Merge pull request #21 from abdallm/dev-purdue-integration

A few little things and some sizeable changes: Added memory partition indexing, some configuration file updates, fixed the texture cache so apps that use tex memory will no longer crash.

13 files changed, 520 insertions, 123 deletions

diff --git a/src/gpgpu-sim/addrdec.cc b/src/gpgpu-sim/addrdec.cc
index 422576d..8651869 100644
--- a/src/gpgpu-sim/addrdec.cc
+++ b/src/gpgpu-sim/addrdec.cc
@@ -62,6 +62,9 @@ void linear_to_raw_address_translation::addrdec_setoption(option_parser_t opp)
    option_parser_register(opp, "-gpgpu_mem_address_mask", OPT_INT32, &gpgpu_mem_address_mask, 
                "0 = old addressing mask, 1 = new addressing mask, 2 = new add. mask + flipped bank sel and chip sel bits",
                "0");
+   option_parser_register(opp, "-memory_partition_indexing", OPT_UINT32, &memory_partition_indexing,
+                  "0 = no indexing, 1 = bitwise xoring, 2 = IPoly, 3 = custom indexing",
+                  "0");
 }
 
 new_addr_type linear_to_raw_address_translation::partition_address( new_addr_type addr ) const 
@@ -103,6 +106,74 @@ void linear_to_raw_address_translation::addrdec_tlx(new_addr_type addr, addrdec_
       tlx->burst= addrdec_packbits(addrdec_mask[BURST], rest_of_addr, addrdec_mkhigh[BURST], addrdec_mklow[BURST]);
    }
 
+    switch(memory_partition_indexing){
+		case CONSECUTIVE:
+			//Do nothing
+			break;
+		case BITWISE_PERMUTATION:
+		{
+			assert(!gap);
+			tlx->chip = (tlx->chip) ^ (tlx->row & (m_n_channel-1));
+			assert(tlx->chip < m_n_channel);
+			break;
+		}
+		case IPOLY:
+		{
+		    /*
+			* Set Indexing function from "Pseudo-randomly interleaved memory."
+			* Rau, B. R et al.
+			* ISCA 1991
+			*
+			* equations are adopted from:
+			* "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu cache management scheme."
+			* Khairy et al.
+			* IEEE TPDS 2017.
+			*/
+			if(m_n_channel == 32) {
+				std::bitset<64> a(tlx->row);
+				std::bitset<5> chip(tlx->chip);
+				chip[0] = a[13]^a[12]^a[11]^a[10]^a[9]^a[6]^a[5]^a[3]^a[0]^chip[0];
+				chip[1] = a[14]^a[13]^a[12]^a[11]^a[10]^a[7]^a[6]^a[4]^a[1]^chip[1];
+				chip[2] = a[14]^a[10]^a[9]^a[8]^a[7]^a[6]^a[3]^a[2]^a[0]^chip[2];
+				chip[3] = a[11]^a[10]^a[9]^a[8]^a[7]^a[4]^a[3]^a[1]^chip[3];
+				chip[4] = a[12]^a[11]^a[10]^a[9]^a[8]^a[5]^a[4]^a[2]^chip[4];
+				tlx->chip = chip.to_ulong();
+				
+			}
+			else{ /* Else incorrect number of channels for the hashing function */
+            assert("\nGPGPU-Sim memory_partition_indexing error: The number of channels should be "
+                    "32 for the hashing IPOLY index function.\n" && 0);
+			}
+			assert(tlx->chip < m_n_channel);
+			break;
+		}
+		case PAE:
+		{
+			//Page Address Entropy
+			//random selected bits from the page and bank bits
+			//similar to
+			//Liu, Yuxi, et al. "Get Out of the Valley: Power-Efficient Address Mapping for GPUs." ISCA 2018
+			std::bitset<64> a(tlx->row);
+			std::bitset<5> chip(tlx->chip);
+			std::bitset<4> b(tlx->bk);
+			chip[0] = a[13]^a[10]^a[9]^a[5]^a[0]^b[3]^b[0]^chip[0];
+			chip[1] = a[12]^a[11]^a[6]^a[1]^b[3]^b[2]^b[1]^chip[1];
+			chip[2] = a[14]^a[9]^a[8]^a[7]^a[2]^b[1]^chip[2];
+			chip[3] = a[11]^a[10]^a[8]^a[3]^b[2]^b[3]^chip[3];
+			chip[4] = a[12]^a[9]^a[8]^a[5]^a[4]^b[1]^b[0]^chip[4];
+			tlx->chip = chip.to_ulong();
+			assert(tlx->chip < m_n_channel);
+			break;
+		}
+		case CUSTOM:
+			/* No custom set function implemented */
+			//Do you custom index here
+			break;
+		default:
+			 assert("\nUndefined set index function.\n" && 0);
+			 break;
+	}
+
    // combine the chip address and the lower bits of DRAM bank address to form the subpartition ID
    unsigned sub_partition_addr_mask = m_n_sub_partition_in_channel - 1; 
    tlx->sub_partition = tlx->chip * m_n_sub_partition_in_channel
diff --git a/src/gpgpu-sim/addrdec.h b/src/gpgpu-sim/addrdec.h
index fd9af8d..bdc5fec 100644
--- a/src/gpgpu-sim/addrdec.h
+++ b/src/gpgpu-sim/addrdec.h
@@ -35,6 +35,14 @@
 
 #include "../abstract_hardware_model.h"
 
+enum partition_index_function{
+	CONSECUTIVE = 0,
+	BITWISE_PERMUTATION,
+	IPOLY,
+	PAE,
+    CUSTOM
+};
+
 struct addrdec_t {
    void print( FILE *fp ) const;
     
@@ -72,6 +80,7 @@ private:
 
    const char *addrdec_option;
    int gpgpu_mem_address_mask;
+   partition_index_function memory_partition_indexing;
    bool run_test; 
 
    int ADDR_CHIP_S;
diff --git a/src/gpgpu-sim/dram.cc b/src/gpgpu-sim/dram.cc
index a57508c..6c11b43 100644
--- a/src/gpgpu-sim/dram.cc
+++ b/src/gpgpu-sim/dram.cc
@@ -199,15 +199,28 @@ dram_req_t::dram_req_t( class mem_fetch *mf, unsigned banks, unsigned dram_bnk_i
 
    const addrdec_t &tlx = mf->get_tlx_addr();
 
