void
spu_cpu_cpp_builtins (struct cpp_reader *pfile)
{
builtin_define_std ("__SPU__");
cpp_assert (pfile, "cpu=spu");
cpp_assert (pfile, "machine=spu");
if (spu_arch == PROCESSOR_CELLEDP)
builtin_define_std ("__SPU_EDP__");
builtin_define_std ("__vector=__attribute__((__spu_vector__))");
if (!flag_iso)
{
/* Define this when supporting context-sensitive keywords. */
cpp_define (pfile, "__VECTOR_KEYWORD_SUPPORTED__");
cpp_define (pfile, "vector=vector");
/* Initialize vector keywords. */
__vector_keyword = get_identifier ("__vector");
C_CPP_HASHNODE (__vector_keyword)->flags |= NODE_CONDITIONAL;
vector_keyword = get_identifier ("vector");
C_CPP_HASHNODE (vector_keyword)->flags |= NODE_CONDITIONAL;
/* Enable context-sensitive macros. */
cpp_get_callbacks (pfile)->macro_to_expand = spu_macro_to_expand;
}
}
void
ix86_target_macros (void)
{
/* 32/64-bit won't change with target specific options, so do the assert and
builtin_define_std calls here. */
if (TARGET_64BIT)
{
cpp_assert (parse_in, "cpu=x86_64");
cpp_assert (parse_in, "machine=x86_64");
cpp_define (parse_in, "__amd64");
cpp_define (parse_in, "__amd64__");
cpp_define (parse_in, "__x86_64");
cpp_define (parse_in, "__x86_64__");
}
else
{
cpp_assert (parse_in, "cpu=i386");
cpp_assert (parse_in, "machine=i386");
builtin_define_std ("i386");
}
ix86_target_macros_internal (ix86_isa_flags,
ix86_arch,
ix86_tune,
ix86_fpmath,
cpp_define);
}
void
ix86_target_macros (void)
{
/* 32/64-bit won't change with target specific options, so do the assert and
builtin_define_std calls here. */
if (TARGET_64BIT)
{
cpp_assert (parse_in, "cpu=x86_64");
cpp_assert (parse_in, "machine=x86_64");
cpp_define (parse_in, "__amd64");
cpp_define (parse_in, "__amd64__");
cpp_define (parse_in, "__x86_64");
cpp_define (parse_in, "__x86_64__");
if (TARGET_X32)
{
cpp_define (parse_in, "_ILP32");
cpp_define (parse_in, "__ILP32__");
}
}
else
{
cpp_assert (parse_in, "cpu=i386");
cpp_assert (parse_in, "machine=i386");
builtin_define_std ("i386");
}
cpp_define_formatted (parse_in, "__ATOMIC_HLE_ACQUIRE=%d", IX86_HLE_ACQUIRE);
cpp_define_formatted (parse_in, "__ATOMIC_HLE_RELEASE=%d", IX86_HLE_RELEASE);
ix86_target_macros_internal (ix86_isa_flags,
ix86_arch,
ix86_tune,
ix86_fpmath,
cpp_define);
}
void
sparc_target_macros (void)
{
builtin_define_std ("sparc");
if (TARGET_64BIT)
{
cpp_assert (parse_in, "cpu=sparc64");
cpp_assert (parse_in, "machine=sparc64");
}
else
{
cpp_assert (parse_in, "cpu=sparc");
cpp_assert (parse_in, "machine=sparc");
}
if (TARGET_VIS3)
{
cpp_define (parse_in, "__VIS__=0x300");
cpp_define (parse_in, "__VIS=0x300");
}
else if (TARGET_VIS2)
{
cpp_define (parse_in, "__VIS__=0x200");
cpp_define (parse_in, "__VIS=0x200");
}
else if (TARGET_VIS)
{
cpp_define (parse_in, "__VIS__=0x100");
cpp_define (parse_in, "__VIS=0x100");
}
}
void
ix86_target_macros (void)
{
/* 32/64-bit won't change with target specific options, so do the assert and
builtin_define_std calls here. */
if (TARGET_64BIT)
{
cpp_assert (parse_in, "cpu=x86_64");
cpp_assert (parse_in, "machine=x86_64");
cpp_define (parse_in, "__amd64");
cpp_define (parse_in, "__amd64__");
cpp_define (parse_in, "__x86_64");
cpp_define (parse_in, "__x86_64__");
if (TARGET_X32)
{
cpp_define (parse_in, "_ILP32");
cpp_define (parse_in, "__ILP32__");
}
}
else
{
cpp_assert (parse_in, "cpu=i386");
cpp_assert (parse_in, "machine=i386");
builtin_define_std ("i386");
}
if (!TARGET_80387)
cpp_define (parse_in, "_SOFT_FLOAT");
if (TARGET_LONG_DOUBLE_64)
cpp_define (parse_in, "__LONG_DOUBLE_64__");
if (TARGET_LONG_DOUBLE_128)
cpp_define (parse_in, "__LONG_DOUBLE_128__");
if (TARGET_128BIT_LONG_DOUBLE)
cpp_define (parse_in, "__SIZEOF_FLOAT80__=16");
else
cpp_define (parse_in, "__SIZEOF_FLOAT80__=12");
cpp_define (parse_in, "__SIZEOF_FLOAT128__=16");
cpp_define_formatted (parse_in, "__ATOMIC_HLE_ACQUIRE=%d", IX86_HLE_ACQUIRE);
cpp_define_formatted (parse_in, "__ATOMIC_HLE_RELEASE=%d", IX86_HLE_RELEASE);
cpp_define (parse_in, "__GCC_ASM_FLAG_OUTPUTS__");
ix86_target_macros_internal (ix86_isa_flags,
ix86_isa_flags2,
ix86_arch,
ix86_tune,
ix86_fpmath,
cpp_define);
cpp_define (parse_in, "__SEG_FS");
cpp_define (parse_in, "__SEG_GS");
}
void
spu_cpu_cpp_builtins (struct cpp_reader *pfile)
{
builtin_define_std ("__SPU__");
