def wrap_temporary_decl(self, decl, scope):
if scope == AddressSpace.LOCAL:
from cgen.cuda import CudaShared
return CudaShared(decl)
elif scope == AddressSpace.PRIVATE:
return decl
else:
raise ValueError("unexpected temporary variable scope: %s" % scope)
def get_kernel(self, diff_op, elgroup, for_benchmark=False):
from cgen import \
Pointer, POD, Value, ArrayOf, Const, \
Module, FunctionDeclaration, FunctionBody, Block, \
Comment, Line, Define, Include, \
Initializer, If, For, Statement, Assign
from pycuda.tools import dtype_to_ctype
from cgen.cuda import CudaShared, CudaGlobal
discr = self.discr
d = discr.dimensions
dims = range(d)
plan = self.plan
given = plan.given
elgroup, = discr.element_groups
float_type = given.float_type
f_decl = CudaGlobal(FunctionDeclaration(Value("void", "apply_diff_mat_smem"),
[Pointer(POD(float_type, "debugbuf")), Pointer(POD(float_type, "field")), ]
+ [Pointer(POD(float_type, "drst%d_global" % i)) for i in dims]
))
par = plan.parallelism
cmod = Module([
Include("pycuda-helpers.hpp"),
])
if float_type == numpy.float64:
cmod.append(Value("texture<fp_tex_double, 1, cudaReadModeElementType>",
"diff_rst_mat_tex"))
elif float_type == numpy.float32:
rst_channels = given.devdata.make_valid_tex_channel_count(d)
cmod.append(Value("texture<float%d, 1, cudaReadModeElementType>"
% rst_channels, "diff_rst_mat_tex"))
else:
raise ValueError("unsupported float type: %s" % float_type)
# only preimage size variation is supported here
assert plan.image_dofs_per_el == given.dofs_per_el()
assert plan.aligned_image_dofs_per_microblock == given.microblock.aligned_floats
# FIXME: aligned_image_dofs_per_microblock must be divisible
# by this, therefore hardcoding for now.
chunk_size = 16
cmod.extend([
Line(),
Define("DIMENSIONS", discr.dimensions),
Define("IMAGE_DOFS_PER_EL", plan.image_dofs_per_el),
Define("PREIMAGE_DOFS_PER_EL", plan.preimage_dofs_per_el),
Define("ALIGNED_IMAGE_DOFS_PER_MB", plan.aligned_image_dofs_per_microblock),
Define("ALIGNED_PREIMAGE_DOFS_PER_MB", plan.aligned_preimage_dofs_per_microblock),
Define("ELS_PER_MB", given.microblock.elements),
Define("IMAGE_DOFS_PER_MB", "(IMAGE_DOFS_PER_EL*ELS_PER_MB)"),
Line(),
Define("CHUNK_SIZE", chunk_size),
Define("CHUNK_DOF", "threadIdx.x"),
Define("PAR_MB_NR", "threadIdx.y"),
Define("CHUNK_NR", "threadIdx.z"),
Define("IMAGE_MB_DOF", "(CHUNK_NR*CHUNK_SIZE+CHUNK_DOF)"),
Define("IMAGE_EL_DOF", "(IMAGE_MB_DOF - mb_el*IMAGE_DOFS_PER_EL)"),
Line(),
Define("MACROBLOCK_NR", "blockIdx.x"),
Line(),
Define("PAR_MB_COUNT", par.parallel),
Define("INLINE_MB_COUNT", par.inline),
Define("SEQ_MB_COUNT", par.serial),
Line(),
Define("GLOBAL_MB_NR_BASE",
"(MACROBLOCK_NR*PAR_MB_COUNT*INLINE_MB_COUNT*SEQ_MB_COUNT)"),
Define("GLOBAL_MB_NR",
"(GLOBAL_MB_NR_BASE"
"+ (seq_mb_number*PAR_MB_COUNT + PAR_MB_NR)*INLINE_MB_COUNT)"),
Define("GLOBAL_MB_IMAGE_DOF_BASE", "(GLOBAL_MB_NR*ALIGNED_IMAGE_DOFS_PER_MB)"),
Define("GLOBAL_MB_PREIMAGE_DOF_BASE", "(GLOBAL_MB_NR*ALIGNED_PREIMAGE_DOFS_PER_MB)"),
Line(),
CudaShared(
ArrayOf(
ArrayOf(
ArrayOf(
POD(float_type, "smem_field"),
"PAR_MB_COUNT"),
"INLINE_MB_COUNT"),
"ALIGNED_PREIMAGE_DOFS_PER_MB")),
Line(),
])
S = Statement
f_body = Block([
Initializer(Const(POD(numpy.uint16, "mb_el")),
"IMAGE_MB_DOF / IMAGE_DOFS_PER_EL"),
Line(),
])
# ---------------------------------------------------------------------
def get_load_code():
mb_img_dofs = plan.aligned_image_dofs_per_microblock