-   if(dram_bnk_indexing_policy == 0) {
-	   bk = tlx.bk;
-   }
-    else if(dram_bnk_indexing_policy == 1) {
-    	 int lbank = log2(banks);
-    	 bk  = tlx.bk ^ (((1<<lbank)-1) & tlx.row);
-    }
-    else
-    	assert(1);
+    switch(dram_bnk_indexing_policy){
+		case LINEAR_BK_INDEX:
+		{
+			bk = tlx.bk;
+			break;
+		}
+		case BITWISE_XORING_BK_INDEX:
+		{
+			//xoring bank bits with lower bits of the page
+			int lbank = log2(banks);
+			bk  = tlx.bk ^ (tlx.row & ((1<<lbank)-1));
+			break;
+		}
+		case CUSTOM_BK_INDEX:
+	        /* No custom set function implemented */
+			//Do you custom index here
+			break;
+		default:
+			 assert("\nUndefined bank index function.\n" && 0);
+			 break;
+	}
+
 
    row = tlx.row; 
    col = tlx.col; 
@@ -730,13 +743,15 @@ void dram_t::print( FILE* simFile) const
    printf("\nwrite_to_read_ratio_blp_rw_average = %.6f", write_to_read_ratio_blp_rw_average /banks_access_rw_total);
    printf("\nGrpLevelPara = %.6f \n", (float)bkgrp_parallsim_rw /banks_access_rw_total);
 
-   printf("\nbwutil = %.6f \n", (float)bwutil/n_cmd);
+   printf("\nBW Util details:\n");
+   printf("bwutil = %.6f \n", (float)bwutil/n_cmd);
    printf("total_CMD = %d \n", n_cmd);
    printf("util_bw = %d \n", util_bw);
    printf("Wasted_Col = %d \n", wasted_bw_col);
-   printf("Wasted_Row %d \n", wasted_bw_row);
-   printf("Idle = %d \n\n", idle_bw);
+   printf("Wasted_Row = %d \n", wasted_bw_row);
+   printf("Idle = %d \n", idle_bw);
 
+   printf("\nBW Util Bottlenecks: \n");
    printf("RCDc_limit = %d \n", RCDc_limit);
    printf("RCDWRc_limit = %d \n", RCDWRc_limit);
    printf("WTRc_limit = %d \n", WTRc_limit);
@@ -747,6 +762,7 @@ void dram_t::print( FILE* simFile) const
    printf("WTRc_limit_alone = %d \n", WTRc_limit_alone);
    printf("RTWc_limit_alone = %d \n", RTWc_limit_alone);
 
+   printf("\nCommands details: \n");
    printf("total_CMD = %d \n", n_cmd);
    printf("n_nop = %d \n", n_nop);
    printf("Read = %d \n", n_rd);
@@ -757,8 +773,9 @@ void dram_t::print( FILE* simFile) const
    printf("n_pre = %d \n", n_pre);
    printf("n_ref = %d \n", n_ref);
    printf("n_req = %d \n", n_req );
-   printf("total_req = %d \n\n", n_rd+n_wr+n_rd_L2_A+n_wr_WB);
+   printf("total_req = %d \n", n_rd+n_wr+n_rd_L2_A+n_wr_WB);
 
+   printf("\nDual Bus Interface Util: \n");
    printf("issued_total_row = %lu \n", issued_total_row);
    printf("issued_total_col = %lu \n", issued_total_col);
    printf("Row_Bus_Util =  %.6f \n", (float)issued_total_row / n_cmd);
@@ -879,10 +896,10 @@ void dram_t::set_dram_power_stats(	unsigned &cmd,
 
 unsigned dram_t::get_bankgrp_number(unsigned i)
 {
-	if(m_config->dram_bnkgrp_indexing_policy == 0) { //higher bits
+	if(m_config->dram_bnkgrp_indexing_policy == HIGHER_BITS) { //higher bits
 		return i>>m_config->bk_tag_length;
 	}
-	else if (m_config->dram_bnkgrp_indexing_policy == 1) { //lower bits
+	else if (m_config->dram_bnkgrp_indexing_policy == LOWER_BITS) { //lower bits
 		return i&((m_config->nbkgrp-1));
 	}
 	else {
diff --git a/src/gpgpu-sim/dram.h b/src/gpgpu-sim/dram.h
index 0d4c0e7..965936b 100644
--- a/src/gpgpu-sim/dram.h
+++ b/src/gpgpu-sim/dram.h
@@ -93,6 +93,17 @@ struct bank_t
    unsigned int bkgrpindex;
 };
 
+enum bank_index_function{
+	LINEAR_BK_INDEX = 0,
+	BITWISE_XORING_BK_INDEX,
+    CUSTOM_BK_INDEX
+};
+
+enum bank_grp_bits_position{
+	HIGHER_BITS = 0,
+	LOWER_BITS
+};
+
 struct mem_fetch;
 
 class dram_t 
diff --git a/src/gpgpu-sim/dram_sched.cc b/src/gpgpu-sim/dram_sched.cc
index f754d36..ff50050 100644
--- a/src/gpgpu-sim/dram_sched.cc
+++ b/src/gpgpu-sim/dram_sched.cc
@@ -109,12 +109,14 @@ dram_req_t *frfcfs_scheduler::schedule( unsigned bank, unsigned curr_row )
    if(m_config->seperate_write_queue_enabled) {
 	   if(m_mode == READ_MODE &&
 			  ((m_num_write_pending >= m_config->write_high_watermark )
-			   || (m_queue[bank].empty() && !m_write_queue[bank].empty()))) {
+			  // || (m_queue[bank].empty() && !m_write_queue[bank].empty())
+			   )) {
 		   m_mode = WRITE_MODE;
 	   }
 	   else if(m_mode == WRITE_MODE &&
 				  (( m_num_write_pending < m_config->write_low_watermark )
-				   || (!m_queue[bank].empty() && m_write_queue[bank].empty()))){
+				 //  || (!m_queue[bank].empty() && m_write_queue[bank].empty())
+				   )){
 		   m_mode = READ_MODE;
 	   }
    }
diff --git a/src/gpgpu-sim/gpu-cache.cc b/src/gpgpu-sim/gpu-cache.cc
index 26a1638..a11853a 100644
--- a/src/gpgpu-sim/gpu-cache.cc
+++ b/src/gpgpu-sim/gpu-cache.cc
@@ -71,10 +71,11 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{
 
     switch(m_set_index_function){
     case FERMI_HASH_SET_FUNCTION:
+    case BITWISE_XORING_FUNCTION:
         /*
         * Set Indexing function from "A Detailed GPU Cache Model Based on Reuse Distance Theory"
         * Cedric Nugteren et al.
-        * ISCA 2014
+        * HPCA 2014
         */
         if(m_nset == 32 || m_nset == 64){
             // Lower xor value is bits 7-11
@@ -97,6 +98,36 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{
         }
         break;
 
