[FORK][FEATURE] Added JIT AVX512/AVX2 FP32 Planar Convolution implementation

Author: dmitry-gorokhov

src/cpu/cpu_convolution_list.cpp

@@ -55,6 +55,7 @@
 #include "cpu/x64/jit_uni_fork_dw_convolution.hpp"
 #include "cpu/x64/jit_uni_x8s8s32x_1x1_convolution.hpp"
 #include "cpu/x64/jit_uni_x8s8s32x_convolution.hpp"
+#include "cpu/x64/jit_uni_planar_convolution.hpp"
 using namespace dnnl::impl::cpu::x64;
 #elif DNNL_AARCH64
 #include "cpu/aarch64/jit_brdgmm_dw_conv.hpp"
@@ -124,12 +125,14 @@ const std::map<pk_dt_impl_key_t, std::vector<impl_list_item_t>> &impl_list_map()
             CPU_INSTANCE_X64(ip_convolution_fwd_t)
             CPU_INSTANCE_AMX(brgemm_1x1_convolution_fwd_t<avx512_core_amx>)
             CPU_INSTANCE_AMX(brgemm_convolution_fwd_t<avx512_core_amx>)
+            CPU_INSTANCE_AVX512(jit_avx512_common_planar_convolution_fwd_t)
             CPU_INSTANCE_AVX512(brgemm_1x1_convolution_fwd_t<avx512_core>)
             CPU_INSTANCE_AVX512(brgemm_convolution_fwd_t<avx512_core>)
             CPU_INSTANCE_AVX512(jit_avx512_common_dw_convolution_fwd_t)
             CPU_INSTANCE_AVX512(jit_avx512_common_fork_dw_convolution_fwd_t)
             CPU_INSTANCE_AVX512(jit_avx512_common_1x1_convolution_fwd_f32_t)
             CPU_INSTANCE_AVX512(jit_avx512_common_convolution_fwd_t<f32>)
+            CPU_INSTANCE_AVX2(jit_avx2_planar_convolution_fwd_t)
             CPU_INSTANCE_AVX2(jit_avx2_dw_convolution_fwd_t)
             CPU_INSTANCE_AVX2(brgemm_1x1_convolution_fwd_t<avx2>)
             CPU_INSTANCE_AVX2(brgemm_convolution_fwd_t<avx2>)

src/cpu/x64/jit_primitive_conf.hpp

@@ -260,6 +260,10 @@ struct jit_conv_conf_t {
 
     bool with_input_zp;
     bool with_weights_zp;
+
+    int oh_block;
+    int oh_block_step;
+    int nb_ow_blocking;
 };
 
 // calculates filter size taking into account dilation
@@ -355,6 +359,7 @@ struct jit_conv_call_s {
     size_t oc_off;
     size_t ic_off;
     size_t oc_off_prf;
+    size_t oh_blocks;
 
     const void *input_zp;
 };