Merge branch 'kernels' of https://github.com/hansungk/vortex-private into kernels

ci/toolchain_env.sh

@@ -24,3 +24,7 @@ export PATH=$SV2V_PATH/bin:$PATH
 
 export YOSYS_PATH=$TOOLDIR/yosys
 export PATH=$YOSYS_PATH/bin:$PATH
+
+export LLVM_VORTEX=$TOOLDIR/llvm-vortex
+export POCL_CC_PATH=$TOOLDIR/pocl/compiler
+export POCL_RT_PATH=$TOOLDIR/pocl/runtime

hw/rtl/core/VX_core.sv

@@ -45,7 +45,7 @@ module VX_core import VX_gpu_pkg::*; #(
     output wire [`NUM_REGS-1:0][`XLEN-1:0] sim_wb_value,
 
     // Status
-    output wire             busy
+    output wire             busy    //stays 1 when busy, 0 when done (termination) detect the negative edge
 );
     VX_schedule_if      schedule_if();
     VX_fetch_if         fetch_if();
@@ -258,7 +258,7 @@ module VX_core import VX_gpu_pkg::*; #(
 
 `endif
 
-`ifdef PERF_ENABLE
+`ifdef PERF_ENABLE  // expose these perf counter to console using $display, %time; flag: --perf=0?
 
     wire [`CLOG2(DCACHE_NUM_REQS+1)-1:0] perf_dcache_rd_req_per_cycle;
     wire [`CLOG2(DCACHE_NUM_REQS+1)-1:0] perf_dcache_wr_req_per_cycle;
@@ -331,7 +331,57 @@ module VX_core import VX_gpu_pkg::*; #(
     assign pipeline_perf_if.stores = perf_stores;
     assign pipeline_perf_if.load_latency = perf_dcache_lat;
     assign pipeline_perf_if.ifetch_latency = perf_icache_lat;
-    assign pipeline_perf_if.load_latency = perf_dcache_lat;
+    real instrs = commit_csr_if.instret;
+    real cycles = sched_csr_if.cycles;
+    real icache_lat = perf_icache_lat;
+    real ifetches = perf_ifetches;
+    real dcache_lat = perf_dcache_lat;
+    real loads = perf_loads;
+    real scheduler_idles = pipeline_perf_if.sched_idles;
+    real scheduler_stalls = pipeline_perf_if.sched_stalls;
+    real ibuf_stalls = pipeline_perf_if.ibf_stalls;
+    real scrb_alu_per_core = pipeline_perf_if.units_uses[`EX_ALU];
+    real scrb_fpu_per_core = pipeline_perf_if.units_uses[`EX_FPU];
+    real scrb_lsu_per_core = pipeline_perf_if.units_uses[`EX_LSU];
+    real scrb_sfu_per_core = pipeline_perf_if.units_uses[`EX_SFU];
+    real scrb_tot = scrb_alu_per_core+scrb_fpu_per_core+scrb_lsu_per_core+scrb_sfu_per_core;
+
+    real scrb_wctl_per_core = pipeline_perf_if.sfu_uses[`SFU_WCTL];
+    real scrb_csrs_per_core = pipeline_perf_if.sfu_uses[`SFU_CSRS];
+    real sfu_tot = scrb_wctl_per_core+scrb_csrs_per_core;
+    
+    always @(negedge busy) begin
+        if (!reset) begin
+        $display("====================CORE : %d===================",CORE_ID);
+        $display("time : %t", $time);
+        $display("perf_dcache_rd_req_per_cycle: %d", perf_dcache_rd_req_per_cycle);
+        $display("perf_dcache_wr_req_per_cycle: %d", perf_dcache_wr_req_per_cycle);
+        $display("perf_dcache_rsp_per_cycle: %d", perf_dcache_rsp_per_cycle);
+        $display("perf_icache_pending_read_cycle: %d", perf_icache_pending_read_cycle);
+        $display("perf_dcache_pending_read_cycle: %d", perf_dcache_pending_read_cycle);
+        $display("perf_icache_pending_reads: %d", perf_icache_pending_reads);
+        $display("perf_dcache_pending_reads: %d", perf_dcache_pending_reads);
+        $display("perf_icache_req_fire: %b", perf_icache_req_fire);
+        $display("perf_icache_rsp_fire: %b", perf_icache_rsp_fire);
+        $display("perf_dcache_rd_req_fire: %b", perf_dcache_rd_req_fire);
+        $display("perf_dcache_rd_req_fire_r: %b", perf_dcache_rd_req_fire_r);
+        $display("perf_dcache_wr_req_fire: %b", perf_dcache_wr_req_fire);
+        $display("perf_dcache_wr_req_fire_r: %b", perf_dcache_wr_req_fire_r);
+        $display("perf_dcache_rsp_fire: %b", perf_dcache_rsp_fire);
+
+        $display("Instructions: %d, Cycles: %d, IPC: %f", commit_csr_if.instret, sched_csr_if.cycles, instrs/cycles);
+        $display("scheduler idle: %d (%f)", pipeline_perf_if.sched_idles, scheduler_idles/cycles);
+        $display("scheduler stalls: %d (%f)", pipeline_perf_if.sched_stalls, scheduler_stalls/cycles);
+        $display("ibuffer stalls: %d (%f)",pipeline_perf_if.ibf_stalls, ibuf_stalls/cycles);
+        $display("issue stalls: %d(alu=%f, fpu=%f, lsu=%f, sfu=%f)",pipeline_perf_if.scb_stalls, scrb_alu_per_core/scrb_tot, scrb_fpu_per_core/scrb_tot, scrb_lsu_per_core/scrb_tot, scrb_sfu_per_core/scrb_tot);
+        $display("sfu stalls: %d (scrs=%f, wctl=%f)",pipeline_perf_if.units_uses[`EX_SFU], scrb_csrs_per_core/sfu_tot, scrb_wctl_per_core/sfu_tot);
+        $display("ifetches: %d", perf_ifetches);
+        $display("ifetch latency: %f Cycles", icache_lat/ifetches);
+        $display("loads: %d", perf_loads);
+        $display("load latency: %f Cycles", dcache_lat/loads);
+        $display("stores: %d", perf_stores);
+        end
+    end
 
