def generic_fmmv(proc_idx, queue, device_id):
a: ArgsFmmv = queue.get()
X1, X2, v, out = a.X1, a.X2, a.v, a.out
kernel, max_mem = a.kernel, a.max_mem
dtype = X1.dtype
ntot, dtot = X1.size()
M, T = v.size()
# GPU Memory Usage:
# ker_gpu : n*M
# v_gpu : M*T
# X1s_gpu : n*d
# X2s_gpu : M*d
# mmv_gpu : n*T
# ----------
# total : n*d + n*(M+T) + d*M + M*T
avail_mem = max_mem / sizeof_dtype(dtype)
n, d = select_dim_over_d(maxD=dtot,
maxN=ntot,
coef_nd=1,
coef_n=M + T,
coef_d=M,
rest=M * T,
tot=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
ker_gpu = torch.empty(n, M, dtype=dtype, device=ddev)
v_gpu = v.to(device=ddev) # M x T
X1s_gpu = create_same_stride((n, d), X1, dtype, ddev)
X2s_gpu = create_same_stride((M, d), X2, dtype, ddev)
mmv_gpu = create_same_stride((n, T), out, dtype, ddev)
for i in range(0, ntot, n):
ic = min(n, ntot - i)
ddd = kernel._prepare(X1.narrow(0, i, ic), X2)
c_g_ker = ker_gpu.narrow(0, 0, ic)
c_g_ker.fill_(0.0)
for k in range(0, dtot, d):
kc = min(d, dtot - k)
c_g_X1s = copy_to_device_noorder(ic, kc, X1, i, k, X1s_gpu, 0,
0)
c_g_X2s = copy_to_device_noorder(M, kc, X2, 0, k, X2s_gpu, 0,
0)
kernel._apply(c_g_X1s, c_g_X2s.T, c_g_ker)
kernel._finalize(c_g_ker, ddd)
# Multiply by the vector v
c_g_mmv = mmv_gpu[:ic, :]
torch.mm(c_g_ker, v_gpu, out=c_g_mmv) # n x T
# Copy back to host
copy_to_host_noorder(ic, T, c_g_mmv, 0, 0, out, i, 0)
return out
def _sparse_fmm(proc_idx, queue, device_id):
a: ArgsFmm = queue.get()
X1: SparseTensor = a.X1
X2: SparseTensor = a.X2
out = a.out
kernel, gpu_dtype = a.kernel, a.gpu_dtype
max_mem = a.max_mem
ntot, dtot = X1.shape
mtot = X2.size(0)
avail_mem = max_mem / sizeof_dtype(gpu_dtype)
# Memory usage:
# X1_chunk : ntot + 2 * D * ntot * density
# X2_chunk : dtot + 2 * D * mtot * density (because is transposed)
# sparse_out : ntot + 2 * ntot * mtot * density (assume density=1 here)
# ker_gpu : mtot * ntot
n, m = select_dim_over_nm_v2(max_n=ntot,
max_m=mtot,
coef_nm=3,
coef_n=2 + 2 * dtot * X1.density,
coef_m=2 * dtot * X2.density,
rest=dtot,
max_mem=avail_mem)
tc_device = torch.device('cuda:%d' % (int(device_id)))
with torch.cuda.device(tc_device):
# Initialize GPU buffers
g_out = create_same_stride((n, m), out, gpu_dtype, tc_device)
cpu_buf = None
if X1.dtype != gpu_dtype:
cpu_buf = create_same_stride((n, m),
out,
gpu_dtype,
'cpu',
pin_memory=True)
for j in range(0, mtot, m):
jc = min(m, mtot - j)
X2_chunk = X2.narrow_rows(j, jc).to(dtype=gpu_dtype)
X2_chunk_d = SparseTensor.from_scipy(
X2_chunk.transpose_csc().to_scipy().tocsr(copy=False)) \
.index_to_int() \
.to(device=tc_device)
for i in range(0, ntot, n):
ic = min(n, ntot - i)
X1_chunk = X1.narrow_rows(i, ic).to(dtype=gpu_dtype)
X1_chunk_d = X1_chunk.index_to_int().to(device=tc_device)
cur_g_out = g_out.narrow(0, 0, ic).narrow(1, 0, jc)
