x64: brgemm bwd_d strided: remove static last_oc_block_size

oneapi-src · Apr 9, 2024 · 0184044 · 0184044
1 parent e4737d9
commit 0184044
Showing 1 changed file with 5 additions and 9 deletions.
diff --git a/src/cpu/x64/jit_brgemm_conv_bwd_utils.cpp b/src/cpu/x64/jit_brgemm_conv_bwd_utils.cpp
@@ -455,7 +455,6 @@ struct brg_blocking_t : public jit_brgemm_conv_conf_t {
     static constexpr int bench_iterations = 1;
 
     int sp, sp_block, nb_sp;
-    static int last_oc_block_size;
 
     void get_from_jcp(const jit_brgemm_conv_conf_t &jcp) { *this = jcp; }
     void save_to_jcp(jit_brgemm_conv_conf_t &jcp) const { jcp = *this; }
@@ -531,7 +530,6 @@ struct brg_blocking_t : public jit_brgemm_conv_conf_t {
 
 unsigned brg_blocking_t::L1;
 unsigned brg_blocking_t::L2;
-int brg_blocking_t::last_oc_block_size;
 
 float brg_blocking_t::io_k(dim_t src, dim_t wei, dim_t dst, float n, float pk,
         bool is_broadcast, bool is_shared) const {
@@ -556,7 +554,7 @@ float brg_blocking_t::io_k(const loop_t loop, const array_in_loop_t arr,
 }
 
 void brg_blocking_t::select_oc_block() {
-    const auto padded_oc = last_oc_block_size * (is_oc_padded ? acc_simd_w : 1);
+    const auto padded_oc = vnni_block * (is_oc_padded ? acc_simd_w : 1);
     oc_block = (exec_type == exec_trans ? rnd_up(oc, padded_oc) : oc);
     nb_oc = utils::div_up(oc, oc_block);
 }
@@ -570,7 +568,7 @@ status_t brg_blocking_t::estimate_brgemm_ur() {
 
     // Configure matrix sizes
     // for amx if oc_block != oc then we use exec_trans so K is oc_block
-    const auto padded_oc = last_oc_block_size * (is_oc_padded ? acc_simd_w : 1);
+    const auto padded_oc = vnni_block * (is_oc_padded ? acc_simd_w : 1);
 
     ocp = rnd_up(oc, padded_oc);
 
@@ -641,7 +639,7 @@ status_t brg_blocking_t::get_brgemm_ur(
                     brg_strides.stride_a = ngroups * oc_without_padding
                             * (dilate_w + 1) * src_dsz;
                     //weights are padded by ic_block and last_oc_block
-                    brg_strides.stride_b = rnd_up(oc, last_oc_block_size)
+                    brg_strides.stride_b = rnd_up(oc, vnni_block)
                             * rnd_up(ic, ic_block) * wei_dsz;
                     const auto strides_ptr = (brg_type == brgemm_strd)
                             ? &brg_strides
@@ -1501,8 +1499,7 @@ status_t init_jcp(jit_brgemm_conv_conf_t &jcp, cpu_isa_t isa,
 
     VDISPATCH_CONV_IC(!jcp.is_bf32, VERBOSE_UNSUPPORTED_DT);
 
-    brg_blocking_t::last_oc_block_size
-            = (jcp.wei_dt == f16 && isa == avx512_core_fp16)
+    jcp.vnni_block = (jcp.wei_dt == f16 && isa == avx512_core_fp16)
             ? 1
             : data_type_vnni_granularity(jcp.wei_dt);
 
@@ -1921,8 +1918,7 @@ status_t init_conf(jit_brgemm_conv_conf_t &jcp, cpu_isa_t isa,
 
     jcp.copy_block_only = true;
 
-    const auto oc_padded_block
-            = jcp.acc_simd_w * brg_blocking_t::last_oc_block_size;
+    const auto oc_padded_block = jcp.acc_simd_w * jcp.vnni_block;
     jcp.is_oc_padded = one_of(jcp.wei_dt, bf16, f16, s8)
             && jcp.oc > oc_padded_block && is_amx(isa);