 `endif
 

kernel/Makefile

@@ -51,10 +51,10 @@ $(PROJECT).dump: $(PROJECT).a
 %.S.o: src/%.S
 	$(CC) $(CFLAGS) -c $< -o $@
 
-%.cpp.o: src/%.cpp
+%.cpp.o: src/%.cpp include/vx_spawn.h
 	$(CXX) $(CFLAGS) -c $< -o $@
 
-%.c.o: src/%.c
+%.c.o: src/%.c include/vx_spawn.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
 $(PROJECT).a: $(OBJS)

kernel/include/vx_spawn.h

@@ -17,6 +17,10 @@
 #include <stdint.h>
 #include <stdio.h>
 
+#ifndef CORES_PER_CLUSTER
+#define CORES_PER_CLUSTER 2
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -48,6 +52,8 @@ void vx_wspawn_wait();
 void vx_spawn_kernel(context_t * ctx, vx_spawn_kernel_cb callback, void * arg);
 
 void vx_spawn_tasks(int num_tasks, vx_spawn_tasks_cb callback, void * arg);
+void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void * arg);
+void vx_spawn_tasks_contiguous(int num_tasks, vx_spawn_tasks_cb callback , void * arg);
 
 void vx_serial(vx_serial_cb callback, void * arg);
 

kernel/src/vx_spawn.c

@@ -74,6 +74,52 @@ static void __attribute__ ((noinline)) spawn_tasks_all_stub() {
   }
 }
 
+static void __attribute__ ((noinline)) spawn_tasks_contiguous_all_stub() {
+  int NT  = vx_num_threads();
+  int NW  = vx_num_warps();
+  int cid = vx_core_id();
+  int wid = vx_warp_id();
+  int tid = vx_thread_id();
+
+  wspawn_tasks_args_t* p_wspawn_args = (wspawn_tasks_args_t*)g_wspawn_args[cid];
+
+  int waves = p_wspawn_args->NWs + (wid < p_wspawn_args->RWs);
+  int offset = p_wspawn_args->offset + (NT * wid + tid);
+
+  vx_spawn_tasks_cb callback = p_wspawn_args->callback;
+  void* arg = p_wspawn_args->arg;
+  for (int wave_id = 0; wave_id < waves; ++wave_id) {
+    int task_id = offset + (wave_id * NT * NW);
+    callback(task_id, arg);
+  }
+}
+
+static void __attribute__ ((noinline)) spawn_tasks_cluster_all_stub() {
+  int NT  = vx_num_threads();
+  int NW  = vx_num_warps();
+  int cid = vx_core_id();
+  int wid = vx_warp_id();
+  int tid = vx_thread_id();
+
+  const int core_id_in_cluster = cid % CORES_PER_CLUSTER;
+  // round-robin warp_id allocation across cores in cluster
+  const int wid_in_cluster = CORES_PER_CLUSTER * wid + core_id_in_cluster;
+
+  wspawn_tasks_args_t* p_wspawn_args = (wspawn_tasks_args_t*)g_wspawn_args[cid];
+
+  int waves = p_wspawn_args->NWs + (wid < p_wspawn_args->RWs);
+  int offset = p_wspawn_args->offset + (NT * wid_in_cluster + tid);
+
+  vx_spawn_tasks_cb callback = p_wspawn_args->callback;
+  void* arg = p_wspawn_args->arg;
+
+  // sequential iterations
+  for (int wave_id = 0; wave_id < waves; ++wave_id) {
+    int task_id = offset + (wave_id * NT * NW * CORES_PER_CLUSTER);
+    callback(task_id, arg);
+  }
+}
+
 static void __attribute__ ((noinline)) spawn_tasks_rem_stub() {
   int cid = vx_core_id();
   int tid = vx_thread_id();
@@ -83,6 +129,47 @@ static void __attribute__ ((noinline)) spawn_tasks_rem_stub() {
   (p_wspawn_args->callback)(task_id, p_wspawn_args->arg);
 }
 