mb_preimg_dofs = plan.aligned_preimage_dofs_per_microblock
preimg_dofs_over_dofs = (mb_preimg_dofs+mb_img_dofs-1) // mb_img_dofs
load_code = []
store_code = []
var_num = 0
for load_block in range(preimg_dofs_over_dofs):
for inl in range(par.inline):
# load and store are split for better pipelining
# compiler can't figure that out because of branch
var = "tmp%d" % var_num
var_num += 1
load_code.append(POD(float_type, var))
block_addr = "%d * ALIGNED_IMAGE_DOFS_PER_MB + IMAGE_MB_DOF" % load_block
load_instr = Assign(var,
"field[GLOBAL_MB_PREIMAGE_DOF_BASE"
" + %d*ALIGNED_PREIMAGE_DOFS_PER_MB"
" + %s]" % (inl, block_addr))
store_instr = Assign(
"smem_field[PAR_MB_NR][%d][%s]" % (inl, block_addr),
var
)
if (load_block+1)*mb_img_dofs >= mb_preimg_dofs:
cond = "%s < ALIGNED_PREIMAGE_DOFS_PER_MB" % block_addr
load_instr = If(cond, load_instr)
store_instr = If(cond, store_instr)
load_code.append(load_instr)
store_code.append(store_instr)
return Block(load_code + [Line()] + store_code)
def get_scalar_diff_code():
code = []
for inl in range(par.inline):
for axis in dims:
code.append(
Initializer(POD(float_type, "d%drst%d" % (inl, axis)), 0))
code.append(Line())
tex_channels = ["x", "y", "z", "w"]
store_code = Block()
for inl in range(par.inline):
for rst_axis in dims:
store_code.append(Assign(
"drst%d_global[GLOBAL_MB_IMAGE_DOF_BASE + "
"%d*ALIGNED_IMAGE_DOFS_PER_MB + IMAGE_MB_DOF]"
% (rst_axis, inl),
"d%drst%d" % (inl, rst_axis)
))
from hedge.backends.cuda.tools import unroll
code.extend([
Comment("everybody needs to be done with the old data"),
S("__syncthreads()"),
Line(),
get_load_code(),
Line(),
Comment("all the new data must be loaded"),
S("__syncthreads()"),
Line(),
])
if float_type == numpy.float32:
code.append(Value("float%d" % rst_channels, "dmat_entries"))
code.extend([
POD(float_type, "field_value%d" % inl)
for inl in range(par.inline)
]+[Line()])
def unroll_body(j):
result = [
Assign("field_value%d" % inl,
"smem_field[PAR_MB_NR][%d][mb_el*PREIMAGE_DOFS_PER_EL+%s]" % (inl, j))
for inl in range(par.inline)
]
if float_type == numpy.float32:
result.append(Assign("dmat_entries",
"tex1Dfetch(diff_rst_mat_tex, IMAGE_EL_DOF + %s*IMAGE_DOFS_PER_EL)" % j))
result.extend(
S("d%drst%d += dmat_entries.%s * field_value%d"
% (inl, axis, tex_channels[axis], inl))
for inl in range(par.inline)
for axis in dims)
elif float_type == numpy.float64:
result.extend(
S("d%(inl)drst%(axis)d += "
"fp_tex1Dfetch(diff_rst_mat_tex, %(axis)d "
"+ DIMENSIONS*(IMAGE_EL_DOF + %(j)d*IMAGE_DOFS_PER_EL))"
"* field_value%(inl)d" % {
"inl": inl,
"axis": axis,
"j": j
})
for inl in range(par.inline)
for axis in dims)
else:
assert False
return result
code.append(If("IMAGE_MB_DOF < IMAGE_DOFS_PER_MB", Block(unroll(unroll_body,
total_number=plan.preimage_dofs_per_el)
+[store_code])))
return code
f_body.extend([
For("unsigned short seq_mb_number = 0",
"seq_mb_number < SEQ_MB_COUNT",
"++seq_mb_number",
Block(get_scalar_diff_code())
)
])
# finish off ----------------------------------------------------------
cmod.append(FunctionBody(f_decl, f_body))
if not for_benchmark and "cuda_dump_kernels" in discr.debug:
from hedge.tools import open_unique_debug_file
open_unique_debug_file("diff", ".cu").write(str(cmod))
mod = SourceModule(cmod,
keep="cuda_keep_kernels" in discr.debug,
#options=["--maxrregcount=16"]
)
func = mod.get_function("apply_diff_mat_smem")
if "cuda_diff" in discr.debug:
print "diff: lmem=%d smem=%d regs=%d" % (
func.local_size_bytes,
func.shared_size_bytes,
func.registers)
diff_rst_mat_texref = mod.get_texref("diff_rst_mat_tex")
gpu_diffmats = self.gpu_diffmats(diff_op, elgroup)
if given.float_type == numpy.float32:
gpu_diffmats.bind_to_texref_ext(diff_rst_mat_texref, rst_channels)
elif given.float_type == numpy.float64:
gpu_diffmats.bind_to_texref_ext(diff_rst_mat_texref,
allow_double_hack=True)
else:
assert False
assert given.microblock.aligned_floats % chunk_size == 0
block = (
chunk_size,
plan.parallelism.parallel,
given.microblock.aligned_floats//chunk_size)
func.prepare(
["PP"] + discr.dimensions*["P"],
texrefs=[diff_rst_mat_texref])
return block, func
def get_kernel(self, diff_op_cls, elgroup, for_benchmark=False):
from cgen import \
Pointer, POD, Value, ArrayOf, \
Module, FunctionDeclaration, FunctionBody, Block, \
Line, Define, Include, \
Initializer, If, For, Statement, Assign
from cgen import dtype_to_ctype
from cgen.cuda import CudaShared, CudaGlobal
discr = self.discr
d = discr.dimensions
dims = range(d)
given = self.plan.given
par = self.plan.parallelism
diffmat_data = self.gpu_diffmats(diff_op_cls, elgroup)
elgroup, = discr.element_groups
float_type = given.float_type
f_decl = CudaGlobal(
FunctionDeclaration(
Value("void", "apply_diff_mat"),
[
Pointer(POD(numpy.uint8, "gmem_diff_rst_mat")),
#Pointer(POD(float_type, "debugbuf")),
] +
[Pointer(POD(float_type, "drst%d_global" % i)) for i in dims]))
rst_channels = given.devdata.make_valid_tex_channel_count(d)
cmod = Module([
Include("pycuda-helpers.hpp"),
Line(),
Value(
"texture<fp_tex_%s, 1, cudaReadModeElementType>" %
dtype_to_ctype(float_type), "field_tex"),
Line(),
Define("DIMENSIONS", discr.dimensions),
Define("DOFS_PER_EL", given.dofs_per_el()),
Line(),
Define("SEGMENT_DOF", "threadIdx.x"),
Define("PAR_MB_NR", "threadIdx.y"),
Line(),
Define("MB_SEGMENT", "blockIdx.x"),
Define("MACROBLOCK_NR", "blockIdx.y"),
Line(),
Define("DOFS_PER_SEGMENT", self.plan.segment_size),
Define("SEGMENTS_PER_MB", self.plan.segments_per_microblock()),
Define("ALIGNED_DOFS_PER_MB", given.microblock.aligned_floats),
Define("ELS_PER_MB", given.microblock.elements),
Line(),
Define("PAR_MB_COUNT", par.parallel),
Define("INLINE_MB_COUNT", par.inline),
Define("SEQ_MB_COUNT", par.serial),
Line(),
Define("THREAD_NUM", "(SEGMENT_DOF+PAR_MB_NR*DOFS_PER_SEGMENT)"),
Define("COALESCING_THREAD_COUNT",
"(PAR_MB_COUNT*DOFS_PER_SEGMENT)"),
Line(),
Define("MB_DOF_BASE", "(MB_SEGMENT*DOFS_PER_SEGMENT)"),
Define("MB_DOF", "(MB_DOF_BASE+SEGMENT_DOF)"),
Define(
"GLOBAL_MB_NR_BASE",
"(MACROBLOCK_NR*PAR_MB_COUNT*INLINE_MB_COUNT*SEQ_MB_COUNT)"),
Define(
"GLOBAL_MB_NR", "(GLOBAL_MB_NR_BASE"
"+ (seq_mb_number*PAR_MB_COUNT + PAR_MB_NR)*INLINE_MB_COUNT)"),
Define("GLOBAL_MB_DOF_BASE", "(GLOBAL_MB_NR*ALIGNED_DOFS_PER_MB)"),
Line(),
Define("DIFFMAT_SEGMENT_FLOATS", diffmat_data.block_floats),
Define("DIFFMAT_SEGMENT_BYTES",
"(DIFFMAT_SEGMENT_FLOATS*%d)" % given.float_size()),
Define("DIFFMAT_COLUMNS", diffmat_data.matrix_columns),
Line(),
CudaShared(
ArrayOf(POD(float_type, "smem_diff_rst_mat"),
"DIFFMAT_COLUMNS*DOFS_PER_SEGMENT")),
Line(),
])
S = Statement
f_body = Block()
f_body.extend_log_block("calculate responsibility data", [
Initializer(POD(numpy.uint16, "mb_el"), "MB_DOF/DOFS_PER_EL"),
])
from hedge.backends.cuda.tools import get_load_code
f_body.extend(
get_load_code(
dest="smem_diff_rst_mat",