cpp_assert (pfile, "cpu=spu");
cpp_assert (pfile, "machine=spu");
if (spu_arch == PROCESSOR_CELLEDP)
builtin_define_std ("__SPU_EDP__");
builtin_define_std ("__vector=__attribute__((__spu_vector__))");
}
void check(const I& input, const K& kernel, const C& conv) const {
static_assert(etl::dimensions<I>() == 2, "Invalid number of dimensions for input of conv2_backward");
static_assert(etl::dimensions<K>() == 2, "Invalid number of dimensions for kernel of conv2_backward");
static_assert(etl::dimensions<C>() == 2, "Invalid number of dimensions for conv of conv2_backward");
cpp_assert(etl::dim(conv, 0) == s1 * (etl::dim(input, 0) - 1) + etl::dim(kernel, 0) - 2 * p1, "Invalid dimensions for conv2_backward");
cpp_assert(etl::dim(conv, 1) == s2 * (etl::dim(input, 1) - 1) + etl::dim(kernel, 1) - 2 * p2, "Invalid dimensions for conv2_backward");
cpp_unused(input);
cpp_unused(kernel);
cpp_unused(conv);
}
/* Define platform dependent macros. */
void
s390_cpu_cpp_builtins (cpp_reader *pfile)
{
struct cl_target_option opts;
cpp_assert (pfile, "cpu=s390");
cpp_assert (pfile, "machine=s390");
cpp_define (pfile, "__s390__");
if (TARGET_ZARCH)
cpp_define (pfile, "__zarch__");
if (TARGET_64BIT)
cpp_define (pfile, "__s390x__");
if (TARGET_LONG_DOUBLE_128)
cpp_define (pfile, "__LONG_DOUBLE_128__");
cl_target_option_save (&opts, &global_options);
s390_cpu_cpp_builtins_internal (pfile, &opts, NULL);
}
void train_batch(std::size_t /*epoch*/, const dll::batch<T>& data_batch, const dll::batch<L>& label_batch) {
cpp_assert(data_batch.size() == label_batch.size(), "Invalid sizes");
auto n = label_batch.size();
decltype(auto) first_layer = dbn.template layer_get<0>();
decltype(auto) first_ctx = first_layer.template get_sgd_context<dbn_t>();
decltype(auto) last_layer = dbn.template layer_get<layers - 1>();
decltype(auto) last_ctx = last_layer.template get_sgd_context<dbn_t>();
using inputs_t = typename input_batch_t<0>::type;
using outputs_t = typename output_batch_t<layers - 1>::type;
inputs_t inputs;
outputs_t labels;
//Copy inputs and labels into suitable data structure
copy_inputs(inputs, data_batch.begin(), data_batch.end());
copy_labels(labels, label_batch.begin(), label_batch.end());
//Feedforward pass
compute_outputs(inputs);
static_assert(
decay_layer_traits<decltype(last_layer)>::is_dense_layer() || decay_layer_traits<decltype(last_layer)>::is_standard_rbm_layer(),
"The last layer must be dense for SGD trainining");
//Compute the errors of the last layer
compute_last_errors(last_layer, last_ctx, labels);
//Compute the gradients of each layer
dbn.for_each_layer_rpair([](auto& r1, auto& r2) {
auto& ctx1 = r1.template get_sgd_context<dbn_t>();
auto& ctx2 = r2.template get_sgd_context<dbn_t>();
this_type::compute_gradients(r2, ctx2, ctx1.output);
this_type::compute_errors(r1, ctx1, r2, ctx2);
});
compute_gradients(first_layer, first_ctx, inputs);
//Apply gradients
dbn.for_each_layer([this, n](auto& layer) {
this->apply_gradients(layer, n);
});
}
void
spu_cpu_cpp_builtins (struct cpp_reader *pfile)
{
cpp_define (pfile, "__SPU__");
cpp_assert (pfile, "cpu=spu");
cpp_assert (pfile, "machine=spu");
if (spu_arch == PROCESSOR_CELLEDP)
cpp_define (pfile, "__SPU_EDP__");
if (cpp_get_options (pfile)->lang != CLK_ASM)
cpp_define (pfile, "__vector=__attribute__((__spu_vector__))");
switch (spu_ea_model)
{
case 32:
cpp_define (pfile, "__EA32__");
break;
case 64:
cpp_define (pfile, "__EA64__");
break;
default:
gcc_unreachable ();
}
if (!flag_iso && cpp_get_options (pfile)->lang != CLK_ASM)
{
/* Define this when supporting context-sensitive keywords. */
cpp_define (pfile, "__VECTOR_KEYWORD_SUPPORTED__");
cpp_define (pfile, "vector=vector");
/* Initialize vector keywords. */
__vector_keyword = get_identifier ("__vector");
C_CPP_HASHNODE (__vector_keyword)->flags |= NODE_CONDITIONAL;
vector_keyword = get_identifier ("vector");
C_CPP_HASHNODE (vector_keyword)->flags |= NODE_CONDITIONAL;
/* Enable context-sensitive macros. */
cpp_get_callbacks (pfile)->macro_to_expand = spu_macro_to_expand;
}
}
batch(Iterator&& first, Iterator&& last)
: first(std::forward<Iterator>(first)),
last(std::forward<Iterator>(last)) {
cpp_assert(std::distance(first, last) > 0, "Batch cannot be empty or reversed");
}
