takes %.3f ms" % ((time.time()-tt)/runs*1000)) ## 11s for 200 iterations 1024x1024 on GTX 1080 #torch.cuda.synchronize() #tt = time.time() ##with torch.autograd.profiler.profile(use_cuda=True) as prof: #for i in range(runs): # y_N = discrete_spectral_transform.dct2_N(x[i%10], perm0=perm, expk0=expk, perm1=perm, expk1=expk) #torch.cuda.synchronize() ##print(prof) #print("dct_N takes %.3f ms" % ((time.time()-tt)/runs*1000)) #dct2func = dct.DCT2(expk, expk, algorithm='2N') #torch.cuda.synchronize() #tt = time.time() ##with torch.autograd.profiler.profile(use_cuda=True) as prof: #for i in range(runs): # y_N = dct2func.forward(x[0]) #torch.cuda.synchronize() ##print(prof) #print("DCT2Function 2N takes %.3f ms" % ((time.time()-tt)/runs*1000)) #dct2func = dct.DCT2(expk, expk, algorithm='N') #torch.cuda.synchronize() #tt = time.time() ##with torch.autograd.profiler.profile(use_cuda=True) as prof: #for i in range(runs): # y_N = dct2func.forward(x[0]) #torch.cuda.synchronize() ##print(prof) #print("DCT2Function takes %.3f ms" % ((time.time()-tt)/runs*1000)) #dct2func = dct_lee.DCT2(expk, expk) #torch.cuda.synchronize() #tt = time.time() ##with torch.autograd.profiler.profile(use_cuda=True) as prof: #for i in range(runs): # y_N = dct2func.forward(x[i%10]) #torch.cuda.synchronize() ##print(prof) #print("DCT2Function lee takes %.3f ms" % ((time.time()-tt)/runs*1000)) #torch.cuda.synchronize() #tt = time.time() ##with torch.autograd.profiler.profile(use_cuda=True) as prof: #for i in range(runs): # y_N = discrete_spectral_transform.idct2_2N(x[i%10], expk0=expk, expk1=expk) #torch.cuda.synchronize() ##print(prof) #print("idct2_2N takes %.3f ms" % ((time.time()-tt)/runs*1000)) idct2func = dct.IDCT2(expk, expk, algorithm='2N') torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = idct2func.forward(x[i%10]) torch.cuda.synchronize() #print(prof) print("IDCT2Function 2N takes %.3f ms" % ((time.time()-tt)/runs*1000)) idct2func = dct.IDCT2(expk, expk, algorithm='N') torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = idct2func.forward(x[i%10]) torch.cuda.synchronize() #print(prof) print("IDCT2Function takes %.3f ms" % ((time.time()-tt)/runs*1000)) exit() #torch.cuda.synchronize() #tt = time.time() ##with torch.autograd.profiler.profile(use_cuda=True) as prof: #for i in range(runs): # y_N = discrete_spectral_transform.idxt(x[i%10], 0, expk=expk) #torch.cuda.synchronize() ##print(prof) #print("idxt takes %.3f ms" % ((time.time()-tt)/runs*1000)) #idxct_func = dct.IDXCT(expk) #torch.cuda.synchronize() #tt = time.time() ##with torch.autograd.profiler.profile(use_cuda=True) as prof: #for i in range(runs): # y_N = idxct_func.forward(x[i%10]) #torch.cuda.synchronize() ##print(prof) #print("IDXCTFunction takes %.3f ms" % ((time.time()-tt)/runs*1000)) torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = torch.rfft(x[i%10].view([1, N, N]), signal_ndim=2, onesided=False) torch.cuda.synchronize() #print(prof) print("fft2 takes %.3f ms" % ((time.time()-tt)/runs*1000)) torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = discrete_spectral_transform.idcct2(x[i%10], expk_0=expk, expk_1=expk) torch.cuda.synchronize() #print(prof) print("idcct2 takes %.3f ms" % ((time.time()-tt)/runs*1000)) func = dct.IDCCT2(expk, expk) torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = func.forward(x[i%10]) torch.cuda.synchronize() #print(prof) print("IDCCT2Function takes %.3f ms" % ((time.time()-tt)/runs*1000)) torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = discrete_spectral_transform.idcst2(x[i%10], expk_0=expk, expk_1=expk) torch.cuda.synchronize() #print(prof) print("idcst2 takes %.3f ms" % ((time.time()-tt)/runs*1000)) func = dct.IDCST2(expk, expk) torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = func.forward(x[i%10]) torch.cuda.synchronize() #print(prof) print("IDCST2Function takes %.3f ms" % ((time.time()-tt)/runs*1000)) torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = discrete_spectral_transform.idsct2(x[i%10], expk_0=expk, expk_1=expk) torch.cuda.synchronize() #print(prof) print("idsct2 takes %.3f ms" % ((time.time()-tt)/runs*1000)) func = dct.IDSCT2(expk, expk) torch.cuda.synchronize() tt = time.time() #with torch.autograd.profiler.profile(use_cuda=True) as prof: for i in range(runs): y_N = func.forward(x[i%10]) torch.cuda.synchronize() #print(prof) print("IDSCT2Function takes %.3f ms" % ((time.time()-tt)/runs*1000))