cur_g_out.fill_(0.0)
ddd = kernel._prepare_sparse(X1_chunk, X2_chunk)
cur_g_out = kernel._apply_sparse(X1_chunk_d, X2_chunk_d,
cur_g_out)
cur_g_out = kernel._finalize(cur_g_out, ddd)
copy_to_host_noorder(ic, jc, cur_g_out, 0, 0, out, i, j,
cpu_buf)
del ddd, X1_chunk_d, X1_chunk
del X2_chunk, X2_chunk_d
del g_out
return out
def sparse_fmmv(proc_idx, queue, device_id):
a: ArgsFmmv = queue.get()
X1: SparseTensor = a.X1
X2: SparseTensor = a.X2
v, out = a.v, a.out
kernel, max_mem = a.kernel, a.max_mem
dtype = X1.dtype
ntot, dtot = X1.shape
mtot, T = v.size()
avail_mem = max_mem / sizeof_dtype(dtype)
# Memory needs:
# X1_chunk : N + 2*D*N*density
# X2_chunk : D + 2*D*M*density (because is transposed)
# sparse_out : N + 2*N*M*(density) (assume density = 1)
# ker_gpu : M*N
# mmv_gpu : N*T
# v_gpu : M*T
# Other: GPU buffer
n, m = select_dim_over_m(
maxM=mtot,
maxN=ntot,
tot=avail_mem,
coef_nm=3,
coef_n=2 + 2 * dtot * X1.density + T,
coef_m=2 * dtot * X2.density + T,
rest=dtot,
)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
v_gpu = v.to(device=ddev) # M x T
mmv_gpu = create_same_stride((n, T), out, dtype, ddev)
# ker_gpu should be fortran-ordered due to cusparse csr2dense function
ker_gpu = create_fortran((n, m), dtype=dtype, device=ddev)
for i in range(0, ntot, n):
ic = min(n, ntot - i)
cur_mmv_gpu = mmv_gpu[:ic] # n x T
cur_mmv_gpu.fill_(0.0)
X1_chunk = X1.narrow_rows(i, ic)
X1_chunk_d = X1_chunk.index_to_int().to(device=ddev)
for j in range(0, mtot, m):
jc = min(m, mtot - j)
X2_chunk = X2.narrow_rows(j, jc)
# Prepare sparse on CPU
ddd = kernel._prepare_sparse(X1_chunk, X2_chunk)
# Transpose X2-chunk and convert it to CSR. This uses lots of RAM
X2_chunk_d = SparseTensor.from_scipy(
X2_chunk.transpose_csc().to_scipy().tocsr(copy=False)) \
.index_to_int() \
.to(device=ddev)
cur_ker_gpu = ker_gpu[:ic, :jc]
cur_ker_gpu.fill_(0.0)
# Run the matrix multiplication (kernel apply)
cur_ker_gpu = kernel._apply_sparse(X1_chunk_d, X2_chunk_d,
cur_ker_gpu)
cur_ker_gpu = kernel._finalize(cur_ker_gpu, ddd)
# Multiply by the vector v
cur_mmv_gpu.addmm_(cur_ker_gpu, v_gpu.narrow(0, j, jc))
del ddd, X2_chunk, X2_chunk_d
# send result to CPU
copy_to_host_noorder(ic, T, cur_mmv_gpu, 0, 0, out, i, 0)
del X1_chunk, X1_chunk_d
return out
def distk_fdmmv(proc_idx, queue, device_id):
a: ArgsFdmmv = queue.get()
X1, X2, v, w, out = a.X1, a.X2, a.v, a.w, a.out
kernel: L2DistanceKernel = a.kernel
max_mem = a.max_mem
N, D = X1.size()
M = X2.size(0)
T = v.size(1) if v is not None else w.size(1)
dtype = X1.dtype
# Memory usage:
# v : M x T
# K : n x M
# X1ss : n x d
# X2s : M x d
# Kv : n x T
# out : M x T
# sq1 : n x 1
# sq2 : M x 1
# ------------
# total : n*d + M*d + n*(M + T + 1) + 2*M*T + M
avail_mem = max_mem / sizeof_dtype(dtype)