+    case HASH_IPOLY_FUNCTION:
+    	/*
+		* Set Indexing function from "Pseudo-randomly interleaved memory."
+		* Rau, B. R et al.
+		* ISCA 1991
+		*
+		* "Sacat: streaming-aware conflict-avoiding thrashing-resistant gpgpu cache management scheme."
+		* Khairy et al.
+		* IEEE TPDS 2017.
+    	*/
+    	if(m_nset == 32 || m_nset == 64){
+		std::bitset<64> a(addr);
+		std::bitset<6> index;
+		index[0] = a[25]^a[24]^a[23]^a[22]^a[21]^a[18]^a[17]^a[15]^a[12]^a[7]; //10
+		index[1] = a[26]^a[25]^a[24]^a[23]^a[22]^a[19]^a[18]^a[16]^a[13]^a[8]; //10
+		index[2] = a[26]^a[22]^a[21]^a[20]^a[19]^a[18]^a[15]^a[14]^a[12]^a[9]; //10
+		index[3] = a[23]^a[22]^a[21]^a[20]^a[19]^a[16]^a[15]^a[13]^a[10]; //9
+		index[4] = a[24]^a[23]^a[22]^a[21]^a[20]^a[17]^a[16]^a[14]^a[11]; //9
+
+		 if(m_nset == 64)
+			 index[5] = a[12];
+
+		set_index = index.to_ulong();
+
+    	}else{ /* Else incorrect number of sets for the hashing function */
+    	            assert("\nGPGPU-Sim cache configuration error: The number of sets should be "
+    	                    "32 or 64 for the hashing set index function.\n" && 0);
+    	 }
+        break;
+
     case CUSTOM_SET_FUNCTION:
         /* No custom set function implemented */
         break;
@@ -104,6 +135,10 @@ unsigned l1d_cache_config::set_index(new_addr_type addr) const{
     case LINEAR_SET_FUNCTION:
         set_index = (addr >> m_line_sz_log2) & (m_nset-1);
         break;
+
+    default:
+    	 assert("\nUndefined set index function.\n" && 0);
+    	 break;
     }
 
     // Linear function selected or custom set index function not implemented
@@ -348,6 +383,8 @@ void tag_array::fill( unsigned index, unsigned time, mem_fetch* mf)
     m_lines[index]->fill(time, mf->get_access_sector_mask());
 }
 
+
+//TODO: we need write back the flushed data to the upper level
 void tag_array::flush() 
 {
     for (unsigned i=0; i < m_config.get_num_lines(); i++)
@@ -357,6 +394,13 @@ void tag_array::flush()
     	}
 }
 
+void tag_array::invalidate()
+{
+    for (unsigned i=0; i < m_config.get_num_lines(); i++)
+    	for(unsigned j=0; j < SECTOR_CHUNCK_SIZE; j++)
+    		m_lines[i]->set_status(INVALID, mem_access_sector_mask_t().set(j)) ;
+}
+
 float tag_array::windowed_miss_rate( ) const
 {
     unsigned n_access    = m_access - m_prev_snapshot_access;
@@ -873,8 +917,8 @@ void baseline_cache::cycle(){
 void baseline_cache::fill(mem_fetch *mf, unsigned time){
 
 	if(m_config.m_mshr_type == SECTOR_ASSOC) {
-	assert(mf->original_mf);
-	extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->original_mf);
+	assert(mf->get_original_mf());
+	extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->get_original_mf());
     assert( e != m_extra_mf_fields.end() );
     e->second.pending_read--;
 
@@ -884,7 +928,7 @@ void baseline_cache::fill(mem_fetch *mf, unsigned time){
     	return;
       } else {
     	mem_fetch *temp = mf;
-    	mf = mf->original_mf;
+    	mf = mf->get_original_mf();
     	delete temp;
       }
 	}
@@ -1128,52 +1172,7 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr,
                         unsigned time, std::list<cache_event> &events,
                         enum cache_request_status status )
 {
-
-	new_addr_type block_addr = m_config.block_addr(addr);
-	    new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr());
-
-
-			//if the request writes to the whole cache line/sector, then, write and set cache line Modified.
-			//and no need to send read request to memory or reserve mshr
-
-			if(miss_queue_full(0)) {
-				m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL);
-				return RESERVATION_FAIL; // cannot handle request this cycle
-			}
-
-			bool wb = false;
-			evicted_block_info evicted;
-
-			cache_request_status m_status =  m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf);
-			assert(m_status != HIT);
-			cache_block_t* block = m_tag_array->get_block(cache_index);
-			block->set_status(MODIFIED, mf->get_access_sector_mask());
-			if(m_status == HIT_RESERVED) {
-				block->set_ignore_on_fill(true, mf->get_access_sector_mask());
-				block->set_modified_on_fill(true, mf->get_access_sector_mask());
-			}
-
-			if(mf->get_access_byte_mask().count() == m_config.get_atom_sz())
-			{
-				block->set_m_readable(true, mf->get_access_sector_mask());
-			} else
-			{
-				block->set_m_readable(false, mf->get_access_sector_mask());
-			}
-
-			if( m_status != RESERVATION_FAIL ){
-				   // If evicted block is modified and not a write-through
-				   // (already modified lower level)
-				   if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) {
-					   mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr,
-						   m_wrbk_type,evicted.m_modified_size,true);
-					   send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events);
-				   }
-				   return MISS;
-			   }
-			return RESERVATION_FAIL;
-
-	/*new_addr_type block_addr = m_config.block_addr(addr);
+    new_addr_type block_addr = m_config.block_addr(addr);
     new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr());
 
 	if(mf->get_access_byte_mask().count() == m_config.get_atom_sz())
@@ -1282,7 +1281,59 @@ data_cache::wr_miss_wa_fetch_on_write( new_addr_type addr,
 				return MISS;
 			}
 	   return RESERVATION_FAIL;
-	}*/
+	}
+}
+
+enum cache_request_status
+data_cache::wr_miss_wa_lazy_fetch_on_read( new_addr_type addr,
+                        unsigned cache_index, mem_fetch *mf,
+                        unsigned time, std::list<cache_event> &events,
+                        enum cache_request_status status )
+{
+
+	    new_addr_type block_addr = m_config.block_addr(addr);
+	    new_addr_type mshr_addr = m_config.mshr_addr(mf->get_addr());
+
+
+		//if the request writes to the whole cache line/sector, then, write and set cache line Modified.
+		//and no need to send read request to memory or reserve mshr
+
+		if(miss_queue_full(0)) {
+			m_stats.inc_fail_stats(mf->get_access_type(), MISS_QUEUE_FULL);
+			return RESERVATION_FAIL; // cannot handle request this cycle
+		}
+
+		bool wb = false;
+		evicted_block_info evicted;
+
+		cache_request_status m_status =  m_tag_array->access(block_addr,time,cache_index,wb,evicted,mf);
+		assert(m_status != HIT);
+		cache_block_t* block = m_tag_array->get_block(cache_index);
+		block->set_status(MODIFIED, mf->get_access_sector_mask());
+		if(m_status == HIT_RESERVED) {
+			block->set_ignore_on_fill(true, mf->get_access_sector_mask());
+			block->set_modified_on_fill(true, mf->get_access_sector_mask());
+		}
+
+		if(mf->get_access_byte_mask().count() == m_config.get_atom_sz())
+		{
+			block->set_m_readable(true, mf->get_access_sector_mask());
+		} else
+		{
+			block->set_m_readable(false, mf->get_access_sector_mask());
+		}
+
+		if( m_status != RESERVATION_FAIL ){
+			   // If evicted block is modified and not a write-through
+			   // (already modified lower level)
+			   if( wb && (m_config.m_write_policy != WRITE_THROUGH) ) {
+				   mem_fetch *wb = m_memfetch_creator->alloc(evicted.m_block_addr,
+					   m_wrbk_type,evicted.m_modified_size,true);
+				   send_write_request(wb, cache_event(WRITE_BACK_REQUEST_SENT, evicted), time, events);
+			   }
+			   return MISS;
+		   }
+		return RESERVATION_FAIL;
 }
 