def test_idct2Random(self): torch.manual_seed(10) M = 4 N = 8 x = torch.tensor(torch.empty(M, N, dtype=torch.int32).random_(0, 10), dtype=dtype) print("2D x") print(x) y = discrete_spectral_transform.dct2_2N(x) golden_value = discrete_spectral_transform.idct2_2N(y).data.numpy() print("2D golden_value") print(golden_value) # test cpu using N-FFT #pdb.set_trace() custom = dct.IDCT2(algorithm='N') dct_value = custom.forward(y) print("2D dct_value") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test cpu using 2N-FFT #pdb.set_trace() custom = dct.IDCT2(algorithm='2N') dct_value = custom.forward(y) print("2D dct_value") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test cpu using dct_lee #pdb.set_trace() custom = dct_lee.IDCT2() dct_value = custom.forward(y) print("2D dct_value") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test gpu custom = dct.IDCT2(algorithm='N') dct_value = custom.forward(y.cuda()).cpu() print("2D dct_value cuda") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test gpu custom = dct.IDCT2(algorithm='2N') dct_value = custom.forward(y.cuda()).cpu() print("2D dct_value cuda") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test gpu custom = dct_lee.IDCT2() dct_value = custom.forward(y.cuda()).cpu() print("2D dct_value cuda") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5)
def compare_different_methods(cuda_flag, M=1024, N=1024, dtype=torch.float64): density_map = torch.empty(M, N, dtype=dtype).uniform_(0, 10.0) if cuda_flag: density_map = density_map.cuda() expkM = discrete_spectral_transform.get_expk(M, dtype, density_map.device) expkN = discrete_spectral_transform.get_expk(N, dtype, density_map.device) exact_expkM = discrete_spectral_transform.get_exact_expk(M, dtype, density_map.device) exact_expkN = discrete_spectral_transform.get_exact_expk(N, dtype, density_map.device) print("M = {}, N = {}".format(M, N)) wu = torch.arange(M, dtype=density_map.dtype, device=density_map.device).mul(2 * np.pi / M).view([M, 1]) wv = torch.arange(N, dtype=density_map.dtype, device=density_map.device).mul(2 * np.pi / N).view([1, N]) wu2_plus_wv2 = wu.pow(2) + wv.pow(2) wu2_plus_wv2[0, 0] = 1.0 # avoid zero-division, it will be zeroed out inv_wu2_plus_wv2_2X = 2.0 / wu2_plus_wv2 inv_wu2_plus_wv2_2X[0, 0] = 0.0 wu_by_wu2_plus_wv2_2X = wu.mul(inv_wu2_plus_wv2_2X) wv_by_wu2_plus_wv2_2X = wv.mul(inv_wu2_plus_wv2_2X) # the first approach is used as the ground truth auv_golden = dct.dct2(density_map, expk0=expkM, expk1=expkN) auv = auv_golden.clone() auv[0, :].mul_(0.5) auv[:, 0].mul_(0.5) auv_by_wu2_plus_wv2_wu = auv.mul(wu_by_wu2_plus_wv2_2X) auv_by_wu2_plus_wv2_wv = auv.mul(wv_by_wu2_plus_wv2_2X) field_map_x_golden = dct.idsct2(auv_by_wu2_plus_wv2_wu, expkM, expkN) field_map_y_golden = dct.idcst2(auv_by_wu2_plus_wv2_wv, expkM, expkN) # compute potential phi # auv / (wu**2 + wv**2) auv_by_wu2_plus_wv2 = auv.mul(inv_wu2_plus_wv2_2X).mul_(2) #potential_map = discrete_spectral_transform.idcct2(auv_by_wu2_plus_wv2, expkM, expkN) potential_map_golden = dct.idcct2(auv_by_wu2_plus_wv2, expkM, expkN) # compute energy energy_golden = potential_map_golden.mul(density_map).sum() if density_map.is_cuda: torch.cuda.synchronize() # the second approach uses the idxst_idct and idct_idxst dct2 = dct2_fft2.DCT2(exact_expkM, exact_expkN) idct2 = dct2_fft2.IDCT2(exact_expkM, exact_expkN) idct_idxst = dct2_fft2.IDCT_IDXST(exact_expkM, exact_expkN) idxst_idct = dct2_fft2.IDXST_IDCT(exact_expkM, exact_expkN) inv_wu2_plus_wv2 = 1.0 / wu2_plus_wv2 inv_wu2_plus_wv2[0, 0] = 0.0 wu_by_wu2_plus_wv2_half = wu.mul(inv_wu2_plus_wv2).mul_(0.5) wv_by_wu2_plus_wv2_half = wv.mul(inv_wu2_plus_wv2).mul_(0.5) buv = dct2.forward(density_map) buv_by_wu2_plus_wv2_wu = buv.mul(wu_by_wu2_plus_wv2_half) buv_by_wu2_plus_wv2_wv = buv.mul(wv_by_wu2_plus_wv2_half) field_map_x = idxst_idct.forward(buv_by_wu2_plus_wv2_wu) field_map_y = idct_idxst.forward(buv_by_wu2_plus_wv2_wv) buv_by_wu2_plus_wv2 = buv.mul(inv_wu2_plus_wv2) potential_map = idct2.forward(buv_by_wu2_plus_wv2) energy = potential_map.mul(density_map).sum() if density_map.is_cuda: torch.cuda.synchronize() # compare results np.testing.assert_allclose(buv.data.cpu().numpy(), auv_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) np.testing.assert_allclose(field_map_x.data.cpu().numpy(), field_map_x_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) np.testing.assert_allclose(field_map_y.data.cpu().numpy(), field_map_y_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) np.testing.assert_allclose(potential_map.data.cpu().numpy(), potential_map_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) np.testing.assert_allclose(energy.data.cpu().numpy(), energy_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) # the third approach uses the dct.idxst_idct and dct.idxst_idct dct2 = dct.DCT2(expkM, expkN) idct2 = dct.IDCT2(expkM, expkN) idct_idxst = dct.IDCT_IDXST(expkM, expkN) idxst_idct = dct.IDXST_IDCT(expkM, expkN) cuv = dct2.forward(density_map) cuv_by_wu2_plus_wv2_wu = cuv.mul(wu_by_wu2_plus_wv2_half) cuv_by_wu2_plus_wv2_wv = cuv.mul(wv_by_wu2_plus_wv2_half) field_map_x = idxst_idct.forward(cuv_by_wu2_plus_wv2_wu) field_map_y = idct_idxst.forward(cuv_by_wu2_plus_wv2_wv) cuv_by_wu2_plus_wv2 = cuv.mul(inv_wu2_plus_wv2) potential_map = idct2.forward(cuv_by_wu2_plus_wv2) energy = potential_map.mul(density_map).sum() if density_map.is_cuda: torch.cuda.synchronize() # compare results np.testing.assert_allclose(cuv.data.cpu().numpy(), auv_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) np.testing.assert_allclose(field_map_x.data.cpu().numpy(), field_map_x_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) np.testing.assert_allclose(field_map_y.data.cpu().numpy(), field_map_y_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) np.testing.assert_allclose(potential_map.data.cpu().numpy(), potential_map_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5) np.testing.assert_allclose(energy.data.cpu().numpy(), energy_golden.data.cpu().numpy(), rtol=1e-6, atol=1e-5)
def test_idct2Random(self): torch.manual_seed(10) M = 4 N = 8 x = torch.empty(M, N, dtype=torch.int32).random_(0, 10).double() print("2D x") print(x) expkM = discrete_spectral_transform.get_exact_expk(M, dtype=x.dtype, device=x.device) expkN = discrete_spectral_transform.get_exact_expk(N, dtype=x.dtype, device=x.device) y = discrete_spectral_transform.dct2_2N(x) golden_value = discrete_spectral_transform.idct2_2N(y).data.numpy() print("2D idct golden_value") print(golden_value) # test cpu using N-FFT # pdb.set_trace() custom = dct.IDCT2(algorithm='N') dct_value = custom.forward(y) print("2D idct_value") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test cpu using 2N-FFT # pdb.set_trace() custom = dct.IDCT2(algorithm='2N') dct_value = custom.forward(y) print("2D idct_value") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test cpu using dct_lee # pdb.set_trace() custom = dct_lee.IDCT2() dct_value = custom.forward(y) print("2D idct_value") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test cpu using fft2 custom = dct2_fft2.IDCT2(expkM, expkN) dct_value = custom.forward(y) print("2D idct_value cuda") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) if torch.cuda.device_count(): # test gpu custom = dct.IDCT2(algorithm='N') dct_value = custom.forward(y.cuda()).cpu() print("2D idct_value cuda") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test gpu custom = dct.IDCT2(algorithm='2N') dct_value = custom.forward(y.cuda()).cpu() print("2D idct_value cuda") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test gpu custom = dct_lee.IDCT2() dct_value = custom.forward(y.cuda()).cpu() print("2D idct_value cuda") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5) # test gpu using ifft2 custom = dct2_fft2.IDCT2(expkM.cuda(), expkN.cuda()) dct_value = custom.forward(y.cuda()).cpu() print("2D idct_value cuda") print(dct_value.data.numpy()) np.testing.assert_allclose(dct_value.data.numpy(), golden_value, rtol=1e-6, atol=1e-5)