Fix batch matching for batch mat mul (tensorflow#7062)

BUG * fix * lint * fix
Linchenn · Jan 9, 2023 · cf948c1 · cf948c1
1 parent 5b74d0a
commit cf948c1
Show file tree

Hide file tree

Showing 4 changed files with 24 additions and 15 deletions.
diff --git a/tfjs-backend-cpu/src/kernels/BatchMatMul.ts b/tfjs-backend-cpu/src/kernels/BatchMatMul.ts
@@ -106,12 +106,14 @@ export function batchMatMul(args: {
               let sum = 0.0;
 
               for (let k = k0; k < kBlock; k++) {
-                const batchOffsetA = Math.min(bi, batchDimA - 1) * aBatch;
-                const batchOffsetB = Math.min(bi, batchDimB - 1) * bBatch;
+                const batchIndexA = bi % batchDimA;
+                const batchIndexB = bi % batchDimB;
                 const aVal =
-                    a3dValues[batchOffsetA + i * aOuterStep + k * aInnerStep];
+                    // tslint:disable-next-line: max-line-length
+                    a3dValues[batchIndexA * aBatch + i * aOuterStep + k * aInnerStep];
                 const bVal =
-                    b3dValues[k * bInnerStep + j * bOuterStep + batchOffsetB];
+                    // tslint:disable-next-line: max-line-length
+                    b3dValues[k * bInnerStep + j * bOuterStep + batchIndexB * bBatch];
                 sum += aVal * bVal;
               }
               resVals[bi * size + (i * rightDim + j)] += sum;

diff --git a/tfjs-backend-wasm/src/cc/batch_mat_mul_impl.cc b/tfjs-backend-wasm/src/cc/batch_mat_mul_impl.cc
@@ -174,7 +174,7 @@ void slow_batch_matmul(const size_t a_id, const size_t* a_shape_ptr,
   const size_t shared_dim = transpose_a ? a_shape_ptr[1] : a_shape_ptr[2];
   const size_t left_dim = transpose_a ? a_shape_ptr[2] : a_shape_ptr[1];
   const size_t right_dim = transpose_b ? b_shape_ptr[1] : b_shape_ptr[2];
-  const size_t batch_dim = a_shape_ptr[0];
+  const size_t batch_dim = std::max(a_shape_ptr[0], b_shape_ptr[0]);
 
   std::vector<size_t> a_shape(a_shape_ptr, a_shape_ptr + a_shape_len);
   std::vector<size_t> b_shape(b_shape_ptr, b_shape_ptr + b_shape_len);
@@ -235,14 +235,12 @@ void slow_batch_matmul(const size_t a_id, const size_t* a_shape_ptr,
               float sum = 0.0;
 
               for (size_t k = k0; k < k_block; ++k) {
-                const size_t batch_offset_a =
-                    std::min(b, a_shape[0] - 1) * a_batch;
-                const size_t batch_offset_b =
-                    std::min(b, b_shape[0] - 1) * b_batch;
-                sum +=
-                    a_buf[batch_offset_a + i * a_outer_step +
-                          k * a_inner_step] *
-                    b_buf[k * b_inner_step + j * b_outer_step + batch_offset_b];
+                const size_t batch_index_a = b % a_shape[0];
+                const size_t batch_index_b = b % b_shape[0];
+                sum += a_buf[batch_index_a * a_batch + i * a_outer_step +
+                             k * a_inner_step] *
+                       b_buf[k * b_inner_step + j * b_outer_step +
+                             batch_index_b * b_batch];
               }
               size_t innermost_dim = i * right_dim + j;
               size_t out_buf_index = b * size + innermost_dim;

diff --git a/tfjs-backend-webgl/src/mulmat_packed_gpu.ts b/tfjs-backend-webgl/src/mulmat_packed_gpu.ts
@@ -78,9 +78,9 @@ export class MatMulPackedProgram implements GPGPUProgram {
     let batchASnippet = 'rc.x';
     let batchBSnippet = 'rc.x';
     if (aShape[0] < bShape[0]) {
-      batchASnippet = `int(min(float(rc.x), ${aShape[0] - 1}.))`;
+      batchASnippet = `imod(rc.x, ${aShape[0]})`;
     } else if (bShape[0] < aShape[0]) {
-      batchBSnippet = `int(min(float(rc.x), ${bShape[0] - 1}.))`;
+      batchBSnippet = `imod(rc.x, ${bShape[0]})`;
     }
 
     this.userCode = `

diff --git a/tfjs-core/src/ops/mat_mul_test.ts b/tfjs-core/src/ops/mat_mul_test.ts
@@ -894,6 +894,15 @@ describeWithFlags('matmulBatch', ALL_ENVS, () => {
     ]);
   });
 
+  it('A has more batch dimensions than B', async () => {
+    const a = tf.tensor4d(
+        [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16], [2, 2, 2, 2]);
+    const b = tf.tensor3d([1, 2, 3, 4], [2, 2, 1]);
+
+    const c = tf.matMul(a, b);
+    expectArraysClose(await c.data(), [5, 11, 39, 53, 29, 35, 95, 109]);
+  });
+
   it('batch dimensions do not match', () => {
     const a = tf.tensor3d(
         [