refine heuristic mode inside

ROCm · Mar 1, 2022 · 090d5c1 · 090d5c1
1 parent 3301314
commit 090d5c1
Showing 1 changed file with 4 additions and 34 deletions.
diff --git a/driver/igemm_bwd_gtc_driver.h b/driver/igemm_bwd_gtc_driver.h
@@ -534,6 +534,9 @@ class igemm_bwd_gtc_t : public igemm_driver_base_t{
             return result;
         }
 
+        if(this->driver_mode == driver_mode_heuristic)
+            current_gks = tunable->gemm_k_global_split ? current_gks : 0;
+
         int hi = arg->get_int("in_h");
         int wi = arg->get_int("in_w");
         int n = arg->get_int("batchsize");
@@ -840,7 +843,7 @@ class igemm_bwd_gtc_t : public igemm_driver_base_t{
         result_t result;
         result.kernel_name = kernel_name;
 
-        if(this->driver_mode == driver_mode_normal){
+        if(true){
             float min_duration = FLT_MAX;
             float duration = .0;
             int selected_gks = 0;
@@ -955,39 +958,6 @@ class igemm_bwd_gtc_t : public igemm_driver_base_t{
             result.return_code = 0;
             result.duration_ms = min_duration;
             result.gks         = selected_gks;
-        }else if(this->driver_mode == driver_mode_heuristic){
-            int gks   = tunable->gemm_k_global_split ? current_gks : 0;  // sync with is_tunable_predicted
-            size_t grid_size = get_grid_size(arg, tunable) * (1 << gks);
-            if(tunable->multihead){
-                if(tunable->tensor_layout == "nhwc"){
-                    int gemm_m = n * h_tilda_slice * w_tilda_slice;
-                    int gemm_n = c / group;
-                    // This is hacky, but in MIOpen we prefer a heuristic way to set gks, so ok now.
-                    igemm_bwd_gtc_nhwc_karg_t *karg = (igemm_bwd_gtc_nhwc_karg_t *)(karg_buffer);
-                    magic_div_u32_t mdiv_x_tilda = magic_div_u32_gen(x_tilda);
-                    magic_div_u32_t mdiv_y_tilda = magic_div_u32_gen(y_tilda);
-                    magic_div_u32_t mdiv_group_mn = magic_div_u32_gen(group * utility_integer_divide_ceil(gemm_n, gemm_n_per_block) * utility_integer_divide_ceil(gemm_m, gemm_m_per_block));
-                    karg->dtile_iy = num_of_gemm > 1 ? mdiv_x_tilda.magic : 0;
-                    karg->dtile_ix = num_of_gemm > 1 ? mdiv_x_tilda.shift : 0;
-                    karg->dslice_y = num_of_gemm > 1 ? mdiv_y_tilda.magic : y;
-                    karg->dslice_x = num_of_gemm > 1 ? mdiv_y_tilda.shift : x;
-                    karg->dtile_h  = num_of_gemm > 1 ? mdiv_group_mn.magic : h_tilda;
-                    karg->dtile_w  = num_of_gemm > 1 ? mdiv_group_mn.shift : w_tilda;
-                    karg->ks       = gks;
-                }else{
-                    assert(0);
-                }
-
-                float duration = igemm_launch_kernels({
-                        {kernel_func, karg_buffer, karg_size, {grid_size * block_size, splits, 1}, {block_size, 1, 1}}
-                    }, bwd_prolog, bwd_postlog, this->warmup, this->repeat);
-
-                result.return_code = 0;
-                result.duration_ms = duration;
-                result.gks         = gks;
-            }else{
-                assert(0);  // to be supported
-            }
         }
 
 #ifdef IGEMM_SPLIT_KERNEL