[prelu_opt]: Optimizing RELU/LeakyRelu

  - Using mul_hi in case shift > 15.
  - Using unroll with factor 4.
  - Using Squash to one dim.
  - Split LeakyRelu and PRELU Code to two separate components.

Author: mfarag13

include/mli_config.h

@@ -72,7 +72,7 @@
 
 #if defined(__clang__)
 #define MLI_FORCE_INLINE inline __attribute__((always_inline))
-#define MLI_NO_INLINE __attribute__((noinline))
+#define MLI_NO_INLINE __attribute__((used)) __attribute__((noinline))
 #define MLI_CODE_SECTION_START(x) code(x) 
 #define MLI_CODE_SECTION_END()    code()
 

lib/src/kernels/transform/impl/mli_krn_leaky_relu_dsp.h

@@ -0,0 +1,92 @@
+/*
+* Copyright 2021, Synopsys, Inc.
+* All rights reserved.
+*
+* This source code is licensed under the BSD-3-Clause license found in
+* the LICENSE file in the root directory of this source tree.
+*
+*/
+
+#ifndef _MLI_KRN_LEAKY_RELU_DSP_H_
+#define _MLI_KRN_LEAKY_RELU_DSP_H_
+
+#include "mli_check.h"
+#include "mli_config.h"
+#include "mli_debug.h"
+#include "mli_helpers_api.h"
+#include "mli_math.h"
+#include "mli_prv_dsp.h"
+#include "mli_prv_tensor.h"
+#include "mli_types.h"
+
+namespace mli {
+namespace krn {
+namespace dsp {
+
+static MLI_FORCE_INLINE v2q15_t calc_leaky_relu(
+        const v2q15_t input,
+        const v2q15_t scale_v,
+        const int shift) {
+
+    /* out  = max(0, in) + alpha * min(0, in) */
+    v2q15_t zero = mli_prv_init_v<int16_t, v2q15_t>(0);
+    v2q15_t pos = mli_math_max_fx(zero, input);
+    v2q15_t neg = mli_math_acc_cast_fx<v2q15_t, v2accum40_t>(
+                  mli_math_mul_fx<v2q15_t, v2accum40_t>(scale_v, mli_math_min_fx(zero, input)), shift);
+    return mli_math_add_fx(pos, neg);
+}
+
+template <typename io_T>
+static MLI_FORCE_INLINE void compute_leaky_relu(
+        const MLI_PTR(io_T) vec_in,
+        MLI_OUT_PTR(io_T) vec_out,
+        const io_T scale,
+        const int shift) {
+
+    v2q15_t input = mli_prv_load_1vec(vec_in);
+    v2q15_t scale_v = mli_prv_init_v<io_T, v2q15_t>(scale);
+    mli_prv_store_n_samples(vec_out, calc_leaky_relu(input, scale_v, shift));
+}
+
+template <typename io_T>
+static MLI_FORCE_INLINE void compute_leaky_relu(
+        const MLI_PTR(io_T) vec_in,
+        MLI_OUT_PTR(io_T) vec_out,
+        const io_T scale,
+        const int shift,
+        const int remaining_part) {
+
+    MLI_ASSERT(remaining_part == 1);
+    v2q15_t input = mli_prv_load_1vec(vec_in);
+    v2q15_t scale_v = mli_prv_init_v<io_T, v2q15_t>(scale);
+    mli_prv_store_1_sample(vec_out, calc_leaky_relu(input, scale_v, shift));
+}
+
+static MLI_FORCE_INLINE void compute_leaky_relu(
+        const MLI_PTR(int8_t) vec_in,
+        MLI_OUT_PTR(int8_t) vec_out,
+        const int16_t in_zp,
+        const s8asym_quant_params *identity_params,
+        const s8asym_quant_params *alpha_params) {
+
+    mli::krn::ref::compute_leaky_relu(vec_in, vec_out, in_zp, identity_params, alpha_params);
+    mli::krn::ref::compute_leaky_relu(vec_in + 1, vec_out + 1, in_zp, identity_params, alpha_params);
+}
+
+static MLI_FORCE_INLINE void compute_leaky_relu(
+        const MLI_PTR(int8_t) vec_in,
+        MLI_OUT_PTR(int8_t) vec_out,
+        const int16_t in_zp,
+        const s8asym_quant_params *identity_params,
+        const s8asym_quant_params *alpha_params,
+        const int remaining_part) {
+
+    MLI_ASSERT(remaining_part == 1);
+    mli::krn::ref::compute_leaky_relu(vec_in, vec_out, in_zp, identity_params, alpha_params);
+}
+
+} // namespace dsp
+} // namespace krn
+} // namespace mli
+
+#endif // _MLI_KRN_LEAKY_RELU_DSP_H_