From 4664a75feb55bd734fe942b300ad294d5b6bdad4 Mon Sep 17 00:00:00 2001
From: Jean-Baptiste Keck <Jean-Baptiste.Keck@imag.fr>
Date: Mon, 12 Dec 2016 12:20:53 +0100
Subject: [PATCH] stretching requirements
---
hysop/codegen/base/kernel_codegen.py | 34 +++++++--
.../codegen/kernels/directional_stretching.py | 70 +++++++++++++++--
.../tests/test_directional_stretching.py | 75 +++++++++++++++----
3 files changed, 147 insertions(+), 32 deletions(-)
diff --git a/hysop/codegen/base/kernel_codegen.py b/hysop/codegen/base/kernel_codegen.py
index f9d2791ae..64f40d794 100644
--- a/hysop/codegen/base/kernel_codegen.py
+++ b/hysop/codegen/base/kernel_codegen.py
@@ -61,8 +61,20 @@ class KernelCodeGenerator(KernelBase, OpenClCodeGenerator):
self.gen_kernel_attributes()
- def required_workgroup_cache_size(self):
- return (0,0) # static & dynamic cache
+ #return global_work_size from effective work_size and given local_work_size
+ # /!\ it should be garanted that global_work_size is a multiple of local_work_size
+ def get_global_work_size(self, work_size, local_work_size):
+ work_size = np.asarray(work_size)
+ local_work_size = np.asarray(local_work_size)
+ return ((work_size+local_work_size-1)/local_work_size) * local_work_size
+
+ def min_ghosts(self):
+ ghosts = (0,)*self.work_dim
+ return np.asarray(ghosts)
+
+ #return a tuple of required (static,dynamic) cache bytes per workgroup
+ def required_workgroup_cache_size(self, local_work_size):
+ return (0,0)
def gen_kernel_variables(self):
tg = self.typegen
@@ -73,12 +85,18 @@ class KernelCodeGenerator(KernelBase, OpenClCodeGenerator):
kvars['work_dim'] = CodegenVariable('work_dim','uint', tg, symbolic_mode=sm)
kvars['global_index'] = CodegenVariable('GID', 'int', tg)
kvars['local_index'] = CodegenVariable('LID', 'int', tg)
- kvars['global_size'] = CodegenVectorClBuiltinFunc('global_size', 'G', 'int',work_dim,tg,symbolic_mode=sm)
- kvars['local_size'] = CodegenVectorClBuiltinFunc('local_size', 'L', 'int',work_dim,tg,symbolic_mode=sm)
- kvars['global_id'] = CodegenVectorClBuiltinFunc('global_id', 'gid', 'int',work_dim,tg)
- kvars['local_id'] = CodegenVectorClBuiltinFunc('local_id', 'lid', 'int',work_dim,tg)
- kvars['num_groups'] = CodegenVectorClBuiltinFunc('num_groups', 'ngroups', 'int',work_dim,tg,symbolic_mode=sm)
- kvars['group_id'] = CodegenVectorClBuiltinFunc('group_id', 'group_id','int',work_dim,tg)
+ kvars['global_size'] = CodegenVectorClBuiltinFunc('global_size', 'G',
+ 'int',work_dim,tg,symbolic_mode=sm)
+ kvars['local_size'] = CodegenVectorClBuiltinFunc('local_size', 'L',
+ 'int',work_dim,tg,symbolic_mode=sm)
+ kvars['global_id'] = CodegenVectorClBuiltinFunc('global_id', 'gid',
+ 'int',work_dim,tg)
+ kvars['local_id'] = CodegenVectorClBuiltinFunc('local_id', 'lid',
+ 'int',work_dim,tg)
+ kvars['num_groups'] = CodegenVectorClBuiltinFunc('num_groups', 'ngroups',
+ 'int',work_dim,tg,symbolic_mode=sm)
+ kvars['group_id'] = CodegenVectorClBuiltinFunc('group_id', 'group_id',
+ 'int',work_dim,tg)
self.update_vars(kvars)
diff --git a/hysop/codegen/kernels/directional_stretching.py b/hysop/codegen/kernels/directional_stretching.py
index 56e0aa695..b3ddac11f 100644
--- a/hysop/codegen/kernels/directional_stretching.py
+++ b/hysop/codegen/kernels/directional_stretching.py
@@ -113,12 +113,59 @@ class DirectionalStretchingKernel(KernelCodeGenerator):
self.gencode()
def min_ghosts(self):
- stencil_ghost = self.order/2
- if self.is_conservative:
- ghosts = self.rk_scheme.stages * stencil_ghost
- else:
