Publishing 2019 R1 content

[platform/upstream/dldt.git] / inference-engine / thirdparty / mkl-dnn / src / cpu / jit_avx2_conv_kernel_f32.hpp

diff --git a/inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx2_conv_kernel_f32.hpp b/inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx2_conv_kernel_f32.hpp
index f370054..0c4eb31 100644 (file)
--- a/inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx2_conv_kernel_f32.hpp
+++ b/inference-engine/thirdparty/mkl-dnn/src/cpu/jit_avx2_conv_kernel_f32.hpp
@@ -18,9 +18,11 @@
 #define JIT_AVX2_CONV_KERNEL_F32_HPP
 
 #include "c_types_map.hpp"
+#include "memory_tracking.hpp"
+
+#include "cpu_memory.hpp"
 #include "jit_generator.hpp"
 #include "jit_primitive_conf.hpp"
-#include "cpu_memory.hpp"
 #include "jit_uni_eltwise.hpp"
 #include "jit_uni_depthwise.hpp"
 
@@ -29,8 +31,9 @@ namespace impl {
 namespace cpu {
 
 struct jit_avx2_conv_fwd_kernel_f32: public jit_generator {
-    jit_avx2_conv_fwd_kernel_f32(jit_conv_conf_t ajcp,
-            const primitive_attr_t &attr): jcp(ajcp), attr_(attr)
+    jit_avx2_conv_fwd_kernel_f32(jit_conv_conf_t ajcp, jit_conv_conf_t ajcp_dw,
+            const primitive_attr_t &attr)
+        : jcp(ajcp), jcp_dw(ajcp_dw), attr_(attr)
     {
         this->generate();
         jit_ker = (void (*)(jit_conv_call_s *))this->getCode();
@@ -54,11 +57,12 @@ struct jit_avx2_conv_fwd_kernel_f32: public jit_generator {
             const convolution_desc_t &cd, const memory_desc_wrapper &src_d,
             const memory_desc_wrapper &weights_d,
             const memory_desc_wrapper &dst_d,
-            const primitive_attr_t &attr,
-            bool with_relu = false,
-            float relu_negative_slope = 0.);
+            const primitive_attr_t &attr);
+    static void init_scratchpad(memory_tracking::registrar_t &scratchpad,
+            const jit_conv_conf_t &jcp, const jit_conv_conf_t &jcp_dw = jit_conv_conf_t());
 
     jit_conv_conf_t jcp;
+    jit_conv_conf_t jcp_dw;
     const primitive_attr_t &attr_;
     void (*jit_ker)(jit_conv_call_s *);
 
@@ -84,7 +88,7 @@ private:
     reg64_t reg_long_offt = r15;
     Xbyak::Reg32 reg_ci_flag = r13d;
 
-    Xbyak::Ymm ymask = Xbyak::Ymm(14);
+    Xbyak::Ymm ytmp = Xbyak::Ymm(14);
 
     reg64_t reg_d_weights = imm_addr64;
     reg64_t reg_d_bias = ki_iter;
@@ -116,6 +120,8 @@ struct jit_avx2_conv_bwd_data_kernel_f32: public jit_generator {
             const convolution_desc_t &cd, const memory_desc_wrapper &diff_src_d,
             const memory_desc_wrapper &weights_d,
             const memory_desc_wrapper &diff_dst_d);
+    static void init_scratchpad(memory_tracking::registrar_t &scratchpad,
+            const jit_conv_conf_t &jcp);
 
     jit_conv_conf_t jcp;
     void (*jit_ker)(jit_conv_call_s *);
@@ -123,33 +129,52 @@ struct jit_avx2_conv_bwd_data_kernel_f32: public jit_generator {
 private:
     using reg64_t = const Xbyak::Reg64;
 
-    reg64_t reg_input      = rax;
     reg64_t reg_ddst       = rax;
-    reg64_t aux_reg_input  = r8;
     reg64_t aux_reg_ddst   = r8;
-    reg64_t aux1_reg_input = r9;
     reg64_t reg_kernel     = rdx;
     reg64_t aux_reg_kernel = r10;
-    reg64_t reg_output     = rsi;
     reg64_t reg_dsrc       = rsi;
-    reg64_t aux_reg_output = rbx;
-    reg64_t aux_reg_dsrc = rbx;
+    reg64_t aux_reg_ddst_oc_loop  = rbx; // used in ndims < 5 case only
+    reg64_t aux_reg_kernel_oc_loop = abi_not_param1; /* used in ndims < 5
+                                                        case only */
 
-    reg64_t aux_reg_dst_d = r12;
-    reg64_t aux_reg_ker_d = r14;
+    reg64_t aux_reg_dst_d = r12; // used in ndims == 5 case only
+    reg64_t aux_reg_ker_d = r14; // used in ndims == 5 case only
 
-    reg64_t reg_ki  = abi_not_param1;
+    reg64_t reg_ki  = abi_not_param1; // used in ndims == 5 case only
     reg64_t kj      = r11;
     reg64_t oi_iter = r12;
     reg64_t reg_kh  = r14;
-    reg64_t ki_iter = r13;
+    reg64_t reg_channel = r13;  // used in ndims < 5 case only
+    reg64_t reg_channel_work = r9;  // used in ndims < 5 case only
     reg64_t reg_long_offt = r15;
-    reg64_t start_off_reg = aux1_reg_input;
 
-    inline void hsw_iter(int ur_w, int l_overflow, int r_overflow,
-            int start_off);
+    inline void compute_loop(int ur_w, int l_overflow, int r_overflow);
 
     void generate();
+
+    inline int get_iw_start(int ki, int l_overflow)
+    {
+        int res = (jcp.iw - 1 + jcp.r_pad) % jcp.stride_w
+                + l_overflow * jcp.stride_w
+                - (jcp.kw - 1 - ki) * (jcp.dilate_w + 1);
+        while (res < 0)
+            res += jcp.stride_w;
+
+        return res;
+    }
+
+    inline int get_iw_end(int ur_w, int ki, int r_overflow)
+    {
+        if (utils::one_of(ur_w, jcp.iw, jcp.ur_w_tail))
+            ur_w += nstl::min(0, jcp.r_pad); // remove negative padding
+        int res = (ur_w - 1 + jcp.l_pad) % jcp.stride_w
+                + r_overflow * jcp.stride_w - ki * (jcp.dilate_w + 1);
+        while (res < 0)
+            res += jcp.stride_w;
+
+        return ur_w - res;
+    }
 };
 
 struct jit_avx2_conv_bwd_weights_kernel_f32: public jit_generator {
@@ -165,6 +190,8 @@ struct jit_avx2_conv_bwd_weights_kernel_f32: public jit_generator {
             const convolution_desc_t &cd, const memory_desc_wrapper &src_d,
             const memory_desc_wrapper &diff_weights_d,
             const memory_desc_wrapper &diff_dst_d);
+    static void init_scratchpad(memory_tracking::registrar_t &scratchpad,
+            const jit_conv_conf_t &jcp);
 
     jit_conv_conf_t jcp;
     void (*jit_ker)(jit_conv_call_s *);