+static void __attribute__ ((noinline)) spawn_tasks_cluster_rem_stub() {
+  int NT  = vx_num_threads();
+  int cid = vx_core_id();
+  int tid = vx_thread_id();
+  int wid = vx_warp_id();
+
+  const int core_id_in_cluster = cid % CORES_PER_CLUSTER;
+  // round-robin warp_id allocation across cores in cluster
+  const int wid_in_cluster = CORES_PER_CLUSTER * wid + core_id_in_cluster;
+
+  wspawn_tasks_args_t* p_wspawn_args = (wspawn_tasks_args_t*)g_wspawn_args[cid];
+  // FIXME: This assumes that all cores but the last one are working with full
+  // warps, and only the last core has a partially-filled warp.
+  int offset = p_wspawn_args->offset + (NT * wid_in_cluster + tid);
+
+  int task_id = offset;
+  (p_wspawn_args->callback)(task_id, p_wspawn_args->arg);
+}
+
+static void __attribute__ ((noinline)) spawn_tasks_contiguous_all_cb() {
+  // activate all threads
+  vx_tmc(-1);
+
+  // call stub routine
+  spawn_tasks_contiguous_all_stub();
+
+  // disable warp
+  vx_tmc_zero();
+}
+
+static void __attribute__ ((noinline)) spawn_tasks_cluster_all_cb() {
+  // activate all threads
+  vx_tmc(-1);
+
+  // call stub routine
+  spawn_tasks_cluster_all_stub();
+
+  // disable warp
+  vx_tmc_zero();
+}
+
 static void __attribute__ ((noinline)) spawn_tasks_all_cb() {  
   // activate all threads
   vx_tmc(-1);
@@ -94,6 +181,171 @@ static void __attribute__ ((noinline)) spawn_tasks_all_cb() {
   vx_tmc_zero();
 }
 