base="gmem_diff_rst_mat + MB_SEGMENT*DIFFMAT_SEGMENT_BYTES",
bytes="DIFFMAT_SEGMENT_BYTES",
descr="load diff mat segment") +
[S("__syncthreads()"), Line()])
# ---------------------------------------------------------------------
def get_scalar_diff_code():
code = []
for inl in range(par.inline):
for axis in dims:
code.append(
Initializer(POD(float_type, "d%drst%d" % (inl, axis)),
0))
code.append(Line())
def get_mat_entry(row, col, axis):
return ("smem_diff_rst_mat["
"%(row)s*DIFFMAT_COLUMNS + %(axis)s*DOFS_PER_EL"
" + %(col)s"
"]" % {
"row": row,
"col": col,
"axis": axis
})
tex_channels = ["x", "y", "z", "w"]
from hedge.backends.cuda.tools import unroll
code.extend([
POD(float_type, "field_value%d" % inl)
for inl in range(par.inline)
] + [Line()] + unroll(
lambda j: [
Assign(
"field_value%d" % inl,
"fp_tex1Dfetch(field_tex, GLOBAL_MB_DOF_BASE + %d*ALIGNED_DOFS_PER_MB "
"+ mb_el*DOFS_PER_EL + %s)" % (inl, j))
for inl in range(par.inline)
] + [Line()] + [
S("d%drst%d += %s * field_value%d" %
(inl, axis, get_mat_entry("SEGMENT_DOF", j, axis), inl))
for axis in dims for inl in range(par.inline)
] + [Line()], given.dofs_per_el(), self.plan.max_unroll))
store_code = Block()
for inl in range(par.inline):
for rst_axis in dims:
store_code.append(
Assign(
"drst%d_global[GLOBAL_MB_DOF_BASE"
" + %d*ALIGNED_DOFS_PER_MB + MB_DOF]" %
(rst_axis, inl),
"d%drst%d" % (inl, rst_axis),
))
code.append(If("MB_DOF < DOFS_PER_EL*ELS_PER_MB", store_code))
return code
f_body.extend([
For("unsigned short seq_mb_number = 0",
"seq_mb_number < SEQ_MB_COUNT", "++seq_mb_number",
Block(get_scalar_diff_code()))
])
# finish off ----------------------------------------------------------
cmod.append(FunctionBody(f_decl, f_body))
if not for_benchmark and "cuda_dump_kernels" in discr.debug:
from hedge.tools import open_unique_debug_file
open_unique_debug_file("diff", ".cu").write(str(cmod))
mod = SourceModule(
cmod,
keep="cuda_keep_kernels" in discr.debug,
#options=["--maxrregcount=10"]
)
field_texref = mod.get_texref("field_tex")
func = mod.get_function("apply_diff_mat")
func.prepare(discr.dimensions * [float_type] + ["P"],
block=(self.plan.segment_size, par.parallel, 1),
texrefs=[field_texref])
if "cuda_diff" in discr.debug:
print "diff: lmem=%d smem=%d regs=%d" % (
func.local_size_bytes, func.shared_size_bytes, func.num_regs)
return func, field_texref
def get_kernel(self, fdata, ilist_data, for_benchmark):
from cgen.cuda import CudaShared, CudaGlobal
from pycuda.tools import dtype_to_ctype
discr = self.discr
given = self.plan.given
fplan = self.plan
d = discr.dimensions
dims = range(d)
elgroup, = discr.element_groups
float_type = given.float_type
f_decl = CudaGlobal(
FunctionDeclaration(Value("void", "apply_flux"), [
Pointer(POD(float_type, "debugbuf")),
Pointer(POD(numpy.uint8, "gmem_facedata")),
] + [
Pointer(POD(float_type, "gmem_fluxes_on_faces%d" % flux_nr))
for flux_nr in range(len(self.fluxes))
]))
cmod = Module()
cmod.append(Include("pycuda-helpers.hpp"))
for dep_expr in self.all_deps:
cmod.extend([
Value(
"texture<%s, 1, cudaReadModeElementType>" %
dtype_to_ctype(float_type, with_fp_tex_hack=True),
"field%d_tex" % self.dep_to_index[dep_expr])
])
if fplan.flux_count != len(self.fluxes):
from warnings import warn
warn(
"Flux count in flux execution plan different from actual flux count.\n"
"You may want to specify the tune_for= kwarg in the Discretization\n"
"constructor.")