# FIXME: There seems to be a bug where if we let avail_mem like it is
# for 32-bit data-types some copy fails. In such case we need
# to free up some more memory and then everything runs fine.
rest_coef = 2 * M * T if v is not None else M * T
n, d = select_dim_over_d(maxD=D,
maxN=N,
coef_nd=1,
coef_n=M + T + 1,
coef_d=M,
rest=rest_coef + M,
tot=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
s1 = tcd.Stream()
s2 = tcd.Stream()
with tcd.device(ddev), tcd.stream(s1):
if v is not None:
v_gpu = create_same_stride((M, T), v, dtype, ddev)
copy_to_device_noorder(M, T, v, 0, 0, v_gpu, 0, 0)
K_gpu = create_same_stride((n, M), X1, dtype, ddev)
X1ss_gpu = create_same_stride((n, d), X1, dtype, ddev)
X2s_gpu = create_same_stride((M, d), X2, dtype, ddev)
Kv_gpu = create_same_stride((n, T), X1, dtype, ddev)
if out.is_cuda:
out_gpu = out
else:
out_gpu = create_same_stride((M, T), out, dtype, ddev)
out_gpu.fill_(0.0)
sq1_gpu = create_same_stride((n, ), X1, dtype, ddev)
sq2_gpu = create_same_stride((M, ), X1, dtype, ddev)
#if (d == D):
# with torch.cuda.stream(s2):
# cur_X2s_gpu = copy_to_device_noorder(M, d, X2, 0, 0, X2s_gpu, 0, 0, s=s2)
# torch.norm(cur_X2s_gpu, p=2, dim=1, keepdim=True, out=sq2_gpu).pow_(2)
for i in range(0, N, n):
nb = min(N - i, n)
cur_K_gpu = K_gpu.narrow(0, 0, nb) # nb x M
cur_K_gpu.fill_(0.0)
for j in range(0, D, d):
db = min(D - j, d)
# Parallelize two matrix transfers (probably pointless)
#if d < D:
with torch.cuda.stream(s2):
cur_X2s_gpu = copy_to_device_noorder(M,
db,
X2,
0,
j,
X2s_gpu,
0,
0,
s=s2)
torch.norm(cur_X2s_gpu,
p=2,
dim=1,
keepdim=True,
out=sq2_gpu).pow_(2)
cur_X1ss_gpu = copy_to_device_noorder(nb,
db,
X1,
i,
j,
X1ss_gpu,
0,
0,
s=s1)
torch.norm(cur_X1ss_gpu, p=2, dim=1, keepdim=True,
out=sq1_gpu).pow_(2)
s2.synchronize()
s1.synchronize()
cur_K_gpu.addmm_(mat1=cur_X1ss_gpu,
mat2=cur_X2s_gpu.T,
alpha=-2.0)
cur_K_gpu.add_(sq1_gpu)
cur_K_gpu.add_(sq2_gpu.T)
cur_K_gpu.clamp_min_(0)
cur_K_gpu = kernel._transform(cur_K_gpu)
if w is not None:
# Copy split w to GPU into cur_Kv_gpu,
cur_Kv_gpu = copy_to_device_noorder(nb,
T,
w,
i,
0,
Kv_gpu,
0,
0,
s=s1) # n x T
if v is not None:
cur_Kv_gpu.addmm_(cur_K_gpu, v_gpu)
else:
# v cannot be None if w is None
cur_Kv_gpu = Kv_gpu.narrow(0, 0, nb) # n x T
torch.mm(cur_K_gpu, v_gpu, out=cur_Kv_gpu) # n x T
# Multiply transposed kernel with the Kv result.
out_gpu.addmm_(cur_K_gpu.T, cur_Kv_gpu) # M x T
s1.synchronize()
s1.synchronize()
if not out.is_cuda:
copy_to_host_noorder(M, T, out_gpu, 0, 0, out, 0, 0)
return out
def distk_fmmv(proc_idx, queue, device_id):
a: ArgsFmmv = queue.get()
X1, X2, v, out = a.X1, a.X2, a.v, a.out
kernel: L2DistanceKernel = a.kernel
max_mem = a.max_mem
N, D = X1.shape
M = X2.shape[0]
T = v.shape[1]
dtype = X1.dtype
# GPU memory usage:
# X1s : n x D
# X2s : m x D
# vs : m x T
# nm : n x m
# out : n x T
# -----------
# total: n*m + n * (D + T) + m * (D + T) = R
avail_mem = max_mem / sizeof_dtype(dtype)
#if sizeof_dtype(dtype) == 4:
# avail_mem /= 2
n, m = select_dim_over_m(maxM=M,
maxN=N,
coef_nm=1.0,
coef_n=D + T,
coef_m=D + T,
tot=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
nm_gpu = create_same_stride((n, m), X1, dtype, ddev)
out_gpu = create_same_stride((n, T), out, dtype, ddev)
X1s_gpu = create_same_stride((n, D), X1, dtype, ddev)
X2s_gpu = create_same_stride((m, D), X2, dtype, ddev)
vs_gpu = create_same_stride((m, T), v, dtype, ddev)
for i in range(0, N, n):
nb = min(n, N - i)
cur_X1s_gpu = copy_to_device_noorder(nb, D, X1, i, 0, X1s_gpu, 0,
0)
sq1 = torch.norm(cur_X1s_gpu, p=2, dim=1, keepdim=True).pow_(2)
cur_out_gpu = out_gpu.narrow(0, 0, nb) # n x T
cur_out_gpu.fill_(0.0)
for j in range(0, M, m):
mb = min(m, M - j)
cur_X2s_gpu = copy_to_device_noorder(mb, D, X2, j, 0, X2s_gpu,
0, 0)
cur_vs_gpu = copy_to_device_noorder(mb, T, v, j, 0, vs_gpu, 0,
0) # m x T
cur_nm_gpu = nm_gpu[:nb, :mb] # n x m
sq2 = torch.norm(cur_X2s_gpu, p=2, dim=1, keepdim=True).pow_(2)
torch.mm(cur_X1s_gpu, cur_X2s_gpu.T, out=cur_nm_gpu)
cur_nm_gpu.mul_(-2.0)
cur_nm_gpu.add_(sq1)
cur_nm_gpu.add_(sq2.T)
cur_nm_gpu.clamp_min_(0)
kernel._transform(cur_nm_gpu)