 /// No write-allocate miss: Simply send write request to lower level memory
@@ -1531,7 +1582,7 @@ enum cache_request_status tex_cache::access( new_addr_type addr, mem_fetch *mf,
     if ( status == MISS ) {
         // we need to send a memory request...
         unsigned rob_index = m_rob.push( rob_entry(cache_index, mf, block_addr) );
-        m_extra_mf_fields[mf] = extra_mf_fields(rob_index);
+        m_extra_mf_fields[mf] = extra_mf_fields(rob_index, m_config);
         mf->set_data_size(m_config.get_line_sz());
         m_tags.fill(cache_index,time,mf); // mark block as valid
         m_request_fifo.push(mf);
@@ -1586,6 +1637,23 @@ void tex_cache::cycle(){
 /// Place returning cache block into reorder buffer
 void tex_cache::fill( mem_fetch *mf, unsigned time )
 {
+	if(m_config.m_mshr_type == SECTOR_TEX_FIFO) {
+	assert(mf->get_original_mf());
+	extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf->get_original_mf());
+    assert( e != m_extra_mf_fields.end() );
+    e->second.pending_read--;
+
+    if(e->second.pending_read > 0) {
+    	//wait for the other requests to come back
+    	delete mf;
+    	return;
+      } else {
+    	mem_fetch *temp = mf;
+    	mf = mf->get_original_mf();
+    	delete temp;
+      }
+	}
+
     extra_mf_fields_lookup::iterator e = m_extra_mf_fields.find(mf);
     assert( e != m_extra_mf_fields.end() );
     assert( e->second.m_valid );
diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h
index 76b07a2..4ed382c 100644
--- a/src/gpgpu-sim/gpu-cache.h
+++ b/src/gpgpu-sim/gpu-cache.h
@@ -355,15 +355,18 @@ struct sector_cache_block : public cache_block_t {
 
 		return m_status[sidx];
 	}
+
     virtual void set_status(enum cache_block_state status, mem_access_sector_mask_t sector_mask)
 	{
 		unsigned sidx = get_sector_index(sector_mask);
 		m_status[sidx] = status;
 	}
+
     virtual unsigned get_last_access_time()
 	{
 		return m_line_last_access_time;
 	}
+
     virtual void set_last_access_time(unsigned time, mem_access_sector_mask_t sector_mask)
 	{
 		unsigned sidx = get_sector_index(sector_mask);
@@ -371,25 +374,30 @@ struct sector_cache_block : public cache_block_t {
 		m_last_sector_access_time[sidx] = time;
 		m_line_last_access_time = time;
 	}
+
     virtual unsigned get_alloc_time()
 	{
 		return m_line_alloc_time;
 	}
+
     virtual void set_ignore_on_fill(bool m_ignore, mem_access_sector_mask_t sector_mask)
 	{
 		unsigned sidx = get_sector_index(sector_mask);
 		m_ignore_on_fill_status[sidx] = m_ignore;
 	}
+
     virtual void set_modified_on_fill(bool m_modified, mem_access_sector_mask_t sector_mask)
 	{
 		unsigned sidx = get_sector_index(sector_mask);
 		m_set_modified_on_fill[sidx] = m_modified;
 	}
+
     virtual void set_m_readable(bool readable, mem_access_sector_mask_t sector_mask)
     {
     	unsigned sidx = get_sector_index(sector_mask);
     	m_readable[sidx] = readable;
     }
+
     virtual bool is_readable(mem_access_sector_mask_t sector_mask) {
     	unsigned sidx = get_sector_index(sector_mask);
     	return m_readable[sidx];
@@ -447,25 +455,30 @@ enum write_policy_t {
 
 enum allocation_policy_t {
     ON_MISS,
-    ON_FILL
+    ON_FILL,
+	STREAMING
 };
 
 
 enum write_allocate_policy_t {
 	NO_WRITE_ALLOCATE,
 	WRITE_ALLOCATE,
-	FETCH_ON_WRITE
+	FETCH_ON_WRITE,
+	LAZY_FETCH_ON_READ
 };
 
 enum mshr_config_t {
-    TEX_FIFO,
+    TEX_FIFO, // Tex cache
     ASSOC, // normal cache
+	SECTOR_TEX_FIFO,  //Tex cache sends requests to high-level sector cache
 	SECTOR_ASSOC // normal cache sends requests to high-level sector cache
 };
 
 enum set_index_function{
-    FERMI_HASH_SET_FUNCTION = 0,
-    LINEAR_SET_FUNCTION,
+	LINEAR_SET_FUNCTION = 0,
+	BITWISE_XORING_FUNCTION,
+	HASH_IPOLY_FUNCTION,
+	FERMI_HASH_SET_FUNCTION,
     CUSTOM_SET_FUNCTION
 };
 
@@ -474,6 +487,12 @@ enum cache_type{
     SECTOR
 };
 
+#define MAX_WARP_PER_SHADER 64
+#define INCT_TOTAL_BUFFER 64
+#define L2_TOTAL 64
+#define MAX_WARP_PER_SHADER 64
+#define MAX_WARP_PER_SHADER 64
+
 class cache_config {
 public:
     cache_config() 
@@ -533,10 +552,27 @@ public:
         switch (ap) {
         case 'm': m_alloc_policy = ON_MISS; break;
         case 'f': m_alloc_policy = ON_FILL; break;
+        case 's': m_alloc_policy = STREAMING; break;
         default: exit_parse_error();
         }
+        if(m_alloc_policy == STREAMING) {
+        	//For streaming cache, we set the alloc policy to be on-fill to remove all line_alloc_fail stalls
+        	//we set the MSHRs to be equal to the cache line. This is possible by moving TAG to be shared between cache line and MSHR enrty (i.e. for each cache line, there is an MSHR rntey associated with it)
+        	// This is the easiest think we can think about to model (mimics) L1 streaming cache in Pascal and Volta
+        	//Based on our microbenchmakrs, MSHRs entries have been increasing substantially in Pascal and Volta
+        	//For more information about streaming cache, see:
+        	// http://on-demand.gputechconf.com/gtc/2017/presentation/s7798-luke-durant-inside-volta.pdf
+        	// https://ieeexplore.ieee.org/document/8344474/
+
+			m_alloc_policy = ON_FILL;
+			m_mshr_entries = m_nset*m_assoc;
+			if(m_cache_type == SECTOR)
+				m_mshr_entries *=  SECTOR_CHUNCK_SIZE;
+			m_mshr_max_merge = MAX_WARP_PER_SM;
+        }
         switch (mshr_type) {
         case 'F': m_mshr_type = TEX_FIFO; assert(ntok==14); break;
+        case 'T': m_mshr_type = SECTOR_TEX_FIFO; assert(ntok==14); break;
         case 'A': m_mshr_type = ASSOC; break;
         case 'S' : m_mshr_type = SECTOR_ASSOC; break;
         default: exit_parse_error();
@@ -553,6 +589,7 @@ public:
         case 'N': m_write_alloc_policy = NO_WRITE_ALLOCATE; break;
         case 'W': m_write_alloc_policy = WRITE_ALLOCATE; break;
         case 'F': m_write_alloc_policy = FETCH_ON_WRITE; break;
+        case 'L': m_write_alloc_policy = LAZY_FETCH_ON_READ; break;
 		default: exit_parse_error();
         }
 
