[codegen] [allocation] fixed issues in HMMA

include/triton/codegen/analysis/tiles.h

@@ -25,6 +25,15 @@ class axes;
 class layout;
 class align;
 
+enum layout_t {
+  SCANLINE,
+  HMMA_C,
+  HMMA_A_COL,
+  HMMA_A_ROW,
+  HMMA_B_COL,
+  HMMA_B_ROW
+};
+
 class tiles {
   typedef std::map<ir::value*, std::map<int, int>> param_map_t;
 private:
@@ -34,7 +43,7 @@ private:
 public:
   tiles(size_t num_warps, analysis::align* align, analysis::axes* axes, analysis::layout* layout);
   void run(ir::module &mod);
-  bool hmma(ir::value *value);
+  layout_t hmma(ir::value *value);
   int mts(ir::value *value, unsigned ax);
   int nts(ir::value *value, unsigned ax);
   int fpw(ir::value *value, unsigned ax);
@@ -52,7 +61,7 @@ private:
   // tile properties
   std::map<int, ir::value*> largest_;
   std::map<int, std::vector<int>> order_;
-  std::map<int, bool> hmma_;
+  std::map<int, layout_t> hmma_;
   std::map<int, int> fpw_;
   std::map<int, int> wpt_;
   std::map<int, int> mts_;

lib/codegen/analysis/allocation.cc

@@ -20,24 +20,14 @@ unsigned allocation::is_ld_padded(ir::value *x) {
     if(trans->get_perm()[0]->get_value() != 0)
       return 4;
   }
-  for(ir::user* user: x->get_users())
-    if(auto dot = dynamic_cast<ir::dot_inst*>(user)){
-      bool is_hmma = tiles_->hmma(user);
-      bool is_op_0 = x == dot->get_operand(0);
-      bool is_op_1 = x == dot->get_operand(1);
-      if(is_hmma && is_op_0){
-        if(dot->is_a_trans())
-          return 8;
-        else
-          return 16;
-      }
-      if(is_hmma && is_op_1){
-        if(!dot->is_b_trans())
-          return 8;
-        else
-          return 16;
-      }
-    }
+  if(tiles_->hmma(x) == HMMA_A_ROW)
+    return 8;
+  if(tiles_->hmma(x) == HMMA_A_COL)
+    return 16;
+  if(tiles_->hmma(x) == HMMA_B_COL)
+    return 8;
+  if(tiles_->hmma(x) == HMMA_B_ROW)
+    return 16;
   if(auto* phi = dynamic_cast<ir::phi_node*>(x)) {
     unsigned result = 0;
     for(unsigned i = 0; i < phi->get_num_incoming(); i++)

lib/codegen/analysis/tiles.cc

@@ -23,7 +23,7 @@ tiles::tiles(size_t num_warps, analysis::align *align, analysis::axes *axes, ana
     num_warps_(num_warps), align_(align), axes_(axes), layout_(layout)
 { }
 
-bool is_hmma(ir::value *v){
+bool is_hmma_c(ir::value *v){
   bool result = false;
   if(auto *x = dynamic_cast<ir::dot_inst*>(v)){
     ir::value *a = x->get_operand(0);
@@ -36,9 +36,44 @@ bool is_hmma(ir::value *v){
   return result;
 }
 
+bool is_hmma_a_col(ir::value* v) {
+  for(ir::user *u: v->get_users())
+    if(is_hmma_c(u)){
+      ir::dot_inst* dot = (ir::dot_inst*)u;
+      if((v == dot->get_operand(0)) && !dot->is_a_trans())
+        return true;
+    }
+}
+
+bool is_hmma_a_row(ir::value* v) {
+  for(ir::user *u: v->get_users())
+    if(is_hmma_c(u)){
+      ir::dot_inst* dot = (ir::dot_inst*)u;
+      if((v == dot->get_operand(0)) && dot->is_a_trans())
+        return true;
+    }
+}
+
+bool is_hmma_b_col(ir::value* v) {
+  for(ir::user *u: v->get_users())
+    if(is_hmma_c(u)){
+      ir::dot_inst* dot = (ir::dot_inst*)u;
+      if((v == dot->get_operand(1)) && !dot->is_b_trans())
+        return true;
+    }
+}
+
+bool is_hmma_b_row(ir::value* v) {
+  for(ir::user *u: v->get_users())
+    if(is_hmma_c(u)){
+      ir::dot_inst* dot = (ir::dot_inst*)u;
+      if((v == dot->get_operand(1)) && dot->is_b_trans())
+        return true;
+    }
+}
 
 
-bool tiles::hmma(ir::value *value) {
+layout_t tiles::hmma(ir::value *value) {
   return hmma_.at(layout_->id(value));
 }
 
