[WIP] move 3rd-derivative of xc to GPU

gpu4pyscf/dft/tests/test_libxc.py

@@ -51,7 +51,7 @@ def _diff(dat, ref):
     return np.min((abs(d/(ref+1e-300)), abs(d)), axis=0)
 
 class KnownValues(unittest.TestCase):
-    def _check_xc(self, xc, spin=0, fxc_tol=1e-10, kxc_tol=1e-10):
+    def _check_xc(self, xc, spin=0, deriv=2, fxc_tol=1e-10, kxc_tol=1e-10):
         ni_cpu = numint_cpu()
         ni_gpu = numint_gpu()
         xctype = ni_cpu._xc_type(xc)
@@ -66,8 +66,15 @@ def _check_xc(self, xc, spin=0, fxc_tol=1e-10, kxc_tol=1e-10):
         if spin != 0:
             rho = (rho, rho)
 
-        exc_cpu, vxc_cpu, fxc_cpu, kxc_cpu = ni_cpu.eval_xc_eff(xc, rho, deriv=2, xctype=xctype)
-        exc_gpu, vxc_gpu, fxc_gpu, kxc_gpu = ni_gpu.eval_xc_eff(xc, cupy.array(rho), deriv=2, xctype=xctype)
+        exc_cpu, vxc_cpu, fxc_cpu, kxc_cpu = ni_cpu.eval_xc_eff(xc, rho, deriv=deriv, xctype=xctype)
+        exc_gpu, vxc_gpu, fxc_gpu, kxc_gpu = ni_gpu.eval_xc_eff(xc, cupy.array(rho), deriv=deriv, xctype=xctype)
+
+        print(f"{xc} {spin} exc", _diff(exc_gpu[:,0].get(), exc_cpu).max())
+        print(f"{xc} {spin} vxc", _diff(vxc_gpu.get(), vxc_cpu).max())
+        if fxc_gpu is not None:
+            print(f"{xc} {spin} fxc", _diff(fxc_gpu.get(), fxc_cpu).max())
+        if kxc_gpu is not None:
+            print(f"{xc} {spin} kxc", _diff(kxc_gpu.get(), kxc_cpu).max())
 
         assert _diff(exc_gpu[:,0].get(), exc_cpu).max() < 1e-10
         assert _diff(vxc_gpu.get(), vxc_cpu).max() < 1e-10
@@ -77,15 +84,15 @@ def _check_xc(self, xc, spin=0, fxc_tol=1e-10, kxc_tol=1e-10):
             assert _diff(kxc_gpu.get(), kxc_cpu).max() < kxc_tol
 
     def test_LDA(self):
-        self._check_xc('LDA_C_VWN')
+        self._check_xc('LDA_C_VWN', deriv=3)
 
     def test_GGA(self):
-        self._check_xc('HYB_GGA_XC_B3LYP')
-        self._check_xc('GGA_X_B88', fxc_tol=1e-10)
-        self._check_xc('GGA_C_PBE', fxc_tol=1e-4)
+        self._check_xc('HYB_GGA_XC_B3LYP', deriv=3, kxc_tol=1e-9)
+        self._check_xc('GGA_X_B88', fxc_tol=1e-10, deriv=3, kxc_tol=1e-8)
+        self._check_xc('GGA_C_PBE', fxc_tol=1e-4, deriv=3, kxc_tol=3e2)
 
     def test_mGGA(self):
-        self._check_xc('MGGA_C_M06', fxc_tol=1e-4)
+        self._check_xc('MGGA_C_M06', fxc_tol=1e-4, deriv=3, kxc_tol=1e-2)
 
     def test_u_LDA(self):
         self._check_xc('LDA_C_VWN', spin=1)
@@ -101,4 +108,4 @@ def test_u_mGGA(self):
 
 if __name__ == "__main__":
     print("Full Tests for xc fun")
-    unittest.main()
+    unittest.main()

requirements.txt

@@ -1,5 +1,5 @@
 cutensor-cu12==2.2.0
-gpu4pyscf-libxc-cuda12x==0.5.0
+gpu4pyscf-libxc-cuda12x==0.7.0
 cupy-cuda12x==13.4.1
 pyscf==2.8.0
 basis-set-exchange==0.11

setup.py

@@ -138,6 +138,6 @@ def initialize_with_default_plat_name(self):
         'pyscf-dispersion',
         f'cupy-cuda{CUDA_VERSION}>=13.0,!=13.4.0', # Due to expm in cupyx.scipy.linalg and cutensor 2.0
         'geometric',
-        f'gpu4pyscf-libxc-cuda{CUDA_VERSION}==0.5',
+        f'gpu4pyscf-libxc-cuda{CUDA_VERSION}==0.7.0',
     ]
 )