cmod.extend([
Line(),
Typedef(POD(float_type, "value_type")),
Line(),
flux_header_struct(float_type, discr.dimensions),
Line(),
face_pair_struct(float_type, discr.dimensions),
Line(),
Define("DIMENSIONS", discr.dimensions),
Define("DOFS_PER_FACE", fplan.dofs_per_face),
Define("THREADS_PER_FACE", fplan.threads_per_face()),
Line(),
Define("CONCURRENT_FACES", fplan.parallel_faces),
Define("BLOCK_MB_COUNT", fplan.mbs_per_block),
Line(),
Define("FACEDOF_NR", "threadIdx.x"),
Define("BLOCK_FACE", "threadIdx.y"),
Line(),
Define("FLUX_COUNT", len(self.fluxes)),
Line(),
Define("THREAD_NUM", "(FACEDOF_NR + BLOCK_FACE*THREADS_PER_FACE)"),
Define("THREAD_COUNT", "(THREADS_PER_FACE*CONCURRENT_FACES)"),
Define(
"COALESCING_THREAD_COUNT",
"(THREAD_COUNT < 0x10 ? THREAD_COUNT : THREAD_COUNT & ~0xf)"),
Line(),
Define("DATA_BLOCK_SIZE", fdata.block_bytes),
Define("ALIGNED_FACE_DOFS_PER_MB",
fplan.aligned_face_dofs_per_microblock()),
Define("ALIGNED_FACE_DOFS_PER_BLOCK",
"(ALIGNED_FACE_DOFS_PER_MB*BLOCK_MB_COUNT)"),
Line(),
Define("FOF_BLOCK_BASE",
"(blockIdx.x*ALIGNED_FACE_DOFS_PER_BLOCK)"),
Line(),
] + ilist_data.code + [
Line(),
Value("texture<index_list_entry_t, 1, cudaReadModeElementType>",
"tex_index_lists"),
Line(),
fdata.struct,
Line(),
CudaShared(Value("flux_data", "data")),
])
if not fplan.direct_store:
cmod.extend([
CudaShared(
ArrayOf(
ArrayOf(POD(float_type, "smem_fluxes_on_faces"),
"FLUX_COUNT"),
"ALIGNED_FACE_DOFS_PER_MB*BLOCK_MB_COUNT")),
Line(),
])
S = Statement
f_body = Block()
from hedge.backends.cuda.tools import get_load_code
f_body.extend(
get_load_code(dest="&data",
base="gmem_facedata + blockIdx.x*DATA_BLOCK_SIZE",
bytes="sizeof(flux_data)",
descr="load face_pair data") +
[S("__syncthreads()"), Line()])
def get_flux_code(flux_writer):
flux_code = Block([])
flux_code.extend([
Initializer(Pointer(Value("face_pair", "fpair")),
"data.facepairs+fpair_nr"),
Initializer(
MaybeUnused(POD(numpy.uint32, "a_index")),
"fpair->a_base + tex1Dfetch(tex_index_lists, "
"fpair->a_ilist_index + FACEDOF_NR)"),
Initializer(
MaybeUnused(POD(numpy.uint32, "b_index")),
"fpair->b_base + tex1Dfetch(tex_index_lists, "
"fpair->b_ilist_index + FACEDOF_NR)"),
Line(),
flux_writer(),
Line(),
S("fpair_nr += CONCURRENT_FACES")
])
return flux_code
flux_computation = Block([
Comment("fluxes for dual-sided (intra-block) interior face pairs"),
While("fpair_nr < data.header.same_facepairs_end",
get_flux_code(lambda: self.write_interior_flux_code(True))),
Line(),
Comment("work around nvcc assertion failure"),
S("fpair_nr+=1"),
S("fpair_nr-=1"),
Line(),
Comment(
"fluxes for single-sided (inter-block) interior face pairs"),
While("fpair_nr < data.header.diff_facepairs_end",
get_flux_code(lambda: self.write_interior_flux_code(False))),
Line(),
Comment("fluxes for single-sided boundary face pairs"),
While(
"fpair_nr < data.header.bdry_facepairs_end",
get_flux_code(
lambda: self.write_boundary_flux_code(for_benchmark))),
])
f_body.extend_log_block("compute the fluxes", [
Initializer(POD(numpy.uint32, "fpair_nr"), "BLOCK_FACE"),
If("FACEDOF_NR < DOFS_PER_FACE", flux_computation)
])
if not fplan.direct_store:
f_body.extend([Line(), S("__syncthreads()"), Line()])
f_body.extend_log_block(
"store fluxes",
[
#Assign("debugbuf[blockIdx.x]", "FOF_BLOCK_BASE"),
#Assign("debugbuf[0]", "FOF_BLOCK_BASE"),
#Assign("debugbuf[0]", "sizeof(face_pair)"),
For(