@@ -570,9 +607,9 @@ public:
             assert(0 && "Invalid cache configuration: Writeback cache cannot allocate new line on fill. "); 
         }
 
-        if(m_write_alloc_policy == FETCH_ON_WRITE && m_alloc_policy == ON_FILL)
+        if((m_write_alloc_policy == FETCH_ON_WRITE || m_write_alloc_policy == LAZY_FETCH_ON_READ )&& m_alloc_policy == ON_FILL)
 		{
-			assert(0 && "Invalid cache configuration: FETCH_ON_WRITE cannot work properly with ON_FILL policy. Cache must be ON_MISS. ");
+			assert(0 && "Invalid cache configuration: FETCH_ON_WRITE and LAZY_FETCH_ON_READ cannot work properly with ON_FILL policy. Cache must be ON_MISS. ");
 		}
         if(m_cache_type == SECTOR)
 		{
@@ -587,6 +624,7 @@ public:
 
         switch(sif){
         case 'H': m_set_index_function = FERMI_HASH_SET_FUNCTION; break;
+        case 'P': m_set_index_function = HASH_IPOLY_FUNCTION; break;
         case 'C': m_set_index_function = CUSTOM_SET_FUNCTION; break;
         case 'L': m_set_index_function = LINEAR_SET_FUNCTION; break;
         default: exit_parse_error();
@@ -709,6 +747,7 @@ class l1d_cache_config : public cache_config{
 public:
 	l1d_cache_config() : cache_config(){}
 	virtual unsigned set_index(new_addr_type addr) const;
+	unsigned l1_latency;
 };
 
 class l2_cache_config : public cache_config {
@@ -739,7 +778,8 @@ public:
     unsigned size() const { return m_config.get_num_lines();}
     cache_block_t* get_block(unsigned idx) { return m_lines[idx];}
 
-    void flush(); // flash invalidate all entries
+    void flush(); // flush all written entries
+    void invalidate(); // invalidate all entries
     void new_window();
 
     void print( FILE *stream, unsigned &total_access, unsigned &total_misses ) const;
@@ -991,6 +1031,7 @@ public:
     mem_fetch *next_access(){return m_mshrs.next_access();}
     // flash invalidate all entries in cache
     void flush(){m_tag_array->flush();}
+    void invalidate(){m_tag_array->invalidate();}
     void print(FILE *fp, unsigned &accesses, unsigned &misses) const;
     void display_state( FILE *fp ) const;
 
@@ -1176,6 +1217,7 @@ public:
         case NO_WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_no_wa; break;
 		case WRITE_ALLOCATE: m_wr_miss = &data_cache::wr_miss_wa_naive; break;
 		case FETCH_ON_WRITE: m_wr_miss = &data_cache::wr_miss_wa_fetch_on_write; break;
+		case LAZY_FETCH_ON_READ: m_wr_miss = &data_cache::wr_miss_wa_lazy_fetch_on_read; break;
         default:
             assert(0 && "Error: Must set valid cache write miss policy\n");
             break; // Need to set a write miss function
@@ -1296,7 +1338,14 @@ protected:
 							mem_fetch *mf,
 							unsigned time,
 							std::list<cache_event> &events,
-							enum cache_request_status status ); // write-allocate with read-fetch-only
+							enum cache_request_status status ); // write-allocate with fetch-on-every-write
+	enum cache_request_status
+				   wr_miss_wa_lazy_fetch_on_read( new_addr_type addr,
+								unsigned cache_index,
+								mem_fetch *mf,
+								unsigned time,
+								std::list<cache_event> &events,
+								enum cache_request_status status ); // write-allocate with read-fetch-only
 	enum cache_request_status
 				wr_miss_wa_write_validate( new_addr_type addr,
 							unsigned cache_index,
@@ -1419,7 +1468,7 @@ public:
     m_result_fifo(config.m_result_fifo_entries)
     {
         m_name = name;
-        assert(config.m_mshr_type == TEX_FIFO);
+        assert(config.m_mshr_type == TEX_FIFO || config.m_mshr_type == SECTOR_TEX_FIFO );
         assert(config.m_write_policy == READ_ONLY);
         assert(config.m_alloc_policy == ON_MISS);
         m_memport=memport;
@@ -1572,13 +1621,15 @@ private:
 
     struct extra_mf_fields {
         extra_mf_fields()  { m_valid = false;}
-        extra_mf_fields( unsigned i ) 
+        extra_mf_fields( unsigned i, const cache_config &m_config )
         {
             m_valid = true;
             m_rob_index = i;
+            pending_read = m_config.m_mshr_type == SECTOR_TEX_FIFO? m_config.m_line_sz/SECTOR_SIZE : 0;
         }
         bool m_valid;
         unsigned m_rob_index;
+        unsigned pending_read;
     };
 
     cache_stats m_stats;
diff --git a/src/gpgpu-sim/gpu-sim.cc b/src/gpgpu-sim/gpu-sim.cc
index c5d4464..08d4525 100644
--- a/src/gpgpu-sim/gpu-sim.cc
+++ b/src/gpgpu-sim/gpu-sim.cc
@@ -250,6 +250,12 @@ void shader_core_config::reg_options(class OptionParser * opp)
                    "per-shader L1 data cache config "
                    " {<nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>,<mshr>:<N>:<merge>,<mq> | none}",
                    "none" );
+    option_parser_register(opp, "-l1_latency", OPT_UINT32, &m_L1D_config.l1_latency,
+                 "L1 Hit Latency",
+                 "0");
+    option_parser_register(opp, "-smem_latency", OPT_UINT32, &smem_latency,
+                 "smem Latency",
+                 "3");
     option_parser_register(opp, "-gpgpu_cache:dl1PrefL1", OPT_CSTR, &m_L1D_config.m_config_stringPrefL1,
                    "per-shader L1 data cache config "
                    " {<nsets>:<bsize>:<assoc>,<rep>:<wr>:<alloc>:<wr_alloc>,<mshr>:<N>:<merge>,<mq> | none}",
@@ -1495,7 +1501,8 @@ void gpgpu_sim::cycle()
           } else {
               mem_fetch* mf = (mem_fetch*) icnt_pop( m_shader_config->mem2device(i) );
               m_memory_sub_partition[i]->push( mf, gpu_sim_cycle + gpu_tot_sim_cycle );
-              partiton_reqs_in_parallel_per_cycle++;
+              if(mf)
+            	  partiton_reqs_in_parallel_per_cycle++;
           }
           m_memory_sub_partition[i]->cache_cycle(gpu_sim_cycle+gpu_tot_sim_cycle);
           m_memory_sub_partition[i]->accumulate_L2cache_stats(m_power_stats->pwr_mem_stat->l2_cache_stats[CURRENT_STAT_IDX]);
@@ -1548,12 +1555,12 @@ void gpgpu_sim::cycle()
 
       issue_block2core();
       