@@ -164,7 +199,18 @@ void tiles::run(ir::module &) {
   // find out which groups require hmma layout
   for(size_t i = 0; i < num_groups; i++) {
     const auto& values = layout_->values(i);
-    hmma_[i] = std::any_of(values.begin(), values.end(), &is_hmma);
+    bool hmma_c = std::any_of(values.begin(), values.end(), &is_hmma_c);
+    bool hmma_a_col = std::any_of(values.begin(), values.end(), &is_hmma_a_col);
+    bool hmma_a_row = std::any_of(values.begin(), values.end(), &is_hmma_a_row);
+    bool hmma_b_col = std::any_of(values.begin(), values.end(), &is_hmma_b_col);
+    bool hmma_b_row = std::any_of(values.begin(), values.end(), &is_hmma_b_row);
+    if(hmma_c)          hmma_[i] = HMMA_C;
+    else if(hmma_a_col) hmma_[i] = HMMA_A_COL;
+    else if(hmma_a_row) hmma_[i] = HMMA_A_ROW;
+    else if(hmma_b_col) hmma_[i] = HMMA_B_COL;
+    else if(hmma_b_row) hmma_[i] = HMMA_B_ROW;
+    else                hmma_[i] = SCANLINE;
+
   }
   // find out which value is the largest in each group
   for(size_t i = 0; i < num_groups; i++) {
@@ -197,7 +243,7 @@ void tiles::run(ir::module &) {
     if(!i->get_type()->is_tile_ty())
       continue;
     /* HMMA parameters*/
-    if(hmma_[x.first])
+    if(hmma_[x.first] == HMMA_C)
       init_hmma_tile(i);
     else
       init_scanline_tile(i);

lib/codegen/selection.cc

@@ -710,7 +710,7 @@ void selection::init_hmma_axes(ir::value *v, IRBuilder<> &builder, Value *u_thre
 
 
 void selection::init_axes(ir::value *v, IRBuilder<> &builder, Value *u_thread_id, Value *u_warp_id) {
-  if(tiles_->hmma(v))
+  if(tiles_->hmma(v) == analysis::HMMA_C)
     init_hmma_axes(v, builder, u_thread_id, u_warp_id);
   else
     init_strided_scan_axes(v, builder, u_thread_id, u_warp_id);
@@ -1241,11 +1241,10 @@ void selection::lower_dot(ir::dot_inst *dot, LLVMContext &ctx, Function *fn, IRB
   if(NK != 1) {
     shared_tile *TA = (shared_tile*)tmap_.at(A);
     shared_tile *TB = (shared_tile*)tmap_.at(B);
-    if(tiles_->hmma(dot))
+    if(tiles_->hmma(dot) == analysis::HMMA_C)
       lower_hmma_dot(dot, ctx, fn, builder, TC, TA, TB, TD, NK);
     else
       lower_scanline_dot(dot, ctx, fn, builder, TC, TA, TB, TD, NK, c_ty, f_mul_add);
-
   }
   else {
     distributed_tile *TA = (distributed_tile*)tmap_.at(A);

lib/codegen/transform/peephole.cc

@@ -104,12 +104,12 @@ bool peephole::rewrite_dot_hmma(ir::dot_inst *dot, ir::builder& builder, bool tr
     BB = ((ir::trans_inst*)B)->get_operand(0);
   }
   else{
-    if(auto *T = dynamic_cast<ir::trans_inst*>(A)){
+    if(auto *T = dynamic_cast<ir::trans_inst*>(B)){
       std::vector<ir::constant_int*> perm(T->get_perm());
       std::swap(perm[0], perm[1]);
-      AA = builder.create_trans(T->get_operand(0), perm);
-      T->replace_all_uses_with(AA);
-      trans_a = true;
+      BB = builder.create_trans(T->get_operand(0), perm);
+      T->replace_all_uses_with(BB);
+      trans_b = true;
     }
   }
   if(!trans_a && !trans_b)

lib/driver/module.cc

@@ -241,7 +241,6 @@ std::string cu_module::compile_llvm_module(std::unique_ptr<llvm::Module> module,
 cu_module::cu_module(driver::context * context, std::unique_ptr<llvm::Module> ll_module): cu_module(context, compile_llvm_module(std::move(ll_module), context->device())) { }
 
 cu_module::cu_module(driver::context * context, std::string const & source) : module(context, CUmodule(), true), source_(source){
-//  std::cout << source << std::endl;
   cu_context::context_switcher ctx_switch(*context);
   // JIT compile source-code
   CUjit_option opt[] = {CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES, CU_JIT_ERROR_LOG_BUFFER};

