From fc1b17b00c80f445e97dcb4e393ed73ad47b8e82 Mon Sep 17 00:00:00 2001
From: "Timothy A. Smith" <tasmith4@illinois.edu>
Date: Mon, 27 May 2019 23:36:28 -0500
Subject: [PATCH] move bulk of compute_flux_derivative_gpu code into test
fixture
---
kernel_fixtures.py | 25 +++++++++++++++++++++++++
setup_fixtures.py | 19 +++++++++++++++++++
test.py | 34 ++++------------------------------
3 files changed, 48 insertions(+), 30 deletions(-)
diff --git a/kernel_fixtures.py b/kernel_fixtures.py
index e905363..0640e8c 100644
--- a/kernel_fixtures.py
+++ b/kernel_fixtures.py
@@ -45,3 +45,28 @@ def compute_flux_derivatives(ctx_factory, params, arrays):
metric_jacobians=arrays.metric_jacobians,
flux_derivatives=flux_derivatives_dev)
return flux_derivatives_dev.get()
+
+def compute_flux_derivatives_gpu(ctx_factory, params, arrays):
+ prg = fixtures.get_gpu_transformed_weno()
+
+ queue = fixtures.get_queue(ctx_factory)
+
+ nx_halo = params.nx + params.nhalo
+ ny_halo = params.ny + params.nhalo
+ nz_halo = params.nz + params.nhalo
+
+ flux_derivatives_dev = cl.array.empty(queue, (params.nvars, params.ndim,
+ nx_halo, ny_halo, nz_halo), dtype=np.float32, order="F")
+
+ prg = prg.copy(target=lp.PyOpenCLTarget(queue.device))
+
+ if 1:
+ with open("gen-code.cl", "w") as outf:
+ outf.write(lp.generate_code_v2(prg).device_code())
+
+ prg = lp.set_options(prg, no_numpy=True)
+
+ prg(queue, nvars=params.nvars, ndim=params.ndim,
+ states=arrays.states, fluxes=arrays.fluxes, metrics=arrays.metrics,
+ metric_jacobians=arrays.metric_jacobians,
+ flux_derivatives=flux_derivatives_dev)
diff --git a/setup_fixtures.py b/setup_fixtures.py
index 9e24722..b576866 100644
--- a/setup_fixtures.py
+++ b/setup_fixtures.py
@@ -1,5 +1,7 @@
import numpy as np
+import fixtures
+
class FluxDerivativeParams:
def __init__(self, nvars, ndim, nx, ny, nz):
@@ -36,3 +38,20 @@ def random_flux_derivative_arrays(params):
metric_jacobians = random_array(nx_halo, ny_halo, nz_halo)
return FluxDerivativeArrays(states, fluxes, metrics, metric_jacobians)
+
+
+def random_flux_derivative_arrays_on_device(ctx_factory, params):
+ queue = fixtures.get_queue(ctx_factory)
+
+ nvars = params.nvars
+ ndim = params.ndim
+ nx_halo = params.nx + params.nhalo
+ ny_halo = params.ny + params.nhalo
+ nz_halo = params.nz + params.nhalo
+
+ states = fixtures.f_array(queue, nvars, nx_halo, ny_halo, nz_halo)
+ fluxes = fixtures.f_array(queue, nvars, ndim, nx_halo, ny_halo, nz_halo)
+ metrics = fixtures.f_array(queue, ndim, ndim, nx_halo, ny_halo, nz_halo)
+ metric_jacobians = fixtures.f_array(queue, nx_halo, ny_halo, nz_halo)
+
+ return FluxDerivativeArrays(states, fluxes, metrics, metric_jacobians)
diff --git a/test.py b/test.py
index 00b21ef..914e029 100644
--- a/test.py
+++ b/test.py
@@ -32,41 +32,15 @@ def test_matvec(ctx_factory):
def test_compute_flux_derivatives(ctx_factory):
params = setup.FluxDerivativeParams(ndim=3, nvars=5, nx=10, ny=10, nz=10)
arrays = setup.random_flux_derivative_arrays(params)
+
kernel.compute_flux_derivatives(ctx_factory, params, arrays)
-@pytest.mark.skip("slow")
def test_compute_flux_derivatives_gpu(ctx_factory):
- prg = fixtures.get_gpu_transformed_weno()
-
- queue = fixtures.get_queue(ctx_factory)
-
- ndim = 3
- nvars = 5
- nx = 10
- ny = 10
- nz = 10
-
- states = fixtures.f_array(queue, nvars, nx+6, ny+6, nz+6)
- fluxes = fixtures.f_array(queue, nvars, ndim, nx+6, ny+6, nz+6)
- metrics = fixtures.f_array(queue, ndim, ndim, nx+6, ny+6, nz+6)
- metric_jacobians = fixtures.f_array(queue, nx+6, ny+6, nz+6)
-
- flux_derivatives_dev = cl.array.empty(queue, (nvars, ndim, nx+6, ny+6,
- nz+6), dtype=np.float32, order="F")
-
- prg = prg.copy(target=lp.PyOpenCLTarget(queue.device))
-
- if 1:
- with open("gen-code.cl", "w") as outf:
- outf.write(lp.generate_code_v2(prg).device_code())
-
- prg = lp.set_options(prg, no_numpy=True)
+ params = setup.FluxDerivativeParams(ndim=3, nvars=5, nx=10, ny=10, nz=10)
+ arrays = setup.random_flux_derivative_arrays_on_device(ctx_factory, params)
- prg(queue, nvars=nvars, ndim=ndim,
- states=states, fluxes=fluxes, metrics=metrics,
- metric_jacobians=metric_jacobians,
- flux_derivatives=flux_derivatives_dev)
+ kernel.compute_flux_derivatives_gpu(ctx_factory, params, arrays)
# This lets you run 'python test.py test_case(cl._csc)' without pytest.
--
GitLab