# Multiply by the vector v
# FIXME: This is the cause of mapping errors in case of float32 calculations.
cur_out_gpu.addmm_(cur_nm_gpu, cur_vs_gpu) # n x T
# send result to CPU
copy_to_host_noorder(nb, T, out_gpu, 0, 0, out, i, 0)
return out
def _generic_fmm(proc_idx, queue, device_id):
a: ArgsFmm = queue.get()
X1: torch.Tensor = a.X1
X2: torch.Tensor = a.X2
out = a.out
kernel, gpu_dtype = a.kernel, a.gpu_dtype
max_mem = a.max_mem
ntot, dtot = X1.shape
mtot = X2.shape[0]
# This function is slightly faster if we limit the sizes
# of the processed blocks slightly. Especially when doing
# a cold run since pinned-memory allocation is extremely slow.
# We don't want to do it if we're memory constrained though.
if max_mem > 4 * 2**30:
max_mem /= 4
avail_mem = max_mem / sizeof_dtype(gpu_dtype)
# Memory usage:
# - gOut : n x m
# - g_ssX1 : n x d
# - g_sX2 : m x d
# total : n*d + m*d + n*m
n, d, m = select_dim_fMM(avail_mem, ntot, dtot, mtot)
tc_device = torch.device('cuda:%d' % (int(device_id)))
with torch.cuda.device(tc_device):
# Initialize GPU buffers
g_out = create_same_stride((n, m), out, gpu_dtype, tc_device)
g_X1d = create_same_stride((n, d), X1, gpu_dtype, tc_device)
g_X2d = create_same_stride((m, d), X2, gpu_dtype, tc_device)
cpu_buf = None
if X1.dtype != gpu_dtype:
cpu_buf = create_same_stride((n, m),
out,
gpu_dtype,
'cpu',
pin_memory=True)
for j in range(0, mtot, m):
jc = min(m, mtot - j)
X2_chunk = cast_tensor(X2.narrow(0, j, jc),
dtype=gpu_dtype,
warn=False).pin_memory()
for i in range(0, ntot, n):
ic = min(n, ntot - i)
if _gpu_tns_same_memory(X1, X2) and j < i:
out[i:i + ic, j:j + jc].copy_(out[j:j + jc, i:i + ic].T)
else:
X1_chunk = cast_tensor(X1.narrow(0, i, ic),
dtype=gpu_dtype,
warn=False).pin_memory()
ddd = kernel._prepare(X1_chunk, X2_chunk)
cur_g_out = g_out.narrow(0, 0, ic).narrow(1, 0, jc)
cur_g_out.fill_(0.0)
for k in range(0, dtot, d):
kc = min(d, dtot - k)
# Move to GPU
cur_g_X1d = g_X1d.narrow(0, 0, ic).narrow(1, 0, kc)
cur_g_X1d.copy_(X1_chunk.narrow(1, k, kc))
cur_g_X2d = g_X2d.narrow(0, 0, jc).narrow(1, 0, kc)
cur_g_X2d.copy_(X2_chunk.narrow(1, k, kc))
# Apply
a.kernel._apply(cur_g_X1d, cur_g_X2d.T, cur_g_out)
a.kernel._finalize(cur_g_out, ddd)
copy_to_host_noorder(ic, jc, cur_g_out, 0, 0, out, i, j,
cpu_buf)
del ddd
del g_out, g_X1d, g_X2d
return out
def distk_fdmmv(proc_idx, queue, device_id):
a: ArgsFdmmv = queue.get()
X1, X2, v, w, out = a.X1, a.X2, a.v, a.w, a.out
kernel: L2DistanceKernel = a.kernel
max_mem = a.max_mem
N, D = X1.size()
M = X2.size(0)
T = v.shape[1] if v is not None else w.shape[1]
dtype = X1.dtype
cuda_inputs = X1.is_cuda
# Memory usage:
# v : M x T
# K : n x M
# X1ss : n x d
# X2s : M x d
# Kv : n x T
# out : M x T
# sq1 : n x 1
# sq2 : M x 1
# ------------
# total : n*d + M*d + n*(M + T + 1) + 2*M*T + M
avail_mem = max_mem / sizeof_dtype(dtype)