-      // Depending on configuration, flush the caches once all of threads are completed.
+      // Depending on configuration, invalidate the caches once all of threads are completed.
       int all_threads_complete = 1;
       if (m_config.gpgpu_flush_l1_cache) {
          for (unsigned i=0;i<m_shader_config->n_simt_clusters;i++) {
             if (m_cluster[i]->get_not_completed() == 0)
-                m_cluster[i]->cache_flush();
+                m_cluster[i]->cache_invalidate();
             else
                all_threads_complete = 0 ;
          }
@@ -1575,7 +1582,7 @@ void gpgpu_sim::cycle()
                int dlc = 0;
                for (unsigned i=0;i<m_memory_config->m_n_mem;i++) {
                   dlc = m_memory_sub_partition[i]->flushL2();
-                  assert (dlc == 0); // need to model actual writes to DRAM here
+                  assert (dlc == 0); // TODO: need to model actual writes to DRAM here
                   printf("Dirty lines flushed from L2 %d is %d\n", i, dlc  );
                }
             }
diff --git a/src/gpgpu-sim/l2cache.cc b/src/gpgpu-sim/l2cache.cc
index b1465a8..25da107 100644
--- a/src/gpgpu-sim/l2cache.cc
+++ b/src/gpgpu-sim/l2cache.cc
@@ -362,10 +362,11 @@ void memory_sub_partition::cache_cycle( unsigned cycle )
            }else{
         	    if(m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE)
         	    {
-			 		assert(mf->original_wr_mf);
-			 		mf->original_wr_mf->set_reply();
-			 		mf->original_wr_mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle);
-			 		m_L2_icnt_queue->push(mf->original_wr_mf);
+        	    	mem_fetch* original_wr_mf = mf->get_original_wr_mf();
+					assert(original_wr_mf);
+					original_wr_mf->set_reply();
+					original_wr_mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle);
+					m_L2_icnt_queue->push(original_wr_mf);
         	    }
 				m_request_tracker.erase(mf);
 				delete mf;
@@ -428,7 +429,7 @@ void memory_sub_partition::cache_cycle( unsigned cycle )
                         m_icnt_L2_queue->pop();
                     }
                 } else if ( status != RESERVATION_FAIL ) {
-                	if(mf->is_write() && m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE && !was_writeallocate_sent(events)) {
+                	if(mf->is_write() && (m_config->m_L2_config.m_write_alloc_policy == FETCH_ON_WRITE || m_config->m_L2_config.m_write_alloc_policy == LAZY_FETCH_ON_READ) && !was_writeallocate_sent(events)) {
                 		mf->set_reply();
                 		mf->set_status(IN_PARTITION_L2_TO_ICNT_QUEUE,gpu_sim_cycle+gpu_tot_sim_cycle);
                 		m_L2_icnt_queue->push(mf);
@@ -568,7 +569,15 @@ unsigned memory_sub_partition::flushL2()
     if (!m_config->m_L2_config.disabled()) {
         m_L2cache->flush(); 
     }
-    return 0; // L2 is read only in this version
+    return 0;   //TODO: write the flushed data to the main memory
+}
+
+unsigned memory_sub_partition::invalidateL2()
+{
+    if (!m_config->m_L2_config.disabled()) {
+        m_L2cache->invalidate();
+    }
+    return 0;
 }
 
 bool memory_sub_partition::busy() const 
@@ -600,7 +609,7 @@ std::vector<mem_fetch*> memory_sub_partition::breakdown_request_to_sector_reques
 		} else
 			{
 			    printf("Invalid sector received, address = 0x%06x, sector mask = %s, data size = %d",
-			    		mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size(), mf->get_data_size());
+			    		mf->get_addr(), mf->get_access_sector_mask(), mf->get_data_size());
 				assert(0 && "Undefined sector mask is received");
 			}
 
@@ -631,7 +640,11 @@ std::vector<mem_fetch*> memory_sub_partition::breakdown_request_to_sector_reques
 			 result.push_back(n_mf);
 			 byte_sector_mask <<= SECTOR_SIZE;
 		}
-	} else assert(0 && "Undefined data size is received");
+	} else {
+		 printf("Invalid sector received, address = 0x%06x, sector mask = %d, byte mask = , data size = %d",
+					    		mf->get_addr(), mf->get_access_sector_mask().count(), mf->get_data_size());
+		 assert(0 && "Undefined data size is received");
+	}
 
 	return result;
 }
diff --git a/src/gpgpu-sim/l2cache.h b/src/gpgpu-sim/l2cache.h
index 2d13918..18c0a8b 100644
--- a/src/gpgpu-sim/l2cache.h
+++ b/src/gpgpu-sim/l2cache.h
@@ -162,6 +162,7 @@ public:
    void set_done( mem_fetch *mf );
 
    unsigned flushL2();
+   unsigned invalidateL2();
 
    // interface to L2_dram_queue
    bool L2_dram_queue_empty() const; 
diff --git a/src/gpgpu-sim/mem_fetch.h b/src/gpgpu-sim/mem_fetch.h
index 278cf32..da67d49 100644
--- a/src/gpgpu-sim/mem_fetch.h
+++ b/src/gpgpu-sim/mem_fetch.h
@@ -115,8 +115,10 @@ public:
    const memory_config *get_mem_config(){return m_mem_config;}
 
    unsigned get_num_flits(bool simt_to_mem);
-   mem_fetch* original_mf;
-   mem_fetch* original_wr_mf;
+
+   mem_fetch* get_original_mf() { return original_mf; }
+   mem_fetch* get_original_wr_mf()  { return original_wr_mf; }
+
 private:
    // request source information
    unsigned m_request_uid;
@@ -148,6 +150,10 @@ private:
 
    const class memory_config *m_mem_config;
    unsigned icnt_flit_size;
+
+   mem_fetch* original_mf;  //this pointer is set up when a request is divided into sector requests at L2 cache (if the req size > L2 sector size), so the pointer refers to the original request
+   mem_fetch* original_wr_mf;  //this pointer refers to the original write req, when fetch-on-write policy is used
+
 };
 