+// This function runs in every core, but with only 1 warp and 1 thread enabled.
+// The logic in this function figures out how many warps/threads this particular
+// core has to enable to fulfill an entire grid of computation.
+void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg) {
+  // device specs
+  const int NC = vx_num_cores();
+  const int NW = vx_num_warps();
+  const int NT = vx_num_threads();
+  // NOTE: assumes divisible
+  const int num_cluster = NC / CORES_PER_CLUSTER;
+
+  // current core id
+  int core_id = vx_core_id();
+  if (core_id >= NUM_CORES_MAX)
+    return;
+  const int cluster_id = core_id / CORES_PER_CLUSTER;
+  const int core_id_in_cluster = core_id % CORES_PER_CLUSTER;
+
+  // Distribute threads equally across as many cores as possible, even if they
+  // don't fill up NW*NT in a single core.  This makes sure the warps get evenly
+  // distributed in a single cluster
+  //
+  // TODO: Try to contain in a single cluster if possible?
+  const int num_active_cores = (num_tasks + (NT - 1)) / NT;
+  if (core_id >= num_active_cores)
+    return; // terminate extra cores
+
+  // FIXME: assumes num_tasks is divisible by num_cluster
+  const int num_tasks_this_cluster = num_tasks / num_cluster;
+  const int num_full_warps = num_tasks_this_cluster / NT;
+  const int rem_threads_in_last_warp = num_tasks_this_cluster % NT;
+  // const int num_warps = (num_tasks_this_cluster + (NT - 1)) / NT;
+
+  int num_warps_this_core = num_full_warps / CORES_PER_CLUSTER;
+  const int num_warps_in_last_row = num_full_warps % CORES_PER_CLUSTER;
+  if (core_id_in_cluster < num_warps_in_last_row) {
+    num_warps_this_core++;
+  }
+  // if 0, last warp is full-threads enabled
+  int rem_threads_in_last_warp_this_core = 0;
+  if (rem_threads_in_last_warp != 0) {
+    if (core_id_in_cluster == num_warps_in_last_row - 1) {
+      rem_threads_in_last_warp_this_core = rem_threads_in_last_warp;
+    }
+  }
+
+  // sequential iterations
+  const int num_full_waves = num_warps_this_core / NW;
+  const int rem_full_warps_in_last_wave = num_warps_this_core % NW;
+
+  const const int offset = cluster_id * num_tasks_this_cluster;
+  wspawn_tasks_args_t wspawn_args = {callback, arg, offset, num_full_waves,
+                                     rem_full_warps_in_last_wave};
+  g_wspawn_args[core_id] = &wspawn_args;
+
+  if (num_warps_this_core > 0) {
+    // execute callback on other warps
+    const int nw = MIN(num_warps_this_core, NW);
+    vx_wspawn(nw, spawn_tasks_cluster_all_cb);
+
+    // activate all threads
+    vx_tmc(-1);
+
+    // call stub routine
+    spawn_tasks_cluster_all_stub();
+
+    // back to single-threaded
+    vx_tmc_one();
+
+    // wait for spawn warps to terminate
+    vx_wspawn_wait();
+  }
+
+  // TODO: Instead of launching an additional wave just to work on remaining
+  // threads, handle this in the last wave amongst other full warps.
+  if (rem_threads_in_last_warp != 0 && core_id_in_cluster == 0) {
+    // adjust offset
+    // FIXME: consider cluster_id here
+    // FIXME: use rem_threads_in_last_warp_this_core
+    wspawn_args.offset += (num_tasks_this_cluster - rem_threads_in_last_warp);
+
+    // activate remaining threads
+    const int tmask = (1 << rem_threads_in_last_warp) - 1;
+    vx_tmc(tmask);
+
+    // call stub routine
+    spawn_tasks_cluster_rem_stub();
+
+    // back to single-threaded
+    vx_tmc_one();
+  }
+}
+
+void vx_spawn_tasks_contiguous(int num_tasks, vx_spawn_tasks_cb callback , void * arg) {
+	// device specs
+  int NC = vx_num_cores();
+  int NW = vx_num_warps();
+  int NT = vx_num_threads();
+
+  // current core id
+  int core_id = vx_core_id();
+  if (core_id >= NUM_CORES_MAX)
+    return;
+
+  // calculate necessary active cores
+  int WT = NW * NT;
+  int nC = (num_tasks > WT) ? (num_tasks / WT) : 1;
+  int nc = MIN(nC, NC);
+  if (core_id >= nc)
+    return; // terminate extra cores
+
+  // number of tasks per core
+  int tasks_per_core = num_tasks / nc;
+  int tasks_per_core_n1 = tasks_per_core;  
+  if (core_id == (nc-1)) {    
+    int rem = num_tasks - (nc * tasks_per_core); 
+    tasks_per_core_n1 += rem; // last core also executes remaining tasks
+  }
+
+  // number of tasks per warp
+  int TW = tasks_per_core_n1 / NT;      // occupied warps
+  int rT = tasks_per_core_n1 - TW * NT; // remaining threads
+  int fW = 1, rW = 0;
+  if (TW >= NW) {
+    fW = TW / NW;			                  // full warps iterations
+    rW = TW - fW * NW;                  // remaining warps
+  }
+
+  wspawn_tasks_args_t wspawn_args = { callback, arg, core_id * tasks_per_core, fW, rW };
+  g_wspawn_args[core_id] = &wspawn_args;
+
+	if (TW >= 1)	{
+    // execute callback on other warps
+    int nw = MIN(TW, NW);
+	  vx_wspawn(nw, spawn_tasks_contiguous_all_cb);
+
+    // activate all threads
+    vx_tmc(-1);
+
+    // call stub routine
+    spawn_tasks_contiguous_all_stub();
+  
+    // back to single-threaded
+    vx_tmc_one();
+    
+    // wait for spawn warps to terminate
+    vx_wspawn_wait();
+	}  
+
+  if (rT != 0) {
+    // adjust offset
+    wspawn_args.offset += (tasks_per_core_n1 - rT);
+    
+    // activate remaining threads  
+    int tmask = (1 << rT) - 1;
+    vx_tmc(tmask);
+
+    // call stub routine
+    spawn_tasks_rem_stub();
+
+    // back to single-threaded
+    vx_tmc_one();
+  }
+}
+
 void vx_spawn_tasks(int num_tasks, vx_spawn_tasks_cb callback , void * arg) {
 	// device specs
   int NC = vx_num_cores();

kernel/src/vx_start.S

@@ -102,6 +102,8 @@ init_regs:
 #endif
   csrr  t0, VX_CSR_MHARTID
   sll   t1, t0, STACK_LOG2_SIZE
+  sll   t2, t0, 2
+  add   t1, t1, t2
   sub   sp, sp, t1
 
   # set thread pointer register

tests/.gitignore

@@ -1 +1,7 @@
 **/*.log
+.depend
+*.bin
+*.dump
+*.elf
+*.o
+*.ll

tests/opencl/convolution/main.cc

@@ -238,9 +238,9 @@ int main (int argc, char **argv) {
   }
 
   // NOTE(hansung): Dump operand buffer to a file
-  if (write_operand_file("matmul.input.a.bin", h_a.data(), nbytes) != 0)
+  if (write_operand_file("convolution.input.input.bin", h_i.data(), i_nbytes) != 0)
     return EXIT_FAILURE;
-  if (write_operand_file("matmul.input.b.bin", h_b.data(), nbytes) != 0)
+  if (write_operand_file("convolution.input.weights.bin", h_w.data(), w_nbytes) != 0)
     return EXIT_FAILURE;
 