"unsigned word_nr = THREAD_NUM",
"word_nr < ALIGNED_FACE_DOFS_PER_MB*BLOCK_MB_COUNT",
"word_nr += COALESCING_THREAD_COUNT",
Block([
Assign(
"gmem_fluxes_on_faces%d[FOF_BLOCK_BASE+word_nr]"
% flux_nr,
"smem_fluxes_on_faces[%d][word_nr]" % flux_nr)
for flux_nr in range(len(self.fluxes))
]
#+[If("isnan(smem_fluxes_on_faces[%d][word_nr])" % flux_nr,
#Block([
#Assign("debugbuf[blockIdx.x]", "word_nr"),
#])
#)
#for flux_nr in range(len(self.fluxes))]
))
])
if False:
f_body.extend([
Assign("debugbuf[blockIdx.x*96+32+BLOCK_FACE*32+threadIdx.x]",
"fpair_nr"),
Assign("debugbuf[blockIdx.x*96+16]",
"data.header.same_facepairs_end"),
Assign("debugbuf[blockIdx.x*96+17]",
"data.header.diff_facepairs_end"),
Assign("debugbuf[blockIdx.x*96+18]",
"data.header.bdry_facepairs_end"),
])
# finish off ----------------------------------------------------------
cmod.append(FunctionBody(f_decl, f_body))
if not for_benchmark and "cuda_dump_kernels" in discr.debug:
from hedge.tools import open_unique_debug_file
open_unique_debug_file("flux_gather", ".cu").write(str(cmod))
#from pycuda.tools import allow_user_edit
mod = SourceModule(
#allow_user_edit(cmod, "kernel.cu", "the flux kernel"),
cmod,
keep="cuda_keep_kernels" in discr.debug)
expr_to_texture_map = dict(
(dep_expr,
mod.get_texref("field%d_tex" % self.dep_to_index[dep_expr]))
for dep_expr in self.all_deps)
index_list_texref = mod.get_texref("tex_index_lists")
index_list_texref.set_address(ilist_data.device_memory,
ilist_data.bytes)
index_list_texref.set_format(
cuda.dtype_to_array_format(ilist_data.type), 1)
index_list_texref.set_flags(cuda.TRSF_READ_AS_INTEGER)
func = mod.get_function("apply_flux")
block = (fplan.threads_per_face(), fplan.parallel_faces, 1)
func.prepare(
(2 + len(self.fluxes)) * "P",
texrefs=expr_to_texture_map.values() + [index_list_texref])
if "cuda_flux" in discr.debug:
print "flux: lmem=%d smem=%d regs=%d" % (
func.local_size_bytes, func.shared_size_bytes, func.num_regs)
return block, func, expr_to_texture_map
def get_kernel(self, with_scaling, for_benchmark=False):
from cgen import \
Pointer, POD, Value, ArrayOf, \
Module, FunctionDeclaration, FunctionBody, Block, \
Line, Define, Include, \
Initializer, If, For, Statement, Assign, \
ArrayInitializer
from cgen import dtype_to_ctype
from cgen.cuda import CudaShared, CudaConstant, CudaGlobal
discr = self.discr
d = discr.dimensions
dims = range(d)
given = self.plan.given
float_type = given.float_type
f_decl = CudaGlobal(
FunctionDeclaration(Value("void", "apply_el_local_mat_smem_mat"), [
Pointer(POD(float_type, "out_vector")),
Pointer(POD(numpy.uint8, "gmem_matrix")),
Pointer(POD(float_type, "debugbuf")),
POD(numpy.uint32, "microblock_count"),
]))
cmod = Module([
Include("pycuda-helpers.hpp"),
Line(),
Value(
"texture<fp_tex_%s, 1, cudaReadModeElementType>" %
dtype_to_ctype(float_type), "in_vector_tex"),
])
if with_scaling:
cmod.append(
Value(
"texture<fp_tex_%s, 1, cudaReadModeElementType>" %
dtype_to_ctype(float_type), "scaling_tex"), )
par = self.plan.parallelism
cmod.extend([
Line(),
Define("DIMENSIONS", discr.dimensions),
Define("DOFS_PER_EL", given.dofs_per_el()),
Define("PREIMAGE_DOFS_PER_EL", self.plan.preimage_dofs_per_el),
Line(),
Define("SEGMENT_DOF", "threadIdx.x"),
Define("PAR_MB_NR", "threadIdx.y"),
Line(),
Define("MB_SEGMENT", "blockIdx.x"),
Define("MACROBLOCK_NR", "blockIdx.y"),
Line(),
