Implementation of the symmetrically quantized LSTM TFLite Op.

PiperOrigin-RevId: 199337082

1 files changed, 260 insertions, 2 deletions

diff --git a/tensorflow/contrib/lite/kernels/internal/kernel_utils.cc b/tensorflow/contrib/lite/kernels/internal/kernel_utils.cc
index 67e3810479..6e62183975 100644
--- a/tensorflow/contrib/lite/kernels/internal/kernel_utils.cc
+++ b/tensorflow/contrib/lite/kernels/internal/kernel_utils.cc
@@ -63,6 +63,8 @@ void RnnBatchStep(const float* input_ptr_batch, const int8_t* input_weights_ptr,
     // Quantize input from float to uint8 + quantization params (scaling
     // factor).
     float unused_min, unused_max;
+    // TODO(mirkov,raziel): replace this for-loop with a MACRO (or function)
+    // whichever is faster.
     for (int b = 0; b < batch_size; ++b) {
       const int offset = b * input_size;
       tensor_utils::SymmetricQuantizeFloats(
@@ -147,6 +149,7 @@ void LstmStep(
         input_to_input_weights_ptr, n_cell, n_input, input_ptr_batch, n_batch,
         input_gate_scratch, /*result_stride=*/1);
   }
+
   tensor_utils::MatrixBatchVectorMultiplyAccumulate(
       input_to_forget_weights_ptr, n_cell, n_input, input_ptr_batch, n_batch,
       forget_gate_scratch, /*result_stride=*/1);
@@ -161,8 +164,7 @@ void LstmStep(
   if (!use_cifg) {
     tensor_utils::MatrixBatchVectorMultiplyAccumulate(
         recurrent_to_input_weights_ptr, n_cell, n_output, output_state_ptr,
-        n_batch, input_gate_scratch,
-        /*result_stride=*/1);
+        n_batch, input_gate_scratch, /*result_stride=*/1);
   }
   tensor_utils::MatrixBatchVectorMultiplyAccumulate(
       recurrent_to_forget_weights_ptr, n_cell, n_output, output_state_ptr,
@@ -253,5 +255,261 @@ void LstmStep(
                            output_state_ptr);
 }
 
+// TODO(alanchiao): move this to tensor_utils.
+void VectorMultiply(const int8_t* vector, const int v_size, const float scale,
+                    float* result) {
+  for (int i = 0; i < v_size; ++i) {
+    *result++ = scale * *vector++;
+  }
+}
+
+void LstmStep(
+    const float* input_ptr_batch, const int8_t* input_to_input_weights_ptr,
+    float input_to_input_weights_scale,
+    const int8_t* input_to_forget_weights_ptr,
+    float input_to_forget_weights_scale,
+    const int8_t* input_to_cell_weights_ptr, float input_to_cell_weights_scale,
+    const int8_t* input_to_output_weights_ptr,
+    float input_to_output_weights_scale,
+    const int8_t* recurrent_to_input_weights_ptr,
+    float recurrent_to_input_weights_scale,
+    const int8_t* recurrent_to_forget_weights_ptr,
+    float recurrent_to_forget_weights_scale,
+    const int8_t* recurrent_to_cell_weights_ptr,
+    float recurrent_to_cell_weights_scale,
+    const int8_t* recurrent_to_output_weights_ptr,
+    float recurrent_to_output_weights_scale,
+    const int8_t* cell_to_input_weights_ptr, float cell_to_input_weights_scale,
+    const int8_t* cell_to_forget_weights_ptr,
+    float cell_to_forget_weights_scale,
+    const int8_t* cell_to_output_weights_ptr,
+    float cell_to_output_weights_scale, const float* input_gate_bias_ptr,
+    const float* forget_gate_bias_ptr, const float* cell_bias_ptr,
+    const float* output_gate_bias_ptr, const int8_t* projection_weights_ptr,
+    float projection_weights_scale, const float* projection_bias_ptr,
+    const TfLiteLSTMParams* params, int n_batch, int n_cell, int n_input,
+    int n_output, float* input_gate_scratch, float* forget_gate_scratch,
+    float* cell_scratch, float* output_gate_scratch, float* scaling_factors,
+    float* product_scaling_factors, float* recovered_cell_weights,
+    int8_t* quantized_input_ptr_batch, int8_t* quantized_output_state_ptr,
+    int8_t* quantized_cell_state_ptr, float* output_state_ptr,
+    float* cell_state_ptr, float* output_ptr_batch) {
+  // Since we have already checked that weights are all there or none, we can
+  // check the existense of only one to the get the condition.
+  const bool use_cifg = (input_to_input_weights_ptr == nullptr);
+  const bool use_peephole = (cell_to_output_weights_ptr != nullptr);
+  // Initialize scratch buffers with bias.