   // Creating command queue

tests/regression/common.mk

@@ -78,17 +78,23 @@ endif
 endif
 endif
 
-all: $(PROJECT) kernel.bin kernel.dump
+all: $(PROJECT) kernel.bin kernel.dump kernel.radiance.dump
 
 kernel.dump: kernel.elf
 	$(VX_DP) -D kernel.elf > kernel.dump
 
-kernel.bin: kernel.elf
+kernel.radiance.dump: kernel.radiance.elf
+	$(VX_DP) -D kernel.radiance.elf > kernel.radiance.dump
+
+kernel.bin: kernel.elf kernel.radiance.elf
 	$(VX_CP) -O binary kernel.elf kernel.bin
 
 kernel.elf: $(VX_SRCS)
 	$(VX_CXX) $(VX_CFLAGS) $(VX_SRCS) $(VX_LDFLAGS) -o kernel.elf
 
+kernel.radiance.elf: $(VX_SRCS)
+	$(VX_CXX) $(VX_CFLAGS) $(VX_SRCS) $(VX_LDFLAGS) -DRADIANCE -o kernel.radiance.elf
+
 $(PROJECT): $(SRCS)
 	$(CXX) $(CXXFLAGS) $^ $(LDFLAGS) -o $@
 
@@ -115,7 +121,7 @@ clean:
 	rm -rf $(PROJECT) *.o .depend
 
 clean-all: clean
-	rm -rf *.elf *.bin *.dump
+	rm -rf kernel.elf kernel.radiance.elf *.dump
 
 ifneq ($(MAKECMDGOALS),clean)
     -include .depend

tests/regression/sgemm_wg/Makefile

@@ -1,6 +1,6 @@
 PROJECT = sgemm_wg
 
-SRCS = main.cpp
+SRCS = main.cpp common.h
 
 VX_SRCS = kernel.cpp
 

tests/regression/sgemm_wg/kernel.cpp

@@ -1,84 +1,192 @@
 #include <stdint.h>
 #include <vx_intrinsics.h>
+#include <vx_print.h>
 #include <vx_spawn.h>
 #include "common.h"
 
-inline void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
+// Constraints on parameters:
-                              const uint32_t tid_in_threadblock_x,
+// * Memory:
-                              const uint32_t tid_in_threadblock_y,
+//   (BM + BN) * BK * sizeof(float) <= sharedmem size.
+//   BM * BK == BN * BK >= threadblock size >= NT * CORES_PER_CLUSTER
+//     When larger, the kernel runs a sequential loop to read into sharedmem;
+//     but smaller case is not handled.
+// * Compute:
+//   ( M* N) / (TM*TN) == grid size >= NC*NW*NT
+//   (BM*BN) / (TM*TN) == threadblock size < NT * NW * CORES_PER_CLUSTER
+//   (BM*BN) / (TM*TN) == threadblock size >= NT * CORES_PER_CLUSTER
+// * Combining BM * BK >= (BM*BN) / (TM*TN) == threadblock yields
+//   BM <= BK*TM*TN
+#define BM 16
+#define BN BM
+#define BK 4
+#define TM 4
+#define TN 4
+
+void threadblock_barrier(unsigned int tid_in_threadblock, unsigned int barrier_id, unsigned int count) {
+    vx_fence();
+    vx_barrier(barrier_id, count);
+}
+
+void thread_block_gemm(kernel_arg_t *__UNIFORM__ arg,
+                              const uint32_t tid_in_threadblock,
                               const uint32_t threadblock_dim_x,
                               const uint32_t threadblock_dim_y,
                               const uint32_t threadblock_id_x,
-                              const uint32_t threadblock_id_y) {
+                              const uint32_t threadblock_id_y,
-  const float *global_a = (const float *)arg->addr_a;
+                              const uint32_t threadblock_id_in_cluster,
-  const float *global_b = (const float *)arg->addr_b;
+                              float *sharedmem_per_threadblock) {
-  float *global_c = (float *)arg->addr_c;
+  const float *A = (const float *)arg->addr_a;
+  const float *B = (const float *)arg->addr_b;
+  float *C = (float *)arg->addr_c;
 
   // assumes NT == NW == matrix_dim
   const uint32_t dim_m = arg->dim_m;
   const uint32_t dim_n = arg->dim_n;
   const uint32_t dim_k = arg->dim_k;
 
-  // FIXME: assumes local block size is square shape
+  // FIXME: Output block size is assumed to be square, i.e. BM == BN
-  const uint32_t local_row = tid_in_threadblock_y;
+  // const uint32_t BM = threadblock_dim_y;
-  const uint32_t local_col = tid_in_threadblock_x;
+  // const uint32_t BN = threadblock_dim_y;
-  const uint32_t global_row = threadblock_id_y * threadblock_dim_y + local_row;
+  // const uint32_t BK = threadblock_dim_x;
-  const uint32_t global_col = threadblock_id_x * threadblock_dim_x + local_col;
+  // constexpr uint32_t BM = 8;
+  // constexpr uint32_t BN = 8;
+  // constexpr uint32_t BK = 2;
 
