Stop recording to the tape while a kernel runs (tensorflow#980)

Some kernels might call higher-level ops to work around limitations. Because of this, we need to stop recording to the tape while the kernel runs.

Also use built-in inversesqrt in GLSL

src/engine.ts

@@ -103,16 +103,18 @@ export class Engine implements TensorManager {
     };
     const scopeName = this.activeScope.name;
 
+    // Stop recording to a tape when running a kernel.
+    this.customGradientDepth++;
     if (!ENV.get('DEBUG')) {
       result = forwardFunc(this.backend, saveFunc);
     } else {
       result = this.profiler.profileKernel(
           scopeName, () => forwardFunc(this.backend, saveFunc));
     }
+    // Continue recording after the kernel is done.
+    this.customGradientDepth--;
 
-    const recordKernel =
-        this.activeTape != null && this.customGradientDepth === 0;
-    if (recordKernel) {
+    if (this.shouldRecord()) {
       const tapeNode: TapeNode = {
         id: this.nextTapeNodeId++,
         name: scopeName,

src/kernels/webgl/batchnorm_gpu.ts

@@ -54,7 +54,7 @@ export class BatchNormProgram implements GPGPUProgram {
         float variance = getVarianceAtOutCoords();
         float offset = ${offsetSnippet};
         float scale = ${scaleSnippet};
-        float inv = scale / sqrt(variance + float(${varianceEpsilon}));
+        float inv = scale * inversesqrt(variance + float(${varianceEpsilon}));
         setOutput((x - mean) * inv + offset);
       }
     `;