+  if (!use_cifg) {
+    tensor_utils::VectorBatchVectorAssign(input_gate_bias_ptr, n_cell, n_batch,
+                                          input_gate_scratch);
+  }
+  tensor_utils::VectorBatchVectorAssign(forget_gate_bias_ptr, n_cell, n_batch,
+                                        forget_gate_scratch);
+  tensor_utils::VectorBatchVectorAssign(cell_bias_ptr, n_cell, n_batch,
+                                        cell_scratch);
+  tensor_utils::VectorBatchVectorAssign(output_gate_bias_ptr, n_cell, n_batch,
+                                        output_gate_scratch);
+
+  if (!tensor_utils::IsZeroVector(input_ptr_batch, n_batch * n_input)) {
+    // Save quantization and matmul computation for all zero input.
+    float unused_min, unused_max;
+    for (int b = 0; b < n_batch; ++b) {
+      const int offset = b * n_input;
+      tensor_utils::SymmetricQuantizeFloats(
+          input_ptr_batch + offset, n_input, quantized_input_ptr_batch + offset,
+          &unused_min, &unused_max, &scaling_factors[b]);
+    }
+    // For each batch and cell: compute input_weight * input.
+    if (!use_cifg) {
+      for (int b = 0; b < n_batch; ++b) {
+        product_scaling_factors[b] =
+            scaling_factors[b] * input_to_input_weights_scale;
+      }
+      tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+          input_to_input_weights_ptr, n_cell, n_input,
+          quantized_input_ptr_batch, product_scaling_factors, n_batch,
+          input_gate_scratch, /*result_stride=*/1);
+    }
+
+    for (int b = 0; b < n_batch; ++b) {
+      product_scaling_factors[b] =
+          scaling_factors[b] * input_to_forget_weights_scale;
+    }
+    tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+        input_to_forget_weights_ptr, n_cell, n_input, quantized_input_ptr_batch,
+        product_scaling_factors, n_batch, forget_gate_scratch,
+        /*result_stride=*/1);
+
+    for (int b = 0; b < n_batch; ++b) {
+      product_scaling_factors[b] =
+          scaling_factors[b] * input_to_cell_weights_scale;
+    }
+    tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+        input_to_cell_weights_ptr, n_cell, n_input, quantized_input_ptr_batch,
+        product_scaling_factors, n_batch, cell_scratch, /*result_stride=*/1);
+
+    for (int b = 0; b < n_batch; ++b) {
+      product_scaling_factors[b] =
+          scaling_factors[b] * input_to_cell_weights_scale;
+    }
+    tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+        input_to_output_weights_ptr, n_cell, n_input, quantized_input_ptr_batch,
+        product_scaling_factors, n_batch, output_gate_scratch,
+        /*result_stride=*/1);
+  }
+
+  if (!tensor_utils::IsZeroVector(output_state_ptr, n_batch * n_output)) {
+    // Save quantization and matmul computation for all zero input.
+    float unused_min, unused_max;
+    for (int b = 0; b < n_batch; ++b) {
+      const int offset = b * n_output;
+      tensor_utils::SymmetricQuantizeFloats(output_state_ptr + offset, n_output,
+                                            quantized_output_state_ptr + offset,
+                                            &unused_min, &unused_max,
+                                            &scaling_factors[b]);
+    }
+    // For each batch and cell: compute recurrent_weight * output_state.
+    if (!use_cifg) {
+      for (int b = 0; b < n_batch; ++b) {
+        product_scaling_factors[b] =
+            scaling_factors[b] * recurrent_to_input_weights_scale;
+      }
+      tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+          recurrent_to_input_weights_ptr, n_cell, n_output,
+          quantized_output_state_ptr, product_scaling_factors, n_batch,
+          input_gate_scratch, /*result_stride=*/1);
+    }
+
+    for (int b = 0; b < n_batch; ++b) {
+      product_scaling_factors[b] =
+          scaling_factors[b] * recurrent_to_forget_weights_scale;
+    }
+    tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+        recurrent_to_forget_weights_ptr, n_cell, n_output,
+        quantized_output_state_ptr, product_scaling_factors, n_batch,
+        forget_gate_scratch, /*result_stride=*/1);
+
+    for (int b = 0; b < n_batch; ++b) {
+      product_scaling_factors[b] =
+          scaling_factors[b] * recurrent_to_cell_weights_scale;
+    }
+    tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+        recurrent_to_cell_weights_ptr, n_cell, n_output,
+        quantized_output_state_ptr, product_scaling_factors, n_batch,
+        cell_scratch, /*result_stride=*/1);
+
+    for (int b = 0; b < n_batch; ++b) {
+      product_scaling_factors[b] =
+          scaling_factors[b] * recurrent_to_output_weights_scale;
+    }
+    tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+        recurrent_to_output_weights_ptr, n_cell, n_output,
+        quantized_output_state_ptr, product_scaling_factors, n_batch,
+        output_gate_scratch, /*result_stride=*/1);
+  }
+
+  // Save quantization and matmul computation for all zero input.
+  const bool is_cell_state_all_zeros =
+      tensor_utils::IsZeroVector(cell_state_ptr, n_batch * n_cell);
+
+  // For each batch and cell: update input gate.