-  // each thread generates one output element
+  const uint32_t local_a_row = tid_in_threadblock / BK;
-  float reg_c = 0.0f;
+  const uint32_t local_a_col = tid_in_threadblock % BK;
+  const uint32_t local_b_row = tid_in_threadblock / BN;
+  const uint32_t local_b_col = tid_in_threadblock % BN;
+  const uint32_t global_a_row = BM * threadblock_id_y + local_a_row;
+  const uint32_t global_b_col = BN * threadblock_id_x + local_b_col;
 
-  for (uint32_t k = 0; k < dim_k; k += threadblock_dim_x) {
+  const uint32_t local_c_row = tid_in_threadblock / (BN / TN);
-    float *local_a = (float *)DEV_SMEM_START_ADDR;
+  const uint32_t local_c_col = tid_in_threadblock % (BN / TN);
-    size_t local_a_elems = threadblock_dim_x * threadblock_dim_y;
-    float *local_b = (float *)DEV_SMEM_START_ADDR + local_a_elems;
 
-    uint32_t offset_global_a = dim_k * global_row + (k + local_col);
+  // each thread generates TM output element
-    uint32_t offset_global_b = dim_n * (local_row + k) + global_col;
+  float reg_c[TM * TN] = { 0.0f };
-    // local_a: threadblock_dim_y rows, threadblock_dim_x cols
+  float reg_a[TM] = { 0.0f };
-    // local_b: threadblock_dim_x rows, threadblock_dim_y cols
+  float reg_b[TN] = { 0.0f };
-    // threadblock_dim_x == block_k, threadblock_dim_y == block_m == block_n
-    local_a[threadblock_dim_x * local_row + local_col] = global_a[offset_global_a];
-    local_b[threadblock_dim_y * local_col + local_row] = global_b[offset_global_b];
 
-    vx_barrier(0, threadblock_dim_y);
+  volatile float *local_a = sharedmem_per_threadblock;
-    vx_fence();
+  // const size_t local_a_elems = threadblock_dim_x * threadblock_dim_y;
+  const size_t local_a_elems = (BM * BK);
+  volatile float *local_b = sharedmem_per_threadblock + local_a_elems;
 
-    for (uint32_t local_k = 0; local_k < threadblock_dim_x; local_k++) {
+  constexpr uint32_t stride_a = (BM * BN) / BK / (TM * TN);
-      reg_c += local_a[threadblock_dim_x * local_row + local_k] *
+  constexpr uint32_t stride_b = (BM * BN) / BN / (TM * TN);
-               local_b[threadblock_dim_y * local_col + local_k];
+
+  for (uint32_t k = 0; k < dim_k; k += BK) {
+    // Data move from GMEM to SMEM
+    //
+    // Make sure global offset values for A and B are contiguous between
+    // neighboring threads to ensure GMEM coalescing.
+    for (uint32_t load_offset = 0; load_offset < BM; load_offset += stride_a) {
+      const uint32_t global_a_offset =
+          dim_k * (global_a_row + load_offset) + (k + local_a_col);
+      local_a[BK * (local_a_row + load_offset) + local_a_col] =
+          A[global_a_offset];
+    }
+// #pragma GCC unroll 1
+    for (uint32_t load_offset = 0; load_offset < BK; load_offset += stride_b) {
+      const uint32_t global_b_offset =
+          dim_n * (k + local_b_row + load_offset) + global_b_col;
+      local_b[BN * (local_b_row + load_offset) + local_b_col] =
+          B[global_b_offset];
     }
 
-    vx_barrier(0, threadblock_dim_y);
+    threadblock_barrier(tid_in_threadblock, threadblock_id_in_cluster,
-    vx_fence();
+                        threadblock_dim_y);
+
+    // Compute single tile*tile matmul
+#pragma GCC unroll 2
+    for (uint32_t local_k = 0; local_k < BK; local_k++) {
+      // First, pump data from SMEM->RF
+#pragma GCC unroll TM
+      for (uint32_t res_idx_m = 0; res_idx_m < TM; res_idx_m++) {
+        reg_a[res_idx_m] =
+            local_a[BK * (TM * local_c_row + res_idx_m) + local_k];
+      }
+#pragma GCC unroll TN
+      for (uint32_t res_idx_n = 0; res_idx_n < TN; res_idx_n++) {
+        reg_b[res_idx_n] =
+            local_b[BN * local_k + (TN * local_c_col + res_idx_n)];
+      }
+
+      // Next, compute multiple result elements (TM*TN) by reusing data in RF
+#pragma GCC unroll TM
+      for (uint32_t res_idx_m = 0; res_idx_m < TM; res_idx_m++) {
+#pragma GCC unroll TN
+        for (uint32_t res_idx_n = 0; res_idx_n < TN; res_idx_n++) {
+          // NOTE use of local_b_row
+          reg_c[TN * res_idx_m + res_idx_n] +=
+              reg_a[res_idx_m] * reg_b[res_idx_n];
+          // reg_c[TN * res_idx_m + res_idx_n] +=
+          //     local_a[BK * (TM * local_c_row + res_idx_m) + local_k] *
+          //     local_b[BN * local_k + (TN * local_c_col + res_idx_n)];
+        }
+      }
+    }
+
+    threadblock_barrier(tid_in_threadblock, threadblock_id_in_cluster,
+                        threadblock_dim_y);
   }
 
