Formatting

src/onnx/parse_attention.cpp

@@ -162,8 +162,8 @@ struct parse_attention : op_parser<parse_attention>
         auto gemm4 = info.add_instruction(migraphx::make_op("dot"), softmax, v_t);
 
         // scratch3 is BxNxSxH, transpose to output BxSxNxH
-        gemm4 = info.add_instruction(
-            migraphx::make_op("transposectx", {{"head_size", head_size}}), gemm4);
+        gemm4 = info.add_instruction(migraphx::make_op("transposectx", {{"head_size", head_size}}),
+                                     gemm4);
         gemm4 = info.add_instruction(
             make_op("reshape", {{"dims", {batch_size, sequence_length, num_heads * head_size}}}),
             info.make_contiguous(gemm4));

src/targets/gpu/jit/bert_transpose.cpp

@@ -89,7 +89,6 @@ __global__ void transposeqkv_kernel(void* input_p, void* output_p)
 } // namespace migraphx
 )__migraphx__";
 
-
 struct transposeqkv_compiler : compiler<transposeqkv_compiler>
 {
     std::vector<std::string> names() const { return {"transposeqkv"}; }
@@ -121,7 +120,6 @@ struct transposeqkv_compiler : compiler<transposeqkv_compiler>
     }
 };
 
-
 } // namespace gpu
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/targets/gpu/kernels/include/migraphx/kernels/transposectx.hpp

@@ -40,17 +40,20 @@ __device__ void transposectx(const T& input_t, const U& output_t, Settings st)
     const int in_offset = s * head_size + n * sequence_length * head_size + b * NHS;
 
     int out_offset = 0;
-    if (reversed_bs) {
+    if(reversed_bs)
+    {
         const int batch_size = gridDim.y;
         const int BNH        = NH * batch_size;
         out_offset           = n * head_size + b * NH + s * BNH;
-    } else {
+    }
+    else
+    {
         out_offset = n * head_size + s * NH + b * NHS;
-
     }
 
     const int i = threadIdx.x;
-    if (i < head_size) {
+    if(i < head_size)
+    {
         output_t[out_offset + i] = input_t[in_offset + i];
     }
 }

src/targets/gpu/kernels/include/migraphx/kernels/transposeqkv.hpp

@@ -43,16 +43,20 @@ __device__ void transposeqkv(const T& input_t, const U& output_t, Settings st)
     const int NHS             = NH * sequence_length;
 
     int in_offset = 0;
-    if (reversed_bs) {
+    if(reversed_bs)
+    {
         const int BNH = NH * batch_size;
         in_offset     = n * H + (m + b * chunk_num) * NH + s * BNH * chunk_num;
-    } else {
+    }
+    else
+    {
         in_offset = n * H + (m + s * chunk_num) * NH + b * NHS * chunk_num;
     }
     const int out_offset = s * H + n * sequence_length * H + b * NHS + m * NHS * batch_size;
 
     const int i = threadIdx.x;
-    if (i < H) {
+    if(i < H)
+    {
         output_t[out_offset + i] = input_t[in_offset + i];
     }
 }

test/verify/0transposectx_test.cpp

@@ -10,7 +10,8 @@ struct test_transposectx : verify_program<test_transposectx>
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        auto x = mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {1, 12, 128, 64}});
+        auto x =
+            mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {1, 12, 128, 64}});
         mm->add_instruction(migraphx::make_op("transposectx", {{"head_size", 64}}), x);
         p.debug_print();
         return p;

test/verify/0transposeqkv_test.cpp

@@ -9,7 +9,8 @@ struct test_transposeqkv : verify_program<test_transposeqkv>
     {
         migraphx::program p;
         auto* mm = p.get_main_module();
-        auto x = mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {1, 1, 3, 12, 64}});
+        auto x =
+            mm->add_parameter("x", migraphx::shape{migraphx::shape::float_type, {1, 1, 3, 12, 64}});
         mm->add_instruction(migraphx::make_op("transposeqkv", {{"head_size", 64}}), x);
         p.debug_print();
         return p;