Define("DOFS_PER_SEGMENT", self.plan.segment_size),
Define("SEGMENTS_PER_MB", self.plan.segments_per_microblock()),
Define("ALIGNED_DOFS_PER_MB", given.microblock.aligned_floats),
Define("ALIGNED_PREIMAGE_DOFS_PER_MB",
self.plan.aligned_preimage_dofs_per_microblock),
Define("MB_EL_COUNT", given.microblock.elements),
Line(),
Define("PAR_MB_COUNT", par.parallel),
Define("INLINE_MB_COUNT", par.inline),
Define("SEQ_MB_COUNT", par.serial),
Line(),
Define("THREAD_NUM", "(SEGMENT_DOF+PAR_MB_NR*DOFS_PER_SEGMENT)"),
Define("COALESCING_THREAD_COUNT",
"(PAR_MB_COUNT*DOFS_PER_SEGMENT)"),
Line(),
Define("MB_DOF_BASE", "(MB_SEGMENT*DOFS_PER_SEGMENT)"),
Define("MB_DOF", "(MB_DOF_BASE+SEGMENT_DOF)"),
Define(
"GLOBAL_MB_NR_BASE",
"(MACROBLOCK_NR*PAR_MB_COUNT*INLINE_MB_COUNT*SEQ_MB_COUNT)"),
Define(
"GLOBAL_MB_NR", "(GLOBAL_MB_NR_BASE"
"+ (seq_mb_number*PAR_MB_COUNT + PAR_MB_NR)*INLINE_MB_COUNT)"),
Define("GLOBAL_MB_DOF_BASE", "(GLOBAL_MB_NR*ALIGNED_DOFS_PER_MB)"),
Define("GLOBAL_MB_PREIMG_DOF_BASE",
"(GLOBAL_MB_NR*ALIGNED_PREIMAGE_DOFS_PER_MB)"),
Line(),
Define("MATRIX_COLUMNS", self.plan.gpu_matrix_columns()),
Define("MATRIX_SEGMENT_FLOATS",
self.plan.gpu_matrix_block_floats()),
Define("MATRIX_SEGMENT_BYTES",
"(MATRIX_SEGMENT_FLOATS*%d)" % given.float_size()),
Line(),
CudaShared(
ArrayOf(POD(float_type, "smem_matrix"),
"MATRIX_SEGMENT_FLOATS")),
CudaShared(
ArrayOf(
ArrayOf(
ArrayOf(POD(float_type, "dof_buffer"), "PAR_MB_COUNT"),
"INLINE_MB_COUNT"), "DOFS_PER_SEGMENT"), ),
CudaShared(POD(numpy.uint16, "segment_start_el")),
CudaShared(POD(numpy.uint16, "segment_stop_el")),
CudaShared(POD(numpy.uint16, "segment_el_count")),
Line(),
ArrayInitializer(
CudaConstant(
ArrayOf(POD(numpy.uint32, "segment_start_el_lookup"),
"SEGMENTS_PER_MB")),
[(chk * self.plan.segment_size) // given.dofs_per_el()
for chk in range(self.plan.segments_per_microblock())]),
ArrayInitializer(
CudaConstant(
ArrayOf(POD(numpy.uint32, "segment_stop_el_lookup"),
"SEGMENTS_PER_MB")),
[
min(given.microblock.elements,
(chk * self.plan.segment_size +
self.plan.segment_size - 1) // given.dofs_per_el() +
1)
for chk in range(self.plan.segments_per_microblock())
]),
])
S = Statement
f_body = Block()
f_body.extend_log_block(
"calculate this dof's element",
[Initializer(POD(numpy.uint8, "mb_el"), "MB_DOF/DOFS_PER_EL")])
if self.plan.use_prefetch_branch:
f_body.extend_log_block("calculate segment responsibility data", [
If(
"THREAD_NUM==0",
Block([
Assign("segment_start_el",
"segment_start_el_lookup[MB_SEGMENT]"),
Assign("segment_stop_el",
"segment_stop_el_lookup[MB_SEGMENT]"),
Assign("segment_el_count",
"segment_stop_el-segment_start_el"),
])),
S("__syncthreads()")
])
from hedge.backends.cuda.tools import get_load_code
f_body.extend(
get_load_code(dest="smem_matrix",
base=(
"gmem_matrix + MB_SEGMENT*MATRIX_SEGMENT_BYTES"),
bytes="MATRIX_SEGMENT_BYTES",
descr="load matrix segment") +
[S("__syncthreads()")])
# ---------------------------------------------------------------------
def get_batched_fetch_mat_mul_code(el_fetch_count):
result = []
dofs = range(self.plan.preimage_dofs_per_el)
for load_segment_start in range(0, self.plan.preimage_dofs_per_el,
self.plan.segment_size):
result.extend([S("__syncthreads()")] + [
Assign(
"dof_buffer[PAR_MB_NR][%d][SEGMENT_DOF]" %
inl, "fp_tex1Dfetch(in_vector_tex, "
"GLOBAL_MB_PREIMG_DOF_BASE"
" + %d*ALIGNED_PREIMAGE_DOFS_PER_MB"
" + (segment_start_el)*PREIMAGE_DOFS_PER_EL + %d + SEGMENT_DOF)"