-  global_c[dim_n * global_row + global_col] = reg_c;
+  // Store result data from RF to GMEM
+#pragma GCC unroll TM
+  for (uint32_t res_idx_m = 0; res_idx_m < TM; res_idx_m++) {
+#pragma GCC unroll TN
+    for (uint32_t res_idx_n = 0; res_idx_n < TN; res_idx_n++) {
+      // NOTE use of local_b_row and global_b_col here
+      C[dim_n * (BM * threadblock_id_y + TM * local_c_row + res_idx_m) +
+        (BN * threadblock_id_x + TN * local_c_col + res_idx_n)] =
+          reg_c[TN * res_idx_m + res_idx_n];
+    }
+  }
 }
 
-void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
+void kernel_body(int task_id, kernel_arg_t *__UNIFORM__ arg) {
-  const uint32_t dim_n = arg->dim_n;
+  // @perf: All threads are running these compute whose result is mostly same
-  int tid_x = task_id % dim_n;
+  // across the threadblock
-  int tid_y = task_id / dim_n;
 
+  const uint32_t threads_per_threadblock = (BM * BN) / (TM * TN);
+#ifdef RADIANCE
+  const uint32_t threadblocks_per_core = vx_num_threads() * vx_num_warps() /
+                                         threads_per_threadblock *
+                                         CORES_PER_CLUSTER;
+#else
+  const uint32_t threadblocks_per_core =
+      vx_num_threads() * vx_num_warps() / threads_per_threadblock;
+#endif
   const uint32_t threadblock_dim_x = vx_num_threads();
-  const uint32_t threadblock_dim_y = vx_num_warps();
+  const uint32_t threadblock_dim_y = vx_num_warps() / threadblocks_per_core;
-  const uint32_t threads_per_threadblock = threadblock_dim_x * threadblock_dim_y;
   const int threadblock_id = task_id / threads_per_threadblock;
+  const int threadblock_id_in_cluster = threadblock_id % threadblocks_per_core;
+  const int tid_in_threadblock = task_id % threads_per_threadblock;
 
-  const uint32_t dim_n_in_blocks = dim_n / threadblock_dim_x;
+  const uint32_t dim_m = arg->dim_m;
+  const uint32_t dim_n = arg->dim_n;
+  const uint32_t dim_n_in_blocks = dim_n / BN;
   const int threadblock_id_x = threadblock_id % dim_n_in_blocks;
   const int threadblock_id_y = threadblock_id / dim_n_in_blocks;
 
-  const int tid_in_threadblock_x = vx_thread_id();
+  // "static" shared memory allocation.  This would determine threadblock
-  const int tid_in_threadblock_y = vx_warp_id() % threadblock_dim_y;
+  // occupancy of a single cluster
-  thread_block_gemm(arg, tid_in_threadblock_x, tid_in_threadblock_y, threadblock_dim_x,
+  float *sharedmem_per_threadblock =
-                    threadblock_dim_y, threadblock_id_x, threadblock_id_y);
+      (float *)DEV_SMEM_START_ADDR + (2 * BM * BK) * threadblock_id_in_cluster;
+  thread_block_gemm(arg, tid_in_threadblock, threadblock_dim_x,
+                    threadblock_dim_y, threadblock_id_x, threadblock_id_y,
+                    threadblock_id_in_cluster, sharedmem_per_threadblock);
 }
 
 int main() {
   kernel_arg_t *arg = (kernel_arg_t *)KERNEL_ARG_DEV_MEM_ADDR;
-  const uint32_t grid_size = arg->dim_m * arg->dim_n;
+  const uint32_t grid_size = arg->dim_m * arg->dim_n / (TM * TN);
-  vx_spawn_tasks(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
+#ifdef RADIANCE
+  vx_spawn_tasks_cluster(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
+#else
+  // NOTE: This kernel assumes contiguous thread scheduling for efficient shared
+  // memory allocation, and therefore does not work with original vx_spawn_tasks
+  vx_spawn_tasks_contiguous(grid_size, (vx_spawn_tasks_cb)kernel_body, arg);
+#endif
   return 0;
 }

tests/regression/sgemm_wg/main.cpp

@@ -147,8 +147,8 @@ int main(int argc, char *argv[]) {
   RT_CHECK(vx_dev_open(&device));
 
