* limitations under the License.
*******************************************************************************/
-#ifndef GEMM_U8S8S32X_INNER_PRODUCT_HPP
-#define GEMM_U8S8S32X_INNER_PRODUCT_HPP
+#ifndef GEMM_X8S8S32X_INNER_PRODUCT_HPP
+#define GEMM_X8S8S32X_INNER_PRODUCT_HPP
#include <assert.h>
#include "c_types_map.hpp"
-#include "cpu_inner_product_pd.hpp"
-#include "cpu_engine.hpp"
+#include "memory_tracking.hpp"
#include "type_helpers.hpp"
#include "utils.hpp"
-#include "scratchpad.hpp"
-#include "gemm/os_blas.hpp"
+#include "gemm/gemm.hpp"
+#include "jit_generator.hpp"
+
+#include "cpu_inner_product_pd.hpp"
namespace mkldnn {
namespace impl {
namespace cpu {
-template <impl::data_type_t dst_type>
-struct gemm_u8s8s32x_inner_product_fwd_t: public cpu_primitive_t {
+template <impl::data_type_t src_type, impl::data_type_t dst_type>
+struct gemm_x8s8s32x_inner_product_fwd_t: public cpu_primitive_t {
struct pd_t: public cpu_inner_product_fwd_pd_t {
pd_t(engine_t *engine, const inner_product_desc_t *adesc,
const primitive_attr_t *attr,
const inner_product_fwd_pd_t *hint_fwd_pd)
: cpu_inner_product_fwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}
- DECLARE_COMMON_PD_T("gemm:blas", gemm_u8s8s32x_inner_product_fwd_t);
+ DECLARE_COMMON_PD_T(src_type == data_type::u8
+ ? IGEMM_S8U8S32_IMPL_STR
+ : IGEMM_S8S8S32_IMPL_STR,
+ gemm_x8s8s32x_inner_product_fwd_t);
virtual status_t init() override {
using namespace utils;
assert(engine()->kind() == engine_kind::cpu);
bool ok = true
-#if !USE_MKL_IGEMM
- && false
-#endif
&& this->set_default_params() == status::success
&& one_of(desc()->prop_kind, prop_kind::forward_training,
prop_kind::forward_inference)
&& !has_zero_dim_memory()
- && this->desc()->src_desc.data_type == u8
+ && this->desc()->src_desc.data_type == src_type
&& this->desc()->dst_desc.data_type == dst_type
&& this->desc()->weights_desc.data_type == s8
&& IMPLICATION(this->with_bias(), utils::one_of(
attr()->post_ops_.entry_[0].is_relu(true, false))
&& dense_gemm_consitency_check(src_pd(), weights_pd(),
dst_pd());
- return ok ? status::success : status::unimplemented;
+ if (!ok) return status::unimplemented;
+
+ dst_is_acc_ = one_of(dst_type, s32, f32);
+
+ init_scratchpad();
+
+ return status::success;
}
+ bool dst_is_acc_;
+
protected:
virtual status_t set_default_params() override {
using namespace memory_format;
- if (this->src_pd_.desc()->format == any)
- {
+ if (this->src_pd_.desc()->format == any) {
if (ndims() == 4) CHECK(this->src_pd_.set_format(nhwc));
else if (ndims() == 5) CHECK(this->src_pd_.set_format(ndhwc));
else CHECK(this->src_pd_.set_format(nc));
}
if (this->dst_pd_.desc()->format == any)
CHECK(this->dst_pd_.set_format(nc));
- if (this->weights_pd_.desc()->format == any)
- {
+ if (this->weights_pd_.desc()->format == any) {
if (ndims() == 4) CHECK(this->weights_pd_.set_format(hwio));
else if (ndims() == 5) CHECK(this->weights_pd_.set_format(dhwio));
else CHECK(this->weights_pd_.set_format(io));
}
if (this->bias_pd_.desc()->format == any)
CHECK(this->bias_pd_.set_format(x));
+
return status::success;
}
+
+ private:
+ void init_scratchpad() {
+ if (!dst_is_acc_) {
+ auto scratchpad = scratchpad_registry().registrar();
+ scratchpad.book(
+ memory_tracking::names::key_iprod_int_dat_in_acc_dt,
+ sizeof(acc_data_t) * MB() * OC());
+ }
+ }
};
- gemm_u8s8s32x_inner_product_fwd_t(const pd_t *pd, const input_vector &inputs,
+ gemm_x8s8s32x_inner_product_fwd_t(const pd_t *apd, const input_vector &inputs,
const output_vector &outputs)
- : cpu_primitive_t(&conf_, inputs, outputs), conf_(*pd), dst_is_acc_(false),
- scratchpad_(nullptr)
- {
- dst_is_acc_ = utils::one_of(dst_type, data_type::s32, data_type::f32);
- if (!dst_is_acc_) {
- size_t size = conf_.MB() * conf_.OC() * sizeof(acc_data_t);
- scratchpad_ = create_scratchpad(size);
- }
- }
- ~gemm_u8s8s32x_inner_product_fwd_t() { delete scratchpad_; };
+ : cpu_primitive_t(apd, inputs, outputs, true)
+ { pp_kernel_ = new pp_kernel_t(apd, pd()->dst_is_acc_); }
+ ~gemm_x8s8s32x_inner_product_fwd_t() { delete pp_kernel_; }
typedef typename prec_traits<dst_type>::type data_t;
- typedef typename prec_traits<data_type::u8>::type src_data_t;
+ typedef typename prec_traits<src_type>::type src_data_t;
typedef typename prec_traits<data_type::s8>::type wei_data_t;
typedef typename prec_traits<dst_type>::type dst_data_t;
typedef typename prec_traits<data_type::s32>::type acc_data_t;
- virtual void execute(event_t *e) {
+ virtual void execute(event_t *e) const {
execute_forward();
e->set_state(event_t::ready);
}
private:
- void execute_forward();
- pd_t conf_;
- bool dst_is_acc_;
- scratchpad_t *scratchpad_;
+ // XXX: this is throwaway code that will become unnecessary when we have a
+ // sufficiently advanced igemm jit generator that supports quantization,
+ // relu, and whatnot
+ class pp_kernel_t: jit_generator {
+ public:
+ DECLARE_CPU_JIT_AUX_FUNCTIONS(
+ gemm_x8s8s32x_inner_product_fwd_t::pp_kernel);
+ pp_kernel_t(const pd_t *pd, bool dst_is_acc);
+
+ void operator()(dst_data_t *dst, const acc_data_t *acc,
+ const char *bias, const float *scales, float nslope,
+ size_t start, size_t end);
+ private:
+ void generate();
+
+ struct ker_args {
+ dst_data_t *dst;
+ const acc_data_t *acc;
+ const char *bias;
+ const float *scales;
+ float nslope;
+ size_t len;
+ size_t oc_offset;
+ };
+ void (*ker_)(const ker_args *args);
+
+ size_t OC_;
+ data_type_t bias_data_type_;
+ size_t bias_data_type_size_;
+ size_t scale_idx_mult_;
+ round_mode_t rmode_;
+ bool do_bias_;
+ bool do_relu_;
+ };
+
+ void execute_forward() const;
+ const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
+
+ pp_kernel_t *pp_kernel_;
};
}
}
projects / platform / upstream / dldt.git / blobdiff