- ghosts = stencil_ghost
- return ghosts
+ direction = self.direction
+ ghosts = [0]*self.dim
+ if self.boundary == BoundaryCondition.PERIODIC:
+ pass
+ elif self.boundary == BoundaryCondition.NONE:
+ stencil_ghost = self.order/2
+ if self.is_conservative:
+ ghosts[direction] = self.rk_scheme.stages * stencil_ghost
+ else:
+ ghosts[direction] = stencil_ghost
+ return np.asarray(ghosts)
+
+ #return global_work_size from effective work_size and given local_work_size
+ # /!\ it should be garanted that global_work_size is a multiple of local_work_size
+ def get_global_work_size(self, work_size, local_work_size):
+ work_size = np.asarray(work_size)
+ local_work_size = np.asarray(local_work_size)
+
+ cache_ghosts = self.cache_ghosts()
+ local_work = local_work_size - 2*cache_ghosts
+
+ return ((work_size+local_work-1)/local_work) * local_work_size
+
+ #return a tuple of required (static,dynamic) cache bytes per workgroup
+ def required_workgroup_cache_size(self, local_work_size):
+ dim = self.work_dim
+ ftype = self.ftype
+ cached = self.cached
+ direction = self.direction
+ cache_ghosts = self.cache_ghosts()
+ is_conservative = self.is_conservative
+ flt_bytes = self.typegen.FLT_BYTES[ftype]
+
+ sc,dc = 0,0
+ if cached:
+ count = dim*local_work_size[0]
+ if is_conservative:
+ count += local_work_size[0]
+
+ if 'local_size' in self.known_vars:
+ assert (self.known_vars['local_size'] == local_work_size)
+ sc += count
+ else:
+ dc += count
+
+ sc += 2*dim*(2*cache_ghosts)
+ if self.boundary == BoundaryCondition.PERIODIC:
+ sc += 2*dim*(1*cache_ghosts)
+
+ sc *= flt_bytes
+ dc *= flt_bytes
+
+ return (sc,dc)
def build_requirements(self,typegen,work_dim,ftype,order,cached,rk_scheme,direction,
boundary,force_symbolic,formulation,storage):
@@ -189,6 +236,13 @@ class DirectionalStretchingKernel(KernelCodeGenerator):
self.xyz = xyz
return kargs
+
+ def cache_ghosts(self):
+ stencil_ghost = self.order/2
+ if self.is_conservative:
+ return self.rk_scheme.stages * stencil_ghost
+ else:
+ return stencil_ghost
def gencode(self):
s = self
@@ -242,7 +296,7 @@ class DirectionalStretchingKernel(KernelCodeGenerator):
U = CodegenVectorClBuiltin('U',ftype,dim,tg)
cache_ghosts = CodegenVariable('cache_ghosts','int',tg,
- const=True,value=self.min_ghosts())
+ const=True,value=self.cache_ghosts())
local_work = CodegenVariable('lwork','int',tg,const=True)
cached_vars = ArgDict()
@@ -460,7 +514,7 @@ if __name__ == '__main__':
order=4, dim=dim, direction=0,
formulation=StretchingFormulation.GRAD_UW,
rk_scheme=ExplicitRungeKutta('RK2'),
- cached=False,
+ cached=True,
symbolic_mode=True,
boundary=BoundaryCondition.NONE,
known_vars=dict(
diff --git a/hysop/codegen/kernels/tests/test_directional_stretching.py b/hysop/codegen/kernels/tests/test_directional_stretching.py
index fe3bc6a10..9eaf7062f 100644
--- a/hysop/codegen/kernels/tests/test_directional_stretching.py
+++ b/hysop/codegen/kernels/tests/test_directional_stretching.py
@@ -45,20 +45,32 @@ class TestDirectionalStretching(object):
device_buffers = {
'no_ghosts': {
- 'ux': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['no_ghosts']['ux']),
- 'uy': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['no_ghosts']['uy']),
- 'uz': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['no_ghosts']['uz']),
- 'wx': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['no_ghosts']['wx']),
- 'wy': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['no_ghosts']['wy']),