 #endif
diff --git a/src/gpgpu-sim/shader.cc b/src/gpgpu-sim/shader.cc
index 533c6f9..0e2e1c2 100644
--- a/src/gpgpu-sim/shader.cc
+++ b/src/gpgpu-sim/shader.cc
@@ -451,15 +451,15 @@ void shader_core_stats::print( FILE* fout ) const
    fprintf(fout, "gpgpu_n_intrawarp_mshr_merge = %d\n", gpgpu_n_intrawarp_mshr_merge);
    fprintf(fout, "gpgpu_n_cmem_portconflict = %d\n", gpgpu_n_cmem_portconflict);
 
-   fprintf(fout, "gpgpu_stall_shd_mem[c_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]);
-   fprintf(fout, "gpgpu_stall_shd_mem[c_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][MSHR_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[c_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[c_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][DATA_PORT_STALL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[t_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][MSHR_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[t_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[t_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][DATA_PORT_STALL]);
+   fprintf(fout, "gpgpu_stall_shd_mem[c_mem][resource_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][BK_CONF]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][MSHR_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[c_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[C_MEM][DATA_PORT_STALL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][MSHR_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[t_mem][data_port_stall] = %d\n", gpu_stall_shd_mem_breakdown[T_MEM][DATA_PORT_STALL]);
    fprintf(fout, "gpgpu_stall_shd_mem[s_mem][bk_conf] = %d\n", gpu_stall_shd_mem_breakdown[S_MEM][BK_CONF]);
-   fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][bk_conf] = %d\n", 
+   fprintf(fout, "gpgpu_stall_shd_mem[gl_mem][resource_stall] = %d\n",
            gpu_stall_shd_mem_breakdown[G_MEM_LD][BK_CONF] + 
            gpu_stall_shd_mem_breakdown[G_MEM_ST][BK_CONF] + 
            gpu_stall_shd_mem_breakdown[L_MEM_LD][BK_CONF] + 
@@ -477,22 +477,22 @@ void shader_core_stats::print( FILE* fout ) const
            gpu_stall_shd_mem_breakdown[L_MEM_LD][DATA_PORT_STALL] + 
            gpu_stall_shd_mem_breakdown[L_MEM_ST][DATA_PORT_STALL]    
            ); // data port stall at data cache 
-   fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][MSHR_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_CACHE_RSRV_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][MSHR_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_CACHE_RSRV_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][MSHR_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_CACHE_RSRV_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][MSHR_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_ICNT_RC_FAIL]);
-   fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_CACHE_RSRV_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][MSHR_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_LD][WB_CACHE_RSRV_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][MSHR_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[g_mem_st][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[G_MEM_ST][WB_CACHE_RSRV_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][MSHR_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_LD][WB_CACHE_RSRV_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][mshr_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][MSHR_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_st][icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_icnt_rc] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_ICNT_RC_FAIL]);
+   //fprintf(fout, "gpgpu_stall_shd_mem[l_mem_ld][wb_rsrv_fail] = %d\n", gpu_stall_shd_mem_breakdown[L_MEM_ST][WB_CACHE_RSRV_FAIL]);
 
    fprintf(fout, "gpu_reg_bank_conflict_stalls = %d\n", gpu_reg_bank_conflict_stalls);
 
@@ -1488,6 +1488,111 @@ mem_stage_stall_type ldst_unit::process_memory_access_queue( cache_t *cache, war
     return process_cache_access( cache, mf->get_addr(), inst, events, mf, status );
 }
 
+mem_stage_stall_type ldst_unit::process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst )
+{
+    mem_stage_stall_type result = NO_RC_FAIL;
+    if( inst.accessq_empty() )
+        return result;
+
+    mem_fetch *mf = m_mf_allocator->alloc(inst,inst.accessq_back());
+
+    if(m_config->m_L1D_config.l1_latency > 0)
+	{
+    	if((l1_latency_queue[m_config->m_L1D_config.l1_latency-1]) == NULL)
+    	{
+    		l1_latency_queue[m_config->m_L1D_config.l1_latency-1] = mf;
+
+    		if( mf->get_inst().is_store() ) {
+				unsigned inc_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)?
+						(mf->get_data_size()/SECTOR_SIZE) : 1;
+
+				for(unsigned i=0; i< inc_ack; ++i)
+					m_core->inc_store_req( inst.warp_id() );
+			}
+
+    		inst.accessq_pop_back();
+    	}
+    	else
+        {
+        	result = BK_CONF;
+        	delete mf;
+        }
+        if( !inst.accessq_empty() &&  result !=BK_CONF)
+		   result = COAL_STALL;
+	   return result;
+	}
+    else
+    {
+		std::list<cache_event> events;
+		enum cache_request_status status = cache->access(mf->get_addr(),mf,gpu_sim_cycle+gpu_tot_sim_cycle,events);
+		return process_cache_access( cache, mf->get_addr(), inst, events, mf, status );
+    }
+}
+
+void ldst_unit::L1_latency_queue_cycle()
+{
+	//std::deque< std::pair<mem_fetch*,bool> >::iterator it = m_latency_queue.begin();
+	if((l1_latency_queue[0]) != NULL)
+    {
+		    mem_fetch* mf_next = l1_latency_queue[0];
+			std::list<cache_event> events;
+			enum cache_request_status status = m_L1D->access(mf_next->get_addr(),mf_next,gpu_sim_cycle+gpu_tot_sim_cycle,events);
+
+		   bool write_sent = was_write_sent(events);
+		   bool read_sent = was_read_sent(events);
+
+		   if ( status == HIT ) {
+			   assert( !read_sent );
+			   l1_latency_queue[0] = NULL;
+			   if ( mf_next->get_inst().is_load() ) {
+				   for ( unsigned r=0; r < 4; r++)
+					   if (mf_next->get_inst().out[r] > 0)
+					   {
+						   assert(m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]]>0);
+						   unsigned still_pending = --m_pending_writes[mf_next->get_inst().warp_id()][mf_next->get_inst().out[r]];
+						   if(!still_pending)
+						   {
+							m_pending_writes[mf_next->get_inst().warp_id()].erase(mf_next->get_inst().out[r]);
+							m_scoreboard->releaseRegister(mf_next->get_inst().warp_id(),mf_next->get_inst().out[r]);
+							m_core->warp_inst_complete(mf_next->get_inst());
+						   }
+					   }
+			   }
+
+			   //For write hit in WB policy
+			   if(mf_next->get_inst().is_store() && !write_sent)
+			   {
+				   unsigned dec_ack = (m_config->m_L1D_config.get_mshr_type() == SECTOR_ASSOC)?
+				   						(mf_next->get_data_size()/SECTOR_SIZE) : 1;
+
+				   mf_next->set_reply();
+
+				   for(unsigned i=0; i< dec_ack; ++i)
+				      m_core->store_ack(mf_next);
+			   }
+
+			   if( !write_sent )
+				   delete mf_next;
+
+		   } else if ( status == RESERVATION_FAIL ) {
+			   assert( !read_sent );
+			   assert( !write_sent );
+		   } else {
+			   assert( status == MISS || status == HIT_RESERVED );
+			   l1_latency_queue[0] = NULL;
+	   }
+    }
+
+	 for( unsigned stage = 0; stage<m_config->m_L1D_config.l1_latency-1; ++stage)
+	  if( l1_latency_queue[stage] == NULL) {
+		   l1_latency_queue[stage] = l1_latency_queue[stage+1] ;
+		   l1_latency_queue[stage+1] = NULL;
+	   }
+
+}
+
+
+
 bool ldst_unit::constant_cycle( warp_inst_t &inst, mem_stage_stall_type &rc_fail, mem_stage_access_type &fail_type)
 {
    if( inst.empty() || ((inst.space.get_type() != const_space) && (inst.space.get_type() != param_space_kernel)) )
@@ -1561,7 +1666,7 @@ bool ldst_unit::memory_cycle( warp_inst_t &inst, mem_stage_stall_type &stall_rea
        }
    } else {
        assert( CACHE_UNDEFINED != inst.cache_op );
-       stall_cond = process_memory_access_queue(m_L1D,inst);
+       stall_cond = process_memory_access_queue_l1cache(m_L1D,inst);
    }
    if( !inst.accessq_empty() && stall_cond == NO_RC_FAIL)
        stall_cond = COAL_STALL;
@@ -1593,6 +1698,11 @@ void ldst_unit::flush(){
 	m_L1D->flush();
 }
 
