-Bug 146 fix (Adding perfect memory interface)

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 11877]

2 files changed, 40 insertions, 7 deletions

diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc
index 5779537..7bf711f 100644
--- a/src/gpgpu-sim/shader.cc
+++ b/src/gpgpu-sim/shader.cc
@@ -100,7 +100,12 @@ shader_core_ctx::shader_core_ctx( class gpgpu_sim *gpu,
       m_threadState[i].m_active = false;
    }
    
-   m_icnt = new shader_memory_interface(this,cluster);
+   // m_icnt = new shader_memory_interface(this,cluster);
+    if ( m_config->gpgpu_perfect_mem ) {
+        m_icnt = new perfect_memory_interface(this,cluster);
+    } else {
+        m_icnt = new shader_memory_interface(this,cluster);
+    }
    m_mem_fetch_allocator = new shader_core_mem_fetch_allocator(shader_id,tpc_id,mem_config);
 
    // fetch
@@ -924,7 +929,7 @@ pipelined_simd_unit::pipelined_simd_unit( warp_inst_t **result_port, const shade
 	m_pipeline_reg[i] = new warp_inst_t( config );
 }
 
-ldst_unit::ldst_unit( shader_memory_interface *icnt, 
+ldst_unit::ldst_unit( mem_fetch_interface *icnt,
                       shader_core_mem_fetch_allocator *mf_allocator,
                       shader_core_ctx *core, 
                       opndcoll_rfu_t *operand_collector,
@@ -1070,7 +1075,7 @@ void ldst_unit::cycle()
            m_L1C->fill(mf,gpu_sim_cycle+gpu_tot_sim_cycle);
            m_response_fifo.pop_front(); 
        } else {
-           if( mf->get_type() == WRITE_ACK ) {
+    	   if( mf->get_type() == WRITE_ACK || ( m_config->gpgpu_perfect_mem && mf->get_is_write() )) {
                m_core->store_ack(mf);
                m_response_fifo.pop_front();
                delete mf;
@@ -1768,7 +1773,7 @@ void shader_core_ctx::accept_ldst_unit_response(mem_fetch * mf)
 
 void shader_core_ctx::store_ack( class mem_fetch *mf )
 {
-    assert( mf->get_type() == WRITE_ACK );
+	assert( mf->get_type() == WRITE_ACK  || ( m_config->gpgpu_perfect_mem && mf->get_is_write() ) );
     unsigned warp_id = mf->get_wid();
     m_warp[warp_id].dec_store_req();
 }
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index b0a9c3a..1d481dd 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -823,7 +823,7 @@ class cache_t;
 
 class ldst_unit: public pipelined_simd_unit {
 public:
-    ldst_unit( shader_memory_interface *icnt, 
+    ldst_unit( mem_fetch_interface *icnt,
                shader_core_mem_fetch_allocator *mf_allocator,
                shader_core_ctx *core, 
                opndcoll_rfu_t *operand_collector,
@@ -868,7 +868,7 @@ private:
    mem_stage_stall_type process_memory_access_queue( cache_t *cache, warp_inst_t &inst );
 
    const memory_config *m_memory_config;
-   shader_memory_interface *m_icnt;
+   class mem_fetch_interface *m_icnt;
    shader_core_mem_fetch_allocator *m_mf_allocator;
    class shader_core_ctx *m_core;
    unsigned m_sid;
@@ -1213,7 +1213,7 @@ private:
     thread_ctx_t             *m_threadState;
     
     // interconnect interface
-    shader_memory_interface *m_icnt;
+    mem_fetch_interface *m_icnt;
     shader_core_mem_fetch_allocator *m_mem_fetch_allocator;
     
     // fetch
@@ -1263,6 +1263,14 @@ public:
     bool icnt_injection_buffer_full(unsigned size, bool write);
     void icnt_inject_request_packet(class mem_fetch *mf);
 
+    // for perfect memory interface
+    bool response_queue_full() {
+        return ( m_response_fifo.size() >= m_config->n_simt_ejection_buffer_size );
+    }
+    void push_response_fifo(class mem_fetch *mf) {
+        m_response_fifo.push_back(mf);
+    }
+
     void get_pdom_stack_top_info( unsigned sid, unsigned tid, unsigned *pc, unsigned *rpc ) const;
     unsigned max_cta( const kernel_info_t &kernel );
     unsigned get_not_completed() const;
@@ -1302,4 +1310,24 @@ private:
     simt_core_cluster *m_cluster;
 };
 
+class perfect_memory_interface : public mem_fetch_interface {
+public:
+    perfect_memory_interface( shader_core_ctx *core, simt_core_cluster *cluster ) { m_core=core; m_cluster=cluster; }
+    virtual bool full( unsigned size, bool write) const
+    {
+        return m_cluster->response_queue_full();
+    }
+    virtual void push(mem_fetch *mf)
+    {
+        if( !mf->get_inst().empty() )
+            m_core->mem_instruction_stats(mf->get_inst()); // not I$-fetch
+        if ( mf && mf->isatomic() )
+            mf->do_atomic(); // execute atomic inside the "memory subsystem"
+        m_cluster->push_response_fifo(mf);
+    }
+private:
+    shader_core_ctx *m_core;
+    simt_core_cluster *m_cluster;
+};
+
 #endif /* SHADER_H */