% (inl, load_segment_start))
for inl in range(par.inline)
] + [
S("__syncthreads()"),
Line(),
])
for dof in dofs[load_segment_start:load_segment_start +
self.plan.segment_size]:
for inl in range(par.inline):
result.append(
S("result%d += "
"smem_matrix[SEGMENT_DOF*MATRIX_COLUMNS + %d]"
"*"
"dof_buffer[PAR_MB_NR][%d][%d]" %
(inl, dof, inl, dof - load_segment_start)))
result.append(Line())
return result
from hedge.backends.cuda.tools import unroll
def get_direct_tex_mat_mul_code():
return (
[POD(float_type, "fof%d" % inl) for inl in range(par.inline)] +
[POD(float_type, "lm"), Line()] + unroll(
lambda j: [
Assign(
"fof%d" % inl,
"fp_tex1Dfetch(in_vector_tex, "
"GLOBAL_MB_PREIMG_DOF_BASE"
" + %(inl)d * ALIGNED_PREIMAGE_DOFS_PER_MB"
" + mb_el*PREIMAGE_DOFS_PER_EL+%(j)s)" % {
"j": j,
"inl": inl,
"row": "SEGMENT_DOF"
},
) for inl in range(par.inline)
] + [
Assign(
"lm",
"smem_matrix["
"%(row)s*MATRIX_COLUMNS + %(j)s]" % {
"j": j,
"row": "SEGMENT_DOF"
},
)
] + [
S("result%(inl)d += fof%(inl)d*lm" % {"inl": inl})
for inl in range(par.inline)
],
total_number=self.plan.preimage_dofs_per_el,
max_unroll=self.plan.max_unroll) + [Line()])
def get_mat_mul_code(el_fetch_count):
if el_fetch_count == 1:
return get_batched_fetch_mat_mul_code(el_fetch_count)
else:
return get_direct_tex_mat_mul_code()
def mat_mul_outer_loop(fetch_count):
if with_scaling:
inv_jac_multiplier = (
"fp_tex1Dfetch(scaling_tex,"
"(GLOBAL_MB_NR + %(inl)d)*MB_EL_COUNT + mb_el)")
else:
inv_jac_multiplier = "1"
write_condition = "MB_DOF < DOFS_PER_EL*MB_EL_COUNT"
if self.with_index_check:
write_condition += " && GLOBAL_MB_NR < microblock_count"
return For(
"unsigned short seq_mb_number = 0",
"seq_mb_number < SEQ_MB_COUNT", "++seq_mb_number",
Block([
Initializer(POD(float_type, "result%d" % inl), 0)
for inl in range(par.inline)
] + [Line()] + get_mat_mul_code(fetch_count) + [
If(
write_condition,
Block([
Assign(
"out_vector[GLOBAL_MB_DOF_BASE"
" + %d*ALIGNED_DOFS_PER_MB"
" + MB_DOF]" % inl, "result%d * %s" %
(inl, (inv_jac_multiplier % {
"inl": inl
}))) for inl in range(par.inline)
]))
]))
if self.plan.use_prefetch_branch:
from cgen import make_multiple_ifs
f_body.append(
make_multiple_ifs([
("segment_el_count == %d" % fetch_count,
mat_mul_outer_loop(fetch_count)) for fetch_count in range(
1,
self.plan.max_elements_touched_by_segment() + 1)
]))
else:
f_body.append(mat_mul_outer_loop(0))
# finish off ----------------------------------------------------------
cmod.append(FunctionBody(f_decl, f_body))
if not for_benchmark and "cuda_dump_kernels" in discr.debug:
from hedge.tools import open_unique_debug_file
open_unique_debug_file(self.plan.debug_name,
".cu").write(str(cmod))
mod = SourceModule(
cmod,
keep="cuda_keep_kernels" in discr.debug,
#options=["--maxrregcount=12"]
)
func = mod.get_function("apply_el_local_mat_smem_mat")
if self.plan.debug_name in discr.debug:
print "%s: lmem=%d smem=%d regs=%d" % (
self.plan.debug_name, func.local_size_bytes,
func.shared_size_bytes, func.num_regs)
in_vector_texref = mod.get_texref("in_vector_tex")
texrefs = [in_vector_texref]
if with_scaling:
scaling_texref = mod.get_texref("scaling_tex")
texrefs.append(scaling_texref)
else:
scaling_texref = None
func.prepare("PPPI",
block=(self.plan.segment_size,
self.plan.parallelism.parallel, 1),
texrefs=texrefs)
return func, in_vector_texref, scaling_texref