+void ldst_unit::invalidate(){
+	// Flush L1D cache
+	m_L1D->invalidate();
+}
+
 simd_function_unit::simd_function_unit( const shader_core_config *config )
 { 
     m_config=config;
@@ -1766,8 +1876,9 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt,
                       const memory_config *mem_config,  
                       shader_core_stats *stats,
                       unsigned sid,
-                      unsigned tpc ) : pipelined_simd_unit(NULL,config,3,core), m_next_wb(config)
+                      unsigned tpc ) : pipelined_simd_unit(NULL,config,config->smem_latency,core), m_next_wb(config)
 {
+	assert(config->smem_latency > 1);
     init( icnt,
           mf_allocator,
           core, 
@@ -1788,6 +1899,12 @@ ldst_unit::ldst_unit( mem_fetch_interface *icnt,
                               m_icnt,
                               m_mf_allocator,
                               IN_L1D_MISS_QUEUE );
+
+        if(m_config->m_L1D_config.l1_latency > 0)
+	    {
+        	for(int i=0; i<m_config->m_L1D_config.l1_latency; i++ )
+        		l1_latency_queue.push_back((mem_fetch*)NULL);
+	    }
     }
 }
 
@@ -2014,7 +2131,11 @@ void ldst_unit::cycle()
 
    m_L1T->cycle();
    m_L1C->cycle();
-   if( m_L1D ) m_L1D->cycle();
+   if( m_L1D ) {
+	   m_L1D->cycle();
+	   if(m_config->m_L1D_config.l1_latency > 0)
+	   		L1_latency_queue_cycle();
+   }
 
    warp_inst_t &pipe_reg = *m_dispatch_reg;
    enum mem_stage_stall_type rc_fail = NO_RC_FAIL;
@@ -2037,9 +2158,9 @@ void ldst_unit::cycle()
        unsigned warp_id = pipe_reg.warp_id();
        if( pipe_reg.is_load() ) {
            if( pipe_reg.space.get_type() == shared_space ) {
-               if( m_pipeline_reg[2]->empty() ) {
+               if( m_pipeline_reg[m_config->smem_latency-1]->empty() ) {
                    // new shared memory request
-                   move_warp(m_pipeline_reg[2],m_dispatch_reg);
+                   move_warp(m_pipeline_reg[m_config->smem_latency-1],m_dispatch_reg);
                    m_dispatch_reg->clear();
                }
            } else {
@@ -2644,6 +2765,11 @@ void shader_core_ctx::cache_flush()
    m_ldst_unit->flush();
 }
 
+void shader_core_ctx::cache_invalidate()
+{
+   m_ldst_unit->invalidate();
+}
+
 // modifiers
 std::list<opndcoll_rfu_t::op_t> opndcoll_rfu_t::arbiter_t::allocate_reads() 
 {
@@ -3461,6 +3587,12 @@ void simt_core_cluster::cache_flush()
         m_core[i]->cache_flush();
 }
 
+void simt_core_cluster::cache_invalidate()
+{
+    for( unsigned i=0; i < m_config->n_simt_cores_per_cluster; i++ )
+        m_core[i]->cache_invalidate();
+}
+
 bool simt_core_cluster::icnt_injection_buffer_full(unsigned size, bool write)
 {
     unsigned request_size = size;
diff --git a/src/gpgpu-sim/shader.h b/src/gpgpu-sim/shader.h
index ae22eaa..e07096e 100644
--- a/src/gpgpu-sim/shader.h
+++ b/src/gpgpu-sim/shader.h
@@ -1155,6 +1155,7 @@ public:
      
     void fill( mem_fetch *mf );
     void flush();
+    void invalidate();
     void writeback();
 
     // accessors
@@ -1219,6 +1220,7 @@ protected:
                                                       mem_fetch *mf,
                                                       enum cache_request_status status );
    mem_stage_stall_type process_memory_access_queue( cache_t *cache, warp_inst_t &inst );
+   mem_stage_stall_type process_memory_access_queue_l1cache( l1_cache *cache, warp_inst_t &inst );
 
    const memory_config *m_memory_config;
    class mem_fetch_interface *m_icnt;
@@ -1247,6 +1249,9 @@ protected:
    // for debugging
    unsigned long long m_last_inst_gpu_sim_cycle;
    unsigned long long m_last_inst_gpu_tot_sim_cycle;
+
+   std::deque<mem_fetch* > l1_latency_queue;
+   void L1_latency_queue_cycle();
 };
 
 enum pipeline_stage_name_t {
@@ -1398,6 +1403,8 @@ struct shader_core_config : public core_config
 
     int simt_core_sim_order; 
     
+    unsigned smem_latency;
+
     unsigned mem2device(unsigned memid) const { return memid + n_simt_clusters; }
 
     //Jin: concurrent kernel on sm
@@ -1655,6 +1662,7 @@ public:
     void issue_block2core( class kernel_info_t &kernel );
 
     void cache_flush();
+    void cache_invalidate();
     void accept_fetch_response( mem_fetch *mf );
     void accept_ldst_unit_response( class mem_fetch * mf );
     void broadcast_barrier_reduction(unsigned cta_id, unsigned bar_id,warp_set_t warps);
@@ -1947,6 +1955,7 @@ public:
     void reinit();
     unsigned issue_block2core();
     void cache_flush();
+    void cache_invalidate();
     bool icnt_injection_buffer_full(unsigned size, bool write);
     void icnt_inject_request_packet(class mem_fetch *mf);