- 'wz': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['no_ghosts']['wz'])
+ 'ux': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['no_ghosts']['ux']),
+ 'uy': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['no_ghosts']['uy']),
+ 'uz': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['no_ghosts']['uz']),
+ 'wx': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['no_ghosts']['wx']),
+ 'wy': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['no_ghosts']['wy']),
+ 'wz': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['no_ghosts']['wz'])
},
'with_ghosts': {
- 'ux': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['with_ghosts']['ux']),
- 'uy': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['with_ghosts']['uy']),
- 'uz': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['with_ghosts']['uz']),
- 'wx': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['with_ghosts']['wx']),
- 'wy': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['with_ghosts']['wy']),
- 'wz': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR, hostbuf=host_buffers_init['with_ghosts']['wz'])
+ 'ux': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['with_ghosts']['ux']),
+ 'uy': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['with_ghosts']['uy']),
+ 'uz': cl.Buffer(ctx, flags=mf.READ_ONLY | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['with_ghosts']['uz']),
+ 'wx': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['with_ghosts']['wx']),
+ 'wy': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['with_ghosts']['wy']),
+ 'wz': cl.Buffer(ctx, flags=mf.READ_WRITE | mf.COPY_HOST_PTR,
+ hostbuf=host_buffers_init['with_ghosts']['wz'])
}
}
@@ -81,6 +93,7 @@ class TestDirectionalStretching(object):
cls.device_buffers = device_buffers
cls.dt = 0.5
+ cls.local_work_size = np.asarray([16,1,1])
@classmethod
def teardown_class(cls):
@@ -112,14 +125,30 @@ class TestDirectionalStretching(object):
pass
def _cmp_buffers(self):
- self.to_cpu()
+ pass
def _do_compute_cpu(self,order,direction,boundary):
pass
def _do_compute_gpu(self, formulation, rk_scheme, order, direction, boundary, cached):
- known_vars = {}
+ dt = self.dt
+ local_work_size = self.local_work_size
+
+ known_vars = {
+ 'local_size': local_work_size,
+ 'dt': dt
+ }
+
+ kernel_args = []
+ if boundary == BoundaryCondition.PERIODIC:
+ work_size = self.grid_size
+ gpu_buffers = self.gpu_buffers['no_ghosts']
+ elif boundary == BoundaryCondition.NONE:
+ work_size = self.compute_grid_size
+ gpu_buffers = self.gpu_buffers['with_ghosts']
+ else:
+ raise ValueError()
dsk = DirectionalStretchingKernel(
typegen=self.typegen,
@@ -134,12 +163,26 @@ class TestDirectionalStretching(object):
boundary=boundary,
known_vars=known_vars)
+ global_work_size = dsk.get_global_work_size(work_size,local_work_size)
+ (static_shared_bytes, dynamic_shared_bytes) = \
+ dsk.required_workgroup_cache_size(local_work_size)
+
+ for varname in ['vx','vy','vz','wx','wy','wz']:
+ kernel_args.append(device_buffers[varname])
+ if (dynamic_shared_bytes != 0):
+ shared_buffer = cl.LocalMemory(dynamic_shared_bytes)
+ kernel_args.append(shared_buffer)
+
src = dsk.__str__()
- prg = cl.Program(self.typegen.ctx, src)
+ prg = cl.Program(self.typegen.context, src)
kernel = prg.all_kernels()[0]
-
+ kernel.set_args(*kernel_args)
self.to_gpu()
+ evt = cl.enqueue_nd_range_kernel(self.queue, kernel,
+ list(global_work_size), list(local_work_size))
+ evt.wait()
+ self.to_cpu()
def check_kernels(self, formulation, rk_scheme):
--
GitLab