   // FIXME: hardcoded
-  uint32_t dim_m = 16;
+  uint32_t dim_m = 32;
-  uint32_t dim_n = 16;
+  uint32_t dim_n = 32;
   uint32_t dim_k = 32;
 
   generate_source_matrix(dim_m, dim_n, dim_k);

tests/regression/vecaddx/common.h

@@ -1,7 +1,7 @@
 #ifndef _COMMON_H_
 #define _COMMON_H_
 
-#define KERNEL_ARG_DEV_MEM_ADDR 0x7ffff000
+#define KERNEL_ARG_DEV_MEM_ADDR 0x7fff0000
 
 #ifndef TYPE
 #define TYPE float

tests/regression/vecaddx/kernel.cpp

@@ -13,6 +13,10 @@ void kernel_body(int task_id, kernel_arg_t* __UNIFORM__ arg) {
 
 int main() {
 	kernel_arg_t* arg = (kernel_arg_t*)KERNEL_ARG_DEV_MEM_ADDR;
+#ifdef RADIANCE
+	vx_spawn_tasks_cluster(arg->num_points, (vx_spawn_tasks_cb)kernel_body, arg);
+#else
 	vx_spawn_tasks(arg->num_points, (vx_spawn_tasks_cb)kernel_body, arg);
+#endif
 	return 0;
 }

tests/regression/vecaddx/main.cpp

@@ -1,4 +1,5 @@
 #include <iostream>
+#include <fstream>
 #include <unistd.h>
 #include <string.h>
 #include <vector>
@@ -106,9 +107,9 @@ static void parse_args(int argc, char **argv) {
 
 void cleanup() {
   if (device) {    
-    vx_mem_free(device, kernel_arg.src0_addr);
+    // vx_mem_free(device, kernel_arg.src0_addr);
-    vx_mem_free(device, kernel_arg.src1_addr);
+    // vx_mem_free(device, kernel_arg.src1_addr);
-    vx_mem_free(device, kernel_arg.dst_addr);
+    // vx_mem_free(device, kernel_arg.dst_addr);
     vx_dev_close(device);
   }
 }
@@ -181,9 +182,12 @@ int main(int argc, char *argv[]) {
 
   // allocate device memory
   std::cout << "allocate device memory" << std::endl;
-  RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.src0_addr));
+  // RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.src0_addr));
-  RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.src1_addr));
+  // RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.src1_addr));
-  RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.dst_addr));
+  // RT_CHECK(vx_mem_alloc(device, buf_size, VX_MEM_TYPE_GLOBAL, &kernel_arg.dst_addr));
+  kernel_arg.src0_addr = 0x20000UL;
+  kernel_arg.src1_addr = 0x28000UL;
+  kernel_arg.dst_addr = 0xc0000000UL;
 
   kernel_arg.num_points = num_points;
 
@@ -201,10 +205,19 @@ int main(int argc, char *argv[]) {
   memcpy(staging_buf.data(), &kernel_arg, sizeof(kernel_arg_t));
   RT_CHECK(vx_copy_to_dev(device, KERNEL_ARG_DEV_MEM_ADDR, staging_buf.data(), sizeof(kernel_arg_t)));
 
+  std::ofstream file("args.bin", std::ios::binary | std::ios::out);
+  if (!file) {
+    std::cerr << "error: failed to open args.bin for writing\n";
+    exit(EXIT_FAILURE);
+  }
+  file.write(reinterpret_cast<char *>(staging_buf.data()), sizeof(kernel_arg_t));
+  file.close();
+
   // generate source data
   source_data.resize(2 * num_points);
   for (uint32_t i = 0; i < source_data.size(); ++i) {
-    source_data[i] = Comparator<TYPE>::generate();
+    // source_data[i] = Comparator<TYPE>::generate();
+    source_data[i] = static_cast<float>(i);
   }
 
   // upload source buffer0
@@ -215,6 +228,14 @@ int main(int argc, char *argv[]) {
       buf_ptr[i] = source_data[2 * i + 0];
     }
     RT_CHECK(vx_copy_to_dev(device, kernel_arg.src0_addr, staging_buf.data(), buf_size));
+
+    std::ofstream file("input.a.bin", std::ios::binary | std::ios::out);
+    if (!file) {
+      std::cerr << "error: failed to open input.a.bin for writing\n";
+      exit(EXIT_FAILURE);
+    }
+    file.write(reinterpret_cast<char *>(buf_ptr), buf_size);
+    file.close();
   }
 
   // upload source buffer1
@@ -225,6 +246,14 @@ int main(int argc, char *argv[]) {
       buf_ptr[i] = source_data[2 * i + 1];
     }   
     RT_CHECK(vx_copy_to_dev(device, kernel_arg.src1_addr, staging_buf.data(), buf_size));
+
+    std::ofstream file("input.b.bin", std::ios::binary | std::ios::out);
+    if (!file) {
+      std::cerr << "error: failed to open input.b.bin for writing\n";
+      exit(EXIT_FAILURE);
+    }
+    file.write(reinterpret_cast<char *>(buf_ptr), buf_size);
+    file.close();
   }
 
   // clear destination buffer