sync ipex 595b2e1 (#310)

Author: DDEle

CMakeLists.txt

@@ -46,7 +46,7 @@ if (${LOG} STREQUAL "on")
 endif ()
 
 # For large registers mode, enable 256 registers for kernels
-# set(XETLA_OFFLINE_OPTIONS "-doubleGRF")
+set(XETLA_OFFLINE_OPTIONS "-doubleGRF")
 set(XETLA_OFFLINE_OPTIONS "${XETLA_OFFLINE_OPTIONS} -vc-disable-indvars-opt")
 set(XETLA_OFFLINE_OPTIONS "${XETLA_OFFLINE_OPTIONS} -vc-codegen")
 # Enable bank conflict reduction.

include/experimental/group/gemm/compute_policy.hpp

@@ -137,11 +137,11 @@ struct compute_policy_int4_dequantize<
       quant_info_.weight_mem_layout == mem_layout::col_major;
 
   static constexpr uint32_t block_size_y_a = is_col_major_b ? 8 : 16;
-  static constexpr uint32_t block_bytes_x_a = is_col_major_b ? 128 : 32;
+  static constexpr uint32_t block_bytes_x_a = is_col_major_b ? 256 : 32;
   static constexpr uint32_t block_size_x_a =
       block_bytes_x_a / sizeof(dtype_mma_a);
   static constexpr uint32_t block_size_x_b = is_col_major_b ? 1 : 32;
-  static constexpr uint32_t block_bytes_y_b = is_col_major_b ? 128 : 32;
+  static constexpr uint32_t block_bytes_y_b = is_col_major_b ? 256 : 32;
   static constexpr uint32_t block_size_y_b =
       block_bytes_y_b / sizeof(dtype_mma_b);
 

include/subgroup/tile/impl/load_xe.hpp

@@ -524,12 +524,16 @@ tile_load(tile_t& tile, payload_t& payload) {
           }
           reg_sub
               .xetla_select<load_elems * pack_factor, 1>(
-                  sub_block_offset * tile_desc::block_size_x)
+                  sub_block_offset *
+                  (payload_t::mem_transpose ? tile_desc::block_size_y
+                                            : tile_desc::block_size_x))
               .xetla_format<load_dtype>() = reg_tmp_trans;
         } else {
           reg_sub
               .xetla_select<load_elems * pack_factor, 1>(
-                  sub_block_offset * tile_desc::block_size_x)
+                  sub_block_offset *
+                  (payload_t::mem_transpose ? tile_desc::block_size_y
+                                            : tile_desc::block_size_x))
               .xetla_format<load_dtype>() = reg_tmp;
         }
       }

include/subgroup/tile/impl/payload_xe.hpp

@@ -1659,8 +1659,10 @@ struct prefetch_payload_t<
     arch_tag_,
     std::enable_if_t<
         arch_tag_ <= gpu_arch::XeHpg &&
-        ((block_size_y_ != 1 && mem_layout_ == mem_layout::row_major) ||
-         (block_size_x_ != 1 && mem_layout_ == mem_layout::col_major))>> {
+        (((block_size_y_ != 1 || tile_size_y_ != 1) &&
+          mem_layout_ == mem_layout::row_major) ||
+         ((block_size_x_ != 1 || tile_size_x_ != 1) &&
+          mem_layout_ == mem_layout::col_major))>> {
   using dtype = native_type_t<dtype_>;
   using mem_desc_t =
       mem_desc_t<dtype_, mem_layout_, mem_space::global, alignment_>;
@@ -1885,8 +1887,10 @@ struct prefetch_payload_t<
     arch_tag_,
     std::enable_if_t<
         (arch_tag_ == gpu_arch::XeHpc) &&
-        (((block_size_y_ != 1) && mem_layout_ == mem_layout::row_major) ||
-         ((block_size_x_ != 1) && mem_layout_ == mem_layout::col_major))>> {
+        (((block_size_y_ != 1 || tile_size_y_ != 1) &&
+          mem_layout_ == mem_layout::row_major) ||
+         ((block_size_x_ != 1 || tile_size_x_ != 1) &&
+          mem_layout_ == mem_layout::col_major))>> {
   using dtype = dtype_;
   using mem_desc_t =
       mem_desc_t<dtype_, mem_layout_, mem_space::global, alignment_>;
@@ -2176,8 +2180,10 @@ struct prefetch_payload_t<
     num_coop_sg_,
     arch_tag_,
     std::enable_if_t<
-        ((block_size_y_ == 1) && mem_layout_ == mem_layout::row_major) ||
-        ((block_size_x_ == 1) && mem_layout_ == mem_layout::col_major)>> {
+        ((block_size_y_ == 1 && tile_size_y_ == 1) &&
+         mem_layout_ == mem_layout::row_major) ||
+        ((block_size_x_ == 1 && tile_size_x_ == 1) &&
+         mem_layout_ == mem_layout::col_major)>> {
   using dtype = dtype_;
   using mem_desc_t =
       mem_desc_t<dtype_, mem_layout_, mem_space::global, alignment_>;