rest_coef = 2 * M * T if v is not None else M * T
n, d = select_dim_over_nd(max_n=N,
max_d=D,
coef_nd=1,
coef_n=M + T + 1,
coef_d=M,
rest=rest_coef + M,
max_mem=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
s1 = tcd.Stream(ddev)
s2 = tcd.Stream(ddev)
with tcd.device(ddev), tcd.stream(s1):
# First collect necessary memory
mem_needed = n * M + n * T + n + M
if not cuda_inputs:
mem_needed += n * d + M * d
if v is not None:
mem_needed += M * T
if not out.is_cuda:
mem_needed += M * T
# Create flat tensor
flat_gpu_tn = torch.empty(size=(mem_needed, ),
dtype=dtype,
device=ddev)
# Extract the sub-tensors
flat_offset = 0
if v is not None:
if not cuda_inputs:
v_gpu = extract_same_stride(flat_gpu_tn,
size=(M, T),
other=v,
offset=flat_offset)
flat_offset += np.prod(v_gpu.shape)
copy_to_device_noorder(M, T, v, 0, 0, v_gpu, 0, 0)
else:
v_gpu = v
K_gpu = extract_same_stride(flat_gpu_tn,
size=(n, M),
other=X1,
offset=flat_offset)
flat_offset += np.prod(K_gpu.shape)
Kv_gpu = extract_same_stride(flat_gpu_tn,
size=(n, T),
other=X1,
offset=flat_offset)
flat_offset += np.prod(Kv_gpu.shape)
if out.is_cuda:
out_gpu = out
else:
out_gpu = extract_same_stride(flat_gpu_tn,
size=(M, T),
other=out,
offset=flat_offset)
flat_offset += np.prod(out_gpu.shape)
out_gpu.fill_(0.0)
if not cuda_inputs:
X1ss_gpu = extract_same_stride(flat_gpu_tn,
size=(n, d),
other=X1,
offset=flat_offset)
flat_offset += np.prod(X1ss_gpu.shape)
X2s_gpu = extract_same_stride(flat_gpu_tn,
size=(M, d),
other=X2,
offset=flat_offset)
flat_offset += np.prod(X2s_gpu.shape)
sq1_gpu = extract_same_stride(flat_gpu_tn,
size=(n, ),
other=X1,
offset=flat_offset)
flat_offset += np.prod(sq1_gpu.shape)
sq2_gpu = extract_same_stride(flat_gpu_tn,
size=(M, ),
other=X1,
offset=flat_offset)
for i in range(0, N, n):
nb = min(N - i, n)
cur_K_gpu = K_gpu[:nb] # nb x M
cur_K_gpu.fill_(0.0)
for j in range(0, D, d):
db = min(D - j, d)
s1.synchronize(
) # need that the add_(sq2_gpu.T) op is complete to avoid overwrite
# Parallelize two matrix transfers
with tcd.stream(s2):
if cuda_inputs:
cur_X2s_gpu = X2[:, j:j + db]
else:
cur_X2s_gpu = copy_to_device_noorder(M,
db,
X2,
0,
j,
X2s_gpu,
0,
0,
s=s2)
torch.norm(cur_X2s_gpu,
p=2,
dim=1,
keepdim=True,
out=sq2_gpu).pow_(2)
if cuda_inputs:
cur_X1ss_gpu = X1[i:i + nb, j:j + db]
else:
cur_X1ss_gpu = copy_to_device_noorder(nb,
db,
X1,
i,
j,
X1ss_gpu,
0,
0,
s=s1)
torch.norm(cur_X1ss_gpu, p=2, dim=1, keepdim=True,
out=sq1_gpu).pow_(2)
s2.synchronize(
) # need that cur_X2s_gpu and sq2_gpu are available.
cur_K_gpu.addmm_(mat1=cur_X1ss_gpu,
mat2=cur_X2s_gpu.T,
alpha=-2.0)
cur_K_gpu.add_(sq1_gpu)
cur_K_gpu.add_(sq2_gpu.T)
cur_K_gpu.clamp_min_(0)
cur_K_gpu = kernel._transform(cur_K_gpu)
if w is not None:
cur_Kv_gpu = copy_to_device_noorder(nb,
T,
w,
i,
0,
Kv_gpu,
0,
0,
s=s1) # n x T
if v is not None:
cur_Kv_gpu.addmm_(cur_K_gpu, v_gpu)
else:
# v cannot be None if w is None
cur_Kv_gpu = Kv_gpu.narrow(0, 0, nb) # n x T
torch.mm(cur_K_gpu, v_gpu, out=cur_Kv_gpu) # n x T
# Multiply transposed kernel with the Kv result.
out_gpu.addmm_(cur_K_gpu.T, cur_Kv_gpu) # M x T
if not out.is_cuda:
copy_to_host_noorder(M, T, out_gpu, 0, 0, out, 0, 0)
s1.synchronize()
return out
def sparse_fmmv(proc_idx, queue, device_id):
a: ArgsFmmv = queue.get()
X1: SparseTensor = a.X1
X2: SparseTensor = a.X2
v, out = a.v, a.out
kernel, max_mem = a.kernel, a.max_mem
dtype = X1.dtype
cuda_inputs = X1.is_cuda
ntot, dtot = X1.shape
mtot, T = v.size()
avail_mem = max_mem / sizeof_dtype(dtype)
# Memory needs:
# X1_chunk : N + 2*D*N*density
# X2_chunk : D + 2*D*M*density (because is transposed)
# sparse_out : N + 2*N*M*(density) (assume density = 1)
# ker_gpu : M*N
# mmv_gpu : N*T
# v_gpu : M*T
# Other: GPU buffer
n, m = select_dim_over_nm_v2(max_n=ntot,
max_m=mtot,
coef_nm=3,
coef_n=2 + 2 * dtot * X1.density + T,
coef_m=2 * dtot * X2.density + T,
rest=dtot,
max_mem=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
# First collect necessary memory
mem_needed = mtot * T + n * T + n * m
# Create flat tensor
flat_gpu_tn = torch.empty(size=(mem_needed, ),
dtype=dtype,
device=ddev)
# Extract the sub-tensors
flat_offset = 0
v_gpu = extract_same_stride(flat_gpu_tn,
size=(mtot, T),
other=v,
offset=flat_offset)
flat_offset += np.prod(v_gpu.shape)
copy_to_device_noorder(mtot, T, v, 0, 0, v_gpu, 0, 0)
mmv_gpu = extract_same_stride(flat_gpu_tn,
size=(n, T),
other=out,
offset=flat_offset)
flat_offset += np.prod(mmv_gpu.shape)
# ker_gpu should be fortran-ordered due to cusparse csr2dense function
ker_gpu = extract_fortran(flat_gpu_tn, size=(n, m), offset=flat_offset)
flat_offset += np.prod(ker_gpu.shape)
for i in range(0, ntot, n):
ic = min(n, ntot - i)
cur_mmv_gpu = mmv_gpu[:ic] # n x T
cur_mmv_gpu.fill_(0.0)
X1_chunk = X1.narrow_rows(i, ic)
X1_chunk_d = X1_chunk.index_to_int().to(device=ddev)
for j in range(0, mtot, m):
jc = min(m, mtot - j)
X2_chunk = X2.narrow_rows(j, jc)
# Prepare sparse on CPU
ddd = kernel._prepare_sparse(X1_chunk, X2_chunk)
# Transpose X2-chunk and convert it to CSR. This uses lots of RAM
X2_chunk_d = SparseTensor.from_scipy(
X2_chunk.transpose_csc().to_scipy().tocsr(copy=False)) \
.index_to_int() \
.to(device=ddev)
cur_ker_gpu = ker_gpu[:ic, :jc]
cur_ker_gpu.fill_(0.0)
# Run the matrix multiplication (kernel apply)
cur_ker_gpu = kernel._apply_sparse(X1_chunk_d, X2_chunk_d,
cur_ker_gpu)
cur_ker_gpu = kernel._finalize(cur_ker_gpu, ddd)
# Multiply by the vector v
cur_mmv_gpu.addmm_(cur_ker_gpu, v_gpu.narrow(0, j, jc))
del ddd, X2_chunk, X2_chunk_d
# send result to CPU
if not cuda_inputs:
copy_to_host_noorder(ic, T, cur_mmv_gpu, 0, 0, out, i, 0)
del X1_chunk, X1_chunk_d
return out
def distk_fmmv(proc_idx, queue, device_id):
a: ArgsFmmv = queue.get()
X1, X2, v, out = a.X1, a.X2, a.v, a.out
kernel: L2DistanceKernel = a.kernel
max_mem = a.max_mem
N, D = X1.shape
M = X2.shape[0]
T = v.shape[1]
dtype = X1.dtype
cuda_inputs = X1.is_cuda
# GPU memory usage:
# X1s : n x D
# X2s : m x D
# vs : m x T
# nm : n x m
# out : n x T
# -----------
# total: n*m + n * (D + T) + m * (D + T) = R
avail_mem = max_mem / sizeof_dtype(dtype)
n, m = select_dim_over_nm_v2(max_n=N,
max_m=M,
coef_nm=1,
coef_n=D + T,
coef_m=D + T,
rest=0,
max_mem=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
mem_needed = n * m
if not cuda_inputs:
mem_needed += n * T + n * D + m * D + m * T
