From ff858d35ed55680bc2ff53e83d6f16cdb53fa261 Mon Sep 17 00:00:00 2001
From: tensor-tang <tangjian03@baidu.com>
Date: Fri, 14 Sep 2018 14:21:03 +0800
Subject: [PATCH] fix bug and enable on batch mode as well

---
 paddle/fluid/operators/fusion_lstm_op.cc      | 13 +++++-
 .../fluid/operators/math/cpu_lstm_compute.cc  | 44 +------------------
 .../fluid/operators/math/cpu_lstm_compute.h   | 39 +++++++++++++++-
 3 files changed, 52 insertions(+), 44 deletions(-)

diff --git a/paddle/fluid/operators/fusion_lstm_op.cc b/paddle/fluid/operators/fusion_lstm_op.cc
index a23704621cd..8ca79d20ec4 100644
--- a/paddle/fluid/operators/fusion_lstm_op.cc
+++ b/paddle/fluid/operators/fusion_lstm_op.cc
@@ -543,12 +543,23 @@ class FuisonLSTMKernel : public framework::OpKernel<T> {
         MOVE_ONE_STEP;
       }
     } else {
+      // TODO(TJ): unly workaround, clean me
+      std::function<void(T*, const T*, T*, T*)> compute_ctht;
+      if (platform::jit::MayIUse(platform::jit::avx) &&
+          act_gate_str == "sigmoid" && act_cand_str == "tanh" &&
+          act_cell_str == "tanh" && D == 8) {
+        compute_ctht = math::lstm_compute_ctht<T>;
+      } else {
+        compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) {
+          COMPUTE_CtHt(gates, ct_1, ct, ht);
+        };
+      }
       for (int step = tstart; step < max_seq_len; ++step) {
         const int cur_bs = batch_starts[step + 1] - batch_starts[step];
         GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data);
         DEFINE_CUR;
         for (int i = 0; i < cur_bs; ++i) {
-          COMPUTE_CtHt(cur_in_data, cur_prev_c_data, cur_c_out_data,
+          compute_ctht(cur_in_data, cur_prev_c_data, cur_c_out_data,
                        cur_h_out_data);
           MOVE_ONE_BATCH;
         }
diff --git a/paddle/fluid/operators/math/cpu_lstm_compute.cc b/paddle/fluid/operators/math/cpu_lstm_compute.cc
index 83094d01cf6..84e387c974b 100644
--- a/paddle/fluid/operators/math/cpu_lstm_compute.cc
+++ b/paddle/fluid/operators/math/cpu_lstm_compute.cc
@@ -13,49 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/cpu_lstm_compute.h"
-#ifdef __AVX__
-#include <immintrin.h>
-#endif
+
 namespace paddle {
 namespace operators {
-namespace math {
-
-#ifdef __AVX__
-// TODO(TJ): ugly workaround, clean me
-
-namespace detail {
-namespace forward {
-namespace avx {
-__m256 Sigmoid(const __m256 a);
-__m256 Tanh(const __m256 a);
-}  // namespace avx
-}  // namespace forward
-}  // namespace detail
-
-template <>
-void lstm_compute_ctht<float>(float* gates, const float* ct_1, float* ct,
-                              float* ht) {
-  namespace act = detail::forward::avx;
-  // gates: W_ch, W_ih, W_fh, W_oh
-  __m256 c, i, f, o;
-  c = _mm256_loadu_ps(gates);
-  i = _mm256_loadu_ps(gates + 8);
-  f = _mm256_loadu_ps(gates + 16);
-  o = _mm256_loadu_ps(gates + 24);
-
-  /* C_t = C_t-1 * fgated + cand_gated * igated*/
-  c = _mm256_mul_ps(act::Tanh(c), act::Sigmoid(i));
-  i = _mm256_loadu_ps(ct_1);
-  f = _mm256_mul_ps(i, act::Sigmoid(f));
-  f = _mm256_add_ps(c, f);
-  _mm256_storeu_ps(ct, f);
-
-  /* H_t = act_cell(C_t) * ogated */
-  o = _mm256_mul_ps(act::Tanh(f), act::Sigmoid(o));
-  _mm256_storeu_ps(ht, o);
-}
-#endif
-
-}  // namespace math
+namespace math {}  // namespace math
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/fluid/operators/math/cpu_lstm_compute.h b/paddle/fluid/operators/math/cpu_lstm_compute.h
index fe6c01b7d93..00e9e4f32ae 100644
--- a/paddle/fluid/operators/math/cpu_lstm_compute.h
+++ b/paddle/fluid/operators/math/cpu_lstm_compute.h
@@ -16,6 +16,9 @@ limitations under the License. */
 #include <string>
 #include "paddle/fluid/operators/math/cpu_vec.h"
 #include "paddle/fluid/platform/cpu_info.h"
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
 
 namespace paddle {
 namespace operators {
@@ -35,13 +38,47 @@ void lstm_compute_ctht(T* gates, const T* ct_1, T* ct, T* ht) {
     ct[d] = ct_1[d] * f[d] + gates[d] * i[d];
     // H_t = act_cell(C_t) * ogated
     T tmp = ct[d] * 2;
-    tmp = static_cast<T>(0) - (tmp < min) ? min : ((tmp > max) ? max : tmp);
+    tmp = static_cast<T>(0) - ((tmp < min) ? min : ((tmp > max) ? max : tmp));
     vec_exp<T>(1, &tmp, &tmp);
     tmp = static_cast<T>(2) / (static_cast<T>(1) + tmp) - static_cast<T>(1);
     ht[d] = tmp * o[d];
   }
 }
 
+#ifdef __AVX__
+namespace detail {
+namespace forward {
+namespace avx {
+__m256 Sigmoid(const __m256 a);
+__m256 Tanh(const __m256 a);
+}  // namespace avx
+}  // namespace forward
+}  // namespace detail
+
+template <>
+void lstm_compute_ctht<float>(float* gates, const float* ct_1, float* ct,
+                              float* ht) {
+  namespace act = detail::forward::avx;
+  // gates: W_ch, W_ih, W_fh, W_oh
+  __m256 c, i, f, o;
+  c = _mm256_loadu_ps(gates);
+  i = _mm256_loadu_ps(gates + 8);
+  f = _mm256_loadu_ps(gates + 16);
+  o = _mm256_loadu_ps(gates + 24);
+
+  /* C_t = C_t-1 * fgated + cand_gated * igated*/
+  c = _mm256_mul_ps(act::Tanh(c), act::Sigmoid(i));
+  i = _mm256_loadu_ps(ct_1);
+  f = _mm256_mul_ps(i, act::Sigmoid(f));
+  f = _mm256_add_ps(c, f);
+  _mm256_storeu_ps(ct, f);
+
+  /* H_t = act_cell(C_t) * ogated */
+  o = _mm256_mul_ps(act::Tanh(f), act::Sigmoid(o));
+  _mm256_storeu_ps(ht, o);
+}
+#endif
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle
-- 
GitLab