+  if (!use_cifg) {
+    if (use_peephole && !is_cell_state_all_zeros) {
+      VectorMultiply(cell_to_input_weights_ptr, n_cell,
+                     1. / cell_to_input_weights_scale, recovered_cell_weights);
+      tensor_utils::VectorBatchVectorCwiseProductAccumulate(
+          recovered_cell_weights, n_cell, cell_state_ptr, n_batch,
+          input_gate_scratch);
+    }
+    tensor_utils::ApplySigmoidToVector(input_gate_scratch, n_cell * n_batch,
+                                       input_gate_scratch);
+  }
+
+  // For each batch and cell: update forget gate.
+  if (use_peephole && !is_cell_state_all_zeros) {
+    VectorMultiply(cell_to_forget_weights_ptr, n_cell,
+                   1. / cell_to_forget_weights_scale, recovered_cell_weights);
+    tensor_utils::VectorBatchVectorCwiseProductAccumulate(
+        recovered_cell_weights, n_cell, cell_state_ptr, n_batch,
+        forget_gate_scratch);
+  }
+  tensor_utils::ApplySigmoidToVector(forget_gate_scratch, n_cell * n_batch,
+                                     forget_gate_scratch);
+
+  // For each batch and cell: update the cell.
+  tensor_utils::VectorVectorCwiseProduct(forget_gate_scratch, cell_state_ptr,
+                                         n_batch * n_cell, cell_state_ptr);
+  tensor_utils::ApplyActivationToVector(cell_scratch, n_batch * n_cell,
+                                        params->activation, cell_scratch);
+  if (use_cifg) {
+    tensor_utils::Sub1Vector(forget_gate_scratch, n_batch * n_cell,
+                             forget_gate_scratch);
+    tensor_utils::VectorVectorCwiseProductAccumulate(
+        cell_scratch, forget_gate_scratch, n_batch * n_cell, cell_state_ptr);
+  } else {
+    tensor_utils::VectorVectorCwiseProductAccumulate(
+        cell_scratch, input_gate_scratch, n_batch * n_cell, cell_state_ptr);
+  }
+  if (params->cell_clip > 0.0) {
+    tensor_utils::ClipVector(cell_state_ptr, n_batch * n_cell,
+                             params->cell_clip, cell_state_ptr);
+  }
+
+  // For each batch and cell: update the output gate.
+  if (use_peephole && !is_cell_state_all_zeros) {
+    VectorMultiply(cell_to_output_weights_ptr, n_cell,
+                   1. / cell_to_output_weights_scale, recovered_cell_weights);
+    tensor_utils::VectorBatchVectorCwiseProductAccumulate(
+        recovered_cell_weights, n_cell, cell_state_ptr, n_batch,
+        output_gate_scratch);
+  }
+  tensor_utils::ApplySigmoidToVector(output_gate_scratch, n_batch * n_cell,
+                                     output_gate_scratch);
+  tensor_utils::ApplyActivationToVector(cell_state_ptr, n_batch * n_cell,
+                                        params->activation, cell_scratch);
+  tensor_utils::VectorVectorCwiseProduct(output_gate_scratch, cell_scratch,
+                                         n_batch * n_cell, output_gate_scratch);
+
+  // For each batch: update the projection and output_state.
+  const bool use_projection_weight = (projection_weights_ptr != nullptr);
+  const bool use_projection_bias = (projection_bias_ptr != nullptr);
+  if (use_projection_weight) {
+    if (use_projection_bias) {
+      tensor_utils::VectorBatchVectorAssign(projection_bias_ptr, n_output,
+                                            n_batch, output_ptr_batch);
+    } else {
+      tensor_utils::ZeroVector(output_ptr_batch, n_batch * n_output);
+    }
+    if (!tensor_utils::IsZeroVector(output_gate_scratch, n_batch * n_cell)) {
+      // Save quantization and matmul computation for all zero input.
+      float unused_min, unused_max;
+      for (int b = 0; b < n_batch; ++b) {
+        const int offset = b * n_cell;
+        tensor_utils::SymmetricQuantizeFloats(
+            output_gate_scratch + offset, n_cell,
+            quantized_cell_state_ptr + offset, &unused_min, &unused_max,
+            &scaling_factors[b]);
+      }
+      for (int b = 0; b < n_batch; ++b) {
+        product_scaling_factors[b] =
+            scaling_factors[b] * projection_weights_scale;
+      }
+      tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+          projection_weights_ptr, n_output, n_cell, quantized_cell_state_ptr,
+          product_scaling_factors, n_batch, output_ptr_batch,
+          /*result_stride=*/1);
+    }
+    if (params->proj_clip > 0.0) {
+      tensor_utils::ClipVector(output_ptr_batch, n_batch * n_output,
+                               params->proj_clip, output_ptr_batch);
+    }
+  } else {
+    tensor_utils::CopyVector(output_gate_scratch, n_batch * n_output,
+                             output_ptr_batch);
+  }
+  tensor_utils::CopyVector(output_ptr_batch, n_batch * n_output,
+                           output_state_ptr);
+}
+
 }  // namespace kernel_utils
 }  // namespace tflite