flat_gpu_tn = torch.empty(size=(mem_needed, ),
dtype=dtype,
device=ddev)
flat_offset = 0
nm_gpu = extract_same_stride(flat_gpu_tn,
size=(n, m),
other=X1,
offset=flat_offset)
flat_offset += np.prod(nm_gpu.shape)
if not cuda_inputs:
out_gpu = extract_same_stride(flat_gpu_tn,
size=(n, T),
other=out,
offset=flat_offset)
flat_offset += np.prod(out_gpu.shape)
X1s_gpu = extract_same_stride(flat_gpu_tn,
size=(n, D),
other=X1,
offset=flat_offset)
flat_offset += np.prod(X1s_gpu.shape)
X2s_gpu = extract_same_stride(flat_gpu_tn,
size=(m, D),
other=X2,
offset=flat_offset)
flat_offset += np.prod(X2s_gpu.shape)
vs_gpu = extract_same_stride(flat_gpu_tn,
size=(m, T),
other=v,
offset=flat_offset)
flat_offset += np.prod(vs_gpu.shape)
for i in range(0, N, n):
nb = min(n, N - i)
if cuda_inputs:
cur_X1s_gpu = X1.narrow(0, i, nb) # n x D
else:
cur_X1s_gpu = copy_to_device_noorder(nb, D, X1, i, 0, X1s_gpu,
0, 0)
sq1 = torch.norm(cur_X1s_gpu, p=2, dim=1, keepdim=True).pow_(2)
if cuda_inputs:
cur_out_gpu = out.narrow(0, i, nb) # n x T
else:
cur_out_gpu = out_gpu.narrow(0, 0, nb) # n x T
cur_out_gpu.fill_(0.0)
for j in range(0, M, m):
mb = min(m, M - j)
if cuda_inputs:
cur_X2s_gpu = X2.narrow(0, j, mb) # m x D
cur_vs_gpu = v.narrow(0, j, mb) # m x T
else:
cur_X2s_gpu = copy_to_device_noorder(
mb, D, X2, j, 0, X2s_gpu, 0, 0) # m x D
cur_vs_gpu = copy_to_device_noorder(
mb, T, v, j, 0, vs_gpu, 0, 0) # m x T
cur_nm_gpu = nm_gpu[:nb, :mb] # n x m
sq2 = torch.norm(cur_X2s_gpu, p=2, dim=1, keepdim=True).pow_(2)
torch.mm(cur_X1s_gpu, cur_X2s_gpu.T, out=cur_nm_gpu)
cur_nm_gpu.mul_(-2.0)
cur_nm_gpu.add_(sq1)
cur_nm_gpu.add_(sq2.T)
cur_nm_gpu.clamp_min_(0)
kernel._transform(cur_nm_gpu)
# Multiply by the vector v
cur_out_gpu.addmm_(cur_nm_gpu, cur_vs_gpu) # n x T
if not cuda_inputs:
# send result to CPU
copy_to_host_noorder(nb, T, out_gpu, 0, 0, out, i, 0)
return out
def generic_fmmv(proc_idx, queue, device_id):
a: ArgsFmmv = queue.get()
X1, X2, v, out = a.X1, a.X2, a.v, a.out
kernel, max_mem = a.kernel, a.max_mem
dtype = X1.dtype
cuda_inputs = X1.is_cuda
ntot, dtot = X1.size()
M, T = v.size()
# GPU Memory Usage:
# ker_gpu : n*M
# v_gpu : M*T
# X1s_gpu : n*d
# X2s_gpu : M*d
# mmv_gpu : n*T
# ----------
# total : n*d + n*(M+T) + d*M + M*T
avail_mem = max_mem / sizeof_dtype(dtype)
extra_mem = kernel.extra_mem()
n, d = select_dim_over_nd(
max_n=ntot,
max_d=dtot,
coef_nd=1 + extra_mem.get('nd', 0),
coef_n=M + T + extra_mem.get('n', 0) + extra_mem.get('nm', 0) * M,
coef_d=M + extra_mem.get('d', 0) + extra_mem.get('md', 0) * M,
rest=M * T + extra_mem.get('m', 0),
max_mem=avail_mem)
ddev = torch.device('cuda:%d' % int(device_id))
with tcd.device(ddev):
# First collect necessary memory
mem_needed = n * M
if not cuda_inputs:
mem_needed += M * T + n * d + M * d + n * T
# Create flat tensor
flat_gpu_tn = torch.empty(size=(mem_needed, ),
dtype=dtype,
device=ddev)
# Extract the sub-tensors
flat_offset = 0
ker_gpu = extract_same_stride(flat_gpu_tn,
size=(n, M),
other=X1,
offset=flat_offset)
flat_offset += np.prod(ker_gpu.shape)
if not cuda_inputs:
X1s_gpu = extract_same_stride(flat_gpu_tn,
size=(n, d),
other=X1,
offset=flat_offset)
flat_offset += np.prod(X1s_gpu.shape)
X2s_gpu = extract_same_stride(flat_gpu_tn,
size=(M, d),
other=X2,
offset=flat_offset)
flat_offset += np.prod(X2s_gpu.shape)
mmv_gpu = extract_same_stride(flat_gpu_tn,
size=(n, T),
other=out,
offset=flat_offset)
flat_offset += np.prod(mmv_gpu.shape)
v_gpu = extract_same_stride(flat_gpu_tn,
