add simple simd lane set partitioning

1 files changed, 18 insertions, 0 deletions

diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index bb2ac33..01dd92c 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -1572,6 +1572,18 @@ class shader_core_config : public core_config {
     max_warps_per_shader = n_thread_per_shader / warp_size;
     assert(!(n_thread_per_shader % warp_size));
 
+    // Validate and compute SIMD partitioning derived values
+    if (gpgpu_simd_partitioning) {
+      assert(gpgpu_num_simd_sets > 0 && gpgpu_num_simd_sets <= warp_size);
+      assert(warp_size % gpgpu_num_simd_sets == 0);
+      simd_set_width = warp_size / gpgpu_num_simd_sets;
+      printf("SIMD Partitioning: %u sets of %u lanes\n", gpgpu_num_simd_sets,
+             simd_set_width);
+    } else {
+      gpgpu_num_simd_sets = 1;
+      simd_set_width = warp_size;
+    }
+
     set_pipeline_latency();
 
     m_L1I_config.init(m_L1I_config.m_config_string, FuncCachePreferNone);
@@ -1714,6 +1726,11 @@ class shader_core_config : public core_config {
   unsigned max_dp_latency;
   unsigned max_tensor_core_latency;
 
+  // SIMD lane partitioning
+  bool gpgpu_simd_partitioning;
+  unsigned gpgpu_num_simd_sets;
+  unsigned simd_set_width;  // derived: warp_size / num_simd_sets
+
   unsigned n_simt_cores_per_cluster;
   unsigned n_simt_clusters;
   unsigned n_simt_ejection_buffer_size;
@@ -2630,6 +2647,7 @@ class exec_shader_core_ctx : public shader_core_ctx {
   virtual void checkExecutionStatusAndUpdate(warp_inst_t &inst, unsigned t,
                                              unsigned tid);
   virtual void func_exec_inst(warp_inst_t &inst);
+  void execute_warp_inst_per_set(warp_inst_t &inst);
   virtual unsigned sim_init_thread(kernel_info_t &kernel,
                                    ptx_thread_info **thread_info, int sid,
                                    unsigned tid, unsigned threads_left,