lib/runtime/function.cc

@@ -201,17 +201,17 @@ std::unique_ptr<driver::module> function::make_bin(ir::module &module, driver::c
   // create passes
   codegen::transform::cts cts;
   codegen::analysis::align align;
-  codegen::analysis::liveness shmem_liveness;
+  codegen::analysis::liveness liveness;
   codegen::analysis::axes axes;
   codegen::analysis::layout layouts(&axes);
   codegen::transform::coalesce coalesce(&align, &layouts);
   codegen::analysis::tiles tiles(opt.num_warps, &align, &axes, &layouts);
-  codegen::analysis::allocation shmem_allocation(&shmem_liveness, &tiles);
-  codegen::transform::membar shmem_barriers(&shmem_liveness, &shmem_allocation);
+  codegen::analysis::allocation allocation(&liveness, &tiles);
+  codegen::transform::membar barriers(&liveness, &allocation);
   codegen::transform::dce dce;
   codegen::transform::peephole peephole;
   codegen::transform::reassociate reassociate(&align);
-  codegen::selection selection(&shmem_liveness, &shmem_allocation, &tiles, &align, &axes, &layouts, &coalesce, target.get(), opt.num_warps);
+  codegen::selection selection(&liveness, &allocation, &tiles, &align, &axes, &layouts, &coalesce, target.get(), opt.num_warps);
   // run passes
   peephole.run(module);
   dce.run(module);
@@ -226,14 +226,12 @@ std::unique_ptr<driver::module> function::make_bin(ir::module &module, driver::c
   tiles.run(module);
   reassociate.run(module);
   dce.run(module);
-  peephole.run(module);
-  dce.run(module);
   cts.run(module);
-  shmem_liveness.run(module);
-  shmem_allocation.run(module);
-  if(shmem_allocation.allocated_size() > context->device()->max_shared_memory())
+  liveness.run(module);
+  allocation.run(module);
+  if(allocation.allocated_size() > context->device()->max_shared_memory())
     return std::unique_ptr<driver::module>();
-  shmem_barriers.run(module);
+  barriers.run(module);
   dce.run(module);
   dce.run(module);
   axes.run(module);

tests/bench/dot.cc

@@ -27,9 +27,9 @@ inline rt::function::grid_fn_ty grid2d(size_t M, size_t N) {
 
 
 std::vector<double> do_bench(drv::stream* stream, bool AT, bool BT, int32_t M, int32_t N, int32_t K){
-  typedef float NumericT;
-  std::string ty = "float";
-  cublasDataType_t cuty = CUDA_R_32F;
+  typedef half_float::half NumericT;
+  std::string ty = "half";
+  cublasDataType_t cuty = CUDA_R_16F;
   size_t dt_nbytes = sizeof(NumericT);
   drv::context* context = stream->context();
   // leading dimensions
@@ -47,8 +47,8 @@ std::vector<double> do_bench(drv::stream* stream, bool AT, bool BT, int32_t M, i
   opt.defines.push_back({"BT", {BT?"1":"0"}});
   opt.defines.push_back({"TM", {"128"}});
   opt.defines.push_back({"TN", {"128"}});
-  opt.defines.push_back({"TK", {"8"}});
-  opt.num_warps = {8};
+  opt.defines.push_back({"TK", {"16"}});
+  opt.num_warps = {2, 4, 8};
   // create function
   rt::function function(src::dot, opt);
   // benchmark available libraries
@@ -79,7 +79,10 @@ int main() {
   // shapes to benchmark
   typedef std::tuple<bool, bool, int, int, int> config_t;
   std::vector<config_t> configs;
-  for(auto x: std::vector<std::array<bool, 2>>{{false, true}}){
+  for(auto x: std::vector<std::array<bool, 2>>{{false, false},
+                                               {false, true},
+                                               {true, false},
+                                               {true, true}}){
     std::vector<config_t> tmp = {
       config_t{x[0], x[1], 2048, 2048, 2048}
 //      config_t{x[0], x[1], 16, 2048, 2048},

tests/common/src/dot.h

@@ -64,7 +64,7 @@ void dot(TYPE * A, TYPE * B, TYPE * C,
   // epilogue
   int rxc[TM] =  ridx * TM + 0 ... TM;
   int ryc[TN] =  ridy * TN + 0 ... TN;
-  TYPE* pc[TM, TN] = C + rxc[:, newaxis] * ldc + ryc[newaxis, :];
+  TYPE* pc[TM, TN] = C + rxc[:, newaxis] + ryc[newaxis, :] * ldc;
   *pc = c;
 }
 )";

tests/unit/dot.cc

@@ -32,7 +32,7 @@ static void cpu_ref(std::vector<T> &c, const std::vector<T> &a, const std::vecto
     float acc = 0;
     for(size_t k = 0; k < K; k++)
       acc = acc + (AT ? a[k + m*K] : a[m + k*M]) * (BT ? b[n + k*N] : b[k + n*K]);
-    c[m*N + n] = static_cast<T>(acc);
+    c[m + n*M] = static_cast<T>(acc);
   }
 }