size=(M, T),
other=v,
offset=flat_offset)
flat_offset += np.prod(v_gpu.shape)
copy_to_device_noorder(M, T, v, 0, 0, v_gpu, 0, 0)
else:
v_gpu = v
for i in range(0, ntot, n):
ic = min(n, ntot - i)
ddd = kernel._prepare(X1.narrow(0, i, ic), X2)
c_g_ker = ker_gpu.narrow(0, 0, ic)
c_g_ker.fill_(0.0)
for k in range(0, dtot, d):
kc = min(d, dtot - k)
if cuda_inputs:
c_g_X1s = X1[i:i + ic, k:k + kc]
c_g_X2s = X2[:, k:k + kc]
else:
c_g_X1s = copy_to_device_noorder(ic, kc, X1, i, k, X1s_gpu,
0, 0)
c_g_X2s = copy_to_device_noorder(M, kc, X2, 0, k, X2s_gpu,
0, 0)
kernel._apply(c_g_X1s, c_g_X2s.T, c_g_ker)
kernel._finalize(c_g_ker, ddd)
# Multiply by the vector v
if cuda_inputs:
c_g_mmv = out[i:i + ic, :]
else:
c_g_mmv = mmv_gpu[:ic, :]
torch.mm(c_g_ker, v_gpu, out=c_g_mmv) # n x T
# Copy back to host
if not cuda_inputs:
copy_to_host_noorder(ic, T, c_g_mmv, 0, 0, out, i, 0)
return out
def _generic_fmm(proc_idx, queue, device_id):
a: ArgsFmm = queue.get()
X1: torch.Tensor = a.X1
X2: torch.Tensor = a.X2
cuda_inputs = X1.is_cuda
out = a.out
kernel, gpu_dtype = a.kernel, a.gpu_dtype
max_mem = a.max_mem
change_dtype = gpu_dtype != X1.dtype
ntot, dtot = X1.shape
mtot = X2.shape[0]
# This function is slightly faster if we limit the sizes
# of the processed blocks slightly. Especially when doing
# a cold run since pinned-memory allocation is extremely slow.
# We don't want to do it if we're memory constrained though.
if max_mem > 4 * 2**30:
max_mem /= 4
avail_mem = max_mem / sizeof_dtype(gpu_dtype)
# Memory usage:
# - gOut : n x m
# - g_ssX1 : n x d
# - g_sX2 : m x d
# total : n*d + m*d + n*m
if cuda_inputs and not change_dtype:
# No allocation will be performed, so no need to split at all!
n, d, m = ntot, dtot, mtot
else:
n, d, m = select_dim_fMM(avail_mem, ntot, dtot, mtot)
tc_device = torch.device('cuda:%d' % (int(device_id)))
s1 = torch.cuda.Stream(device=tc_device)
with torch.cuda.device(tc_device), torch.cuda.stream(s1):
# Initialize GPU buffers
if not cuda_inputs or change_dtype:
g_X1d = create_same_stride((n, d), X1, gpu_dtype, tc_device)
g_X2d = create_same_stride((m, d), X2, gpu_dtype, tc_device)
g_out = create_same_stride((n, m), out, gpu_dtype, tc_device)
if not cuda_inputs:
cpu_buf = None
if change_dtype:
cpu_buf = create_same_stride((n, m),
out,
gpu_dtype,
'cpu',
pin_memory=True)
for j in range(0, mtot, m):
jc = min(m, mtot - j)
X2_chunk = X2.narrow(0, j, jc)
for i in range(0, ntot, n):
ic = min(n, ntot - i)
if _gpu_tns_same_memory(X1, X2) and j < i:
out[i:i + ic, j:j + jc].copy_(out[j:j + jc, i:i + ic].T)
continue
X1_chunk = X1.narrow(0, i, ic)
ddd = kernel._prepare(X1_chunk, X2_chunk)
if not cuda_inputs or change_dtype:
cur_g_out = g_out.narrow(0, 0, ic).narrow(1, 0, jc)
else:
cur_g_out = out.narrow(0, i, ic).narrow(1, j, jc)
cur_g_out.fill_(0.0)
for k in range(0, dtot, d):
kc = min(d, dtot - k)
# Move to GPU and type-convert
if (not cuda_inputs) or change_dtype:
cur_g_X1d = g_X1d.narrow(0, 0, ic).narrow(1, 0, kc)
cur_g_X1d.copy_(X1_chunk.narrow(1, k, kc))
cur_g_X2d = g_X2d.narrow(0, 0, jc).narrow(1, 0, kc)
cur_g_X2d.copy_(X2_chunk.narrow(1, k, kc))
else:
cur_g_X1d = X1_chunk.narrow(1, k, kc)
cur_g_X2d = X2_chunk.narrow(1, k, kc)
# Apply
a.kernel._apply(cur_g_X1d, cur_g_X2d.T, cur_g_out)
a.kernel._finalize(cur_g_out, ddd)
if not cuda_inputs:
copy_to_host_noorder(ic, jc, cur_g_out, 0, 0, out, i, j,
cpu_buf, s1)
elif change_dtype:
out.narrow(0, i, ic).narrow(1, j, jc).copy_(cur_g_out)
del ddd
return out
