[Codegen][CUDA] Fix make_int4x cuda codegen vectorize (apache#8137)

Co-authored-by: wangyucheng <wangyucheng@sensetime.com>
trevor-m · Jun 17, 2021 · 8117f3f · 8117f3f
1 parent 371c2f8
commit 8117f3f
Show file tree

Hide file tree

Showing 3 changed files with 55 additions and 13 deletions.
diff --git a/src/target/source/codegen_c.cc b/src/target/source/codegen_c.cc
@@ -212,13 +212,18 @@ std::string CodeGenC::GetBufferRef(DataType t, const VarNode* buffer, PrimExpr i
       PrintType(t.element_of(), os);
       os << "*)";
     }
-    os << vid << " + (";
-    PrintExpr(index, os);
-    os << ")";
     if (t.bits() == 4 || (t.bits() == 1 && t.is_int())) {
-      os << " / " << (32 / t.bits());
+      os << vid << ") + (";
+      PrintExpr(index, os);
+      os << ")";
+      os << " / " << t.lanes();
+      os << ")[0]";
+    } else {
+      os << vid << " + (";
+      PrintExpr(index, os);
+      os << ")";
+      os << "))[0]";
     }
-    os << "))[0]";
   }
   return os.str();
 }

diff --git a/src/target/source/codegen_cuda.cc b/src/target/source/codegen_cuda.cc
@@ -809,18 +809,48 @@ void CodeGenCUDA::VisitExpr_(const BroadcastNode* op, std::ostream& os) {  // NO
     return;
   }
 
-  if ((op->dtype.is_int() || op->dtype.is_uint()) && op->dtype.bits() == 4 && op->lanes == 8) {
-    // make_int4x8
+  if ((op->dtype.is_int() || op->dtype.is_uint()) && op->dtype.bits() == 4) {
+    bool fail = false;
     const int64_t* p = as_const_int(op->value);
     ICHECK(p);
     int64_t v = *p & 0xF;
-    v = (v << 28) | (v << 24) | (v << 20) | (v << 16) | (v << 12) | (v << 8) | (v << 4) | v;
-    if (op->dtype.is_uint()) {
-      os << "(uint)" << v;
+
+    if (op->lanes == 4) {
+      v = (v << 12) | (v << 8) | (v << 4) | v;
+      if (op->dtype.is_uint()) {
+        os << "(uint16_t)" << v;
+      } else {
+        os << "(int16_t)" << v;
+      }
     } else {
-      os << "(int)" << v;
+      v = (v << 28) | (v << 24) | (v << 20) | (v << 16) | (v << 12) | (v << 8) | (v << 4) | v;
+      if (op->lanes == 8) {
+        if (op->dtype.is_uint()) {
+          os << "(uint)" << v;
+        } else {
+          os << "(int)" << v;
+        }
+      } else if (op->lanes == 16 || op->lanes == 32) {
+        os << "make_";
+        PrintType(op->dtype, os);
+        os << '(';
+        for (int i = 0; i < op->lanes / 8; ++i) {
+          if (i != 0) os << ", ";
+          if (op->dtype.is_uint()) {
+            os << "(uint)" << v;
+          } else {
+            os << "(int)" << v;
+          }
+        }
+        os << ')';
+      } else {
+        fail = true;
+      }
+    }
+
+    if (!fail) {
+      return;
     }
-    return;
   }
 
   std::string v = PrintExpr(op->value);

diff --git a/tests/python/unittest/test_target_codegen_cuda.py b/tests/python/unittest/test_target_codegen_cuda.py
@@ -215,14 +215,21 @@ def check_cuda(n, value, lanes):
         y, x = s[A].op.axis
         s[A].vectorize(x)
         s[A].bind(y, bx)
-        fun = tvm.build(s, [A], "cuda", name="make_int4x8")
+        kernel_name = "make_int4x" + str(lanes)
+        fun = tvm.build(s, [A], "cuda", name=kernel_name)
         np_a = np.full((n, lanes), value, dtype="int8")
         a = tvm.nd.empty((n, lanes), dtype, dev)
         fun(a)
         np.testing.assert_equal(a.numpy(), np_a)
 
+    check_cuda(64, 1, 4)
+    check_cuda(64, 7, 4)
     check_cuda(64, 1, 8)
     check_cuda(64, 7, 8)
+    check_cuda(64, 1, 16)
+    check_cuda(64, 7, 16)
+    check_cuda(64, 1, 32)
+    check_cuda(64, 7, 32)
 
 
 @tvm.testing.requires_gpu