x64: brgemm bwd_w conv: use ih_block instead of ih for tr_src scratchpad

oneapi-src · Apr 17, 2023 · 8da1083 · 8da1083
1 parent f7acf98
commit 8da1083
Show file tree

Hide file tree

Showing 2 changed files with 17 additions and 14 deletions.
diff --git a/src/cpu/x64/jit_brgemm_conv_bwd_w.cpp b/src/cpu/x64/jit_brgemm_conv_bwd_w.cpp
@@ -134,7 +134,7 @@ status_t brgemm_convolution_bwd_weights_t::pd_t::init(engine_t *engine) {
                 brgattr.max_top_vpad = 0;
                 brgattr.max_bottom_vpad = 0;
 
-                brgattr.LDA2 = jcp_.tr_iw * jcp_.ih * jcp_.id;
+                brgattr.LDA2 = jcp_.tr_iw * jcp_.ih_block * jcp_.id;
                 brgattr.LDB2 = jcp_.tr_ow * jcp_.oc_block * jcp_.oh * jcp_.od;
                 brgattr.LDC2_M = jcp_.oc_block * jcp_.kd * jcp_.kh * jcp_.kw;
                 brgattr.LDC2_N = jcp_.nb_ic * jcp_.ic_block * jcp_.oc_block
@@ -464,7 +464,7 @@ struct brgemm_convolution_bwd_weights_t::thread_info_t {
 
     size_t tr_src_off(int g, int icb, int id, int ih) const {
         const size_t tr_row_size = jcp.tr_iw * jcp.ic_block;
-        const size_t tr_3d_size = tr_row_size * jcp.ih;
+        const size_t tr_3d_size = tr_row_size * jcp.ih_block;
         int adj = (jcp.global_transpose) ? 1 : jcp.nb_ic_blocking;
         // Aligned to buffer end to use guard elements
         return tr_src_buf_number(g, icb) * adj * jcp.tr_src_buf_size
@@ -1024,7 +1024,7 @@ void brgemm_convolution_bwd_weights_t::compute_diff_weights_3d(
                 + _pd->filter_w_to_src(kw) / jcp.stride_w
                 + (kw % jcp.stride_w) * src_stride_w_shift
                 + (bs_ih_s - ih_s) * jcp.tr_iw * jcp.ic_block
-                + (bs_id_s - id_s) * jcp.ih * jcp.tr_iw * jcp.ic_block;
+                + (bs_id_s - id_s) * jcp.ih_block * jcp.tr_iw * jcp.ic_block;
         const void *ptr_B = ((diff_dst_data_t *)p_dst)
                 + (bs_oh_s - oh_s) * jcp.tr_ow * jcp.oc_block
                 + (bs_od_s - od_s) * jcp.oh * jcp.tr_ow * jcp.oc_block;
@@ -1045,7 +1045,7 @@ void brgemm_convolution_bwd_weights_t::compute_diff_weights_3d(
                 ti->brg_batch[odb * bs_h + ohb].ptr.A = (char *)ptr_A
                         + ohb * jcp.typesize_in * jcp.tr_iw * jcp.ic_block
                                 * jcp.stride_h
-                        + odb * jcp.typesize_in * jcp.ih * jcp.tr_iw
+                        + odb * jcp.typesize_in * jcp.ih_block * jcp.tr_iw
                                 * jcp.ic_block * jcp.stride_d;
                 ti->brg_batch[odb * bs_h + ohb].ptr.B = (char *)ptr_B
                         + ohb * jcp.typesize_in * jcp.tr_ow * jcp.oc_block

diff --git a/src/cpu/x64/jit_brgemm_conv_utils.cpp b/src/cpu/x64/jit_brgemm_conv_utils.cpp
@@ -2634,16 +2634,6 @@ void balance_bwd_w(jit_brgemm_conv_conf_t &jcp) {
     jcp.nthr_g = nthr_g;
     jcp.nthr_oc_b = nthr_oc_b;
     jcp.nthr_ic_b = nthr_ic_b;
-
-    // TODO: Optimize memory allocation when threaded on height and depth
-    jcp.tr_src_buf_size = jcp.tr_iw * jcp.ic_block * jcp.ih * jcp.id;
-    jcp.tr_diff_dst_buf_size = jcp.tr_ow * jcp.oc_block * jcp.oh * jcp.od;
-    jcp.tr_src_buf_count = jcp.global_transpose
-            ? jcp.nthr_mb * jcp.nb_ic * jcp.ngroups
-            : jcp.nthr;
-    jcp.tr_diff_dst_buf_count = jcp.global_transpose
-            ? jcp.nthr_mb * jcp.nb_oc * jcp.ngroups
-            : jcp.nthr;
 }
 
 status_t init_conf_bwd_w(jit_brgemm_conv_conf_t &jcp,
@@ -2886,6 +2876,19 @@ status_t init_conf_bwd_w(jit_brgemm_conv_conf_t &jcp,
     // try to split oh by equal oh blocks
     oh_block_limit = div_up(jcp.oh, div_up(jcp.oh, oh_block_limit));
     jcp.oh_block = utils::saturate(1, jcp.oh, oh_block_limit);
+    jcp.ih_block = nstl::min(jcp.ih,
+            jcp.stride_h
+                    * brg_blocking_t::get_inp_size(jcp.ih, jcp.oh_block, jcp.kh,
+                            jcp.stride_h, jcp.dilate_h));
+    // TODO: Optimize memory allocation when threaded on height and depth
+    jcp.tr_src_buf_count = jcp.global_transpose
+            ? jcp.nthr_mb * jcp.nb_ic * jcp.ngroups
+            : jcp.nthr;
+    jcp.tr_diff_dst_buf_count = jcp.global_transpose
+            ? jcp.nthr_mb * jcp.nb_oc * jcp.ngroups
+            : jcp.nthr;
+    jcp.tr_src_buf_size = jcp.tr_iw * jcp.ic_block * jcp.ih_block * jcp.id;
+    jcp.tr_diff_dst_buf_size = jcp.tr_ow * jcp.oc_block * jcp.oh * jcp.od;
 
     const int iframe_size = irow_size * jcp.id;
     const int oframe_size = orow_size * jcp.od;