COMPMID-464 Implement Depthwise convolution 3x3 on NEON

Change-Id: Ie4e1803a52afac6b6c597c6e551729dad2347cd1
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/92607
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>

1 files changed, 2 insertions, 371 deletions

diff --git a/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
index 60a3a1b636..bf23c4b2d2 100644
--- a/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDirectConvolutionLayerKernel.cpp
@@ -22,6 +22,7 @@
  * SOFTWARE.
  */
 #include "arm_compute/core/NEON/kernels/NEDirectConvolutionLayerKernel.h"
+#include "arm_compute/core/NEON/kernels/convolution/NEDirectConvolutionDetail.h"
 
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
@@ -37,6 +38,7 @@
 #include <arm_neon.h>
 
 using namespace arm_compute;
+using namespace arm_compute::detail;
 
 namespace
 {
@@ -428,122 +430,6 @@ public:
 };
 
 #ifdef ARM_COMPUTE_AARCH64_V8_2
-inline float16x8x3_t load_matrix_row(const float16_t *ptr)
-{
-    /* ptr is a pointer to a row in a 3x3 matrix, the function returns 3 vectors holding exactly the same value in all lanes:
-       r.val[0] contains the first element, r.val[1] the second element and r.val[2] the third element (in all lanes) */
-    const float16x8x3_t r =
-    {
-        {
-            vld1q_dup_f16(ptr),
-            vld1q_dup_f16(1 + ptr),
-            vld1q_dup_f16(2 + ptr)
-        }
-    };
-    return r;
-}
-
-template <unsigned int stridex>
-float16x8x2_t convolve_3x3(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
-                           int fixed_point_position);
-
-template <>
-float16x8x2_t convolve_3x3<1>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
-                              int fixed_point_position)
-{
-    ARM_COMPUTE_UNUSED(fixed_point_position);
-
-    const float16x8x3_t vtop =
-    {
-        {
-            vld1q_f16(in_top),
-            vld1q_f16(in_top + 8),
-            vld1q_f16(in_top + 16)
-        }
-    };
-    const float16x8x3_t vmid =
-    {
-        {
-            vld1q_f16(in_mid),
-            vld1q_f16(in_mid + 8),
-            vld1q_f16(in_mid + 16)
-        }
-    };
-    const float16x8x3_t vlow =
-    {
-        {
-            vld1q_f16(in_low),
-            vld1q_f16(in_low + 8),
-            vld1q_f16(in_low + 16)
-        }
-    };
-    float16x8x2_t out =
-    {
-        {
-            vmulq_f16(vtop.val[0], m0.val[0]),
-            vmulq_f16(vtop.val[1], m0.val[0])
-        }
-    };
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vtop.val[0], vtop.val[1], 1), m0.val[1]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vtop.val[0], vtop.val[1], 2), m0.val[2]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vmid.val[0], m1.val[0]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vmid.val[0], vmid.val[1], 1), m1.val[1]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vmid.val[0], vmid.val[1], 2), m1.val[2]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vlow.val[0], m2.val[0]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vlow.val[0], vlow.val[1], 1), m2.val[1]));
-    out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vlow.val[0], vlow.val[1], 2), m2.val[2]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vtop.val[1], vtop.val[2], 1), m0.val[1]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vtop.val[1], vtop.val[2], 2), m0.val[2]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vmid.val[1], m1.val[0]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vmid.val[1], vmid.val[2], 1), m1.val[1]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vmid.val[1], vmid.val[2], 2), m1.val[2]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vlow.val[1], m2.val[0]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vlow.val[1], vlow.val[2], 1), m2.val[1]));
-    out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vlow.val[1], vlow.val[2], 2), m2.val[2]));
-    return out;
-}
-
-template <>
-inline float16x8x2_t convolve_3x3<2>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
-                                     int fixed_point_position)
-{
-    float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 2), out.val[0], 1);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 0), out.val[0], 2);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 2), out.val[0], 3);
-    return out;
-}
-
-template <>
-inline float16x8x2_t convolve_3x3<3>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2,
-                                     int fixed_point_position)
-{
-    float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]        = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 3), out.val[0], 1);
-    return out;
-}
-
-template <unsigned int stridex>
-void store_results(float16_t *buffer, const float16x8x2_t &values);
-
-template <>
-void store_results<1>(float16_t *buffer, const float16x8x2_t &values)
-{
-    vst1q_f16(buffer, values.val[0]);
-    vst1q_f16(buffer + 8, values.val[1]);
-}
-
-template <>
-void store_results<2>(float16_t *buffer, const float16x8x2_t &values)
-{
-    vst1q_f16(buffer, values.val[0]);
-}
-
-template <>
-void store_results<3>(float16_t *buffer, const float16x8x2_t &values)
-{
-    vst1_f16(buffer, vget_low_f16(values.val[0]));
-}
 
 template <unsigned int stridex>
 void accumulate_results(float16_t *buffer, const float16x8x2_t &values);
@@ -569,33 +455,6 @@ void accumulate_results<3>(float16_t *buffer, const float16x8x2_t &values)
 
 #endif /* ARM_COMPUTE_AARCH64_V8_2 */
 
-inline float32x4x3_t load_matrix_row(const float *ptr)
-{
-    const float32x4x3_t r =
-    {
-        {
-            vld1q_dup_f32(ptr),
-            vld1q_dup_f32(1 + ptr),
-            vld1q_dup_f32(2 + ptr)
-        }
-    };
-    return r;
-}
-inline qint8x8x3_t load_matrix_row(const qint8_t *ptr)
-{
-    /* ptr is a pointer to a row in a 3x3 matrix, the function returns 3 vectors holding exactly the same value in all lanes:
-       r.val[0] contains the first element, r.val[1] the second element and r.val[2] the third element (in all lanes) */
-    const qint8x8x3_t r =
-    {
-        {
-            vld1_dup_qs8(ptr),
-            vld1_dup_qs8(1 + ptr),
-            vld1_dup_qs8(2 + ptr)
-        }
-    };
-    return r;
-}
-
 template <unsigned int stridex>
 float32x4x2_t convolve_5x5(const float *in_0, const float *in_1, const float *in_2, const float *in_3, const float *in_4,
                            const float *m0, const float *m1, const float *m2, const float *m3, const float *m4, int fixed_point_position);
@@ -750,213 +609,6 @@ inline float32x4x2_t convolve_5x5<3>(const float *in_0, const float *in_1, const
 }
 
 template <unsigned int stridex>
-float32x4x2_t convolve_3x3(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position);
-
-template <>
-inline float32x4x2_t convolve_3x3<1>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
-{
-    ARM_COMPUTE_UNUSED(fixed_point_position);
-
-    const float32x4x3_t vtop =
-    {
-        {
-            vld1q_f32(in_top),
-            vld1q_f32(in_top + 4),
-            vld1q_f32(in_top + 8)
-        }
-    };
-    const float32x4x3_t vmid =
-    {
-        {
-            vld1q_f32(in_mid),
-            vld1q_f32(in_mid + 4),
-            vld1q_f32(in_mid + 8)
-        }
-    };
-    const float32x4x3_t vlow =
-    {
-        {
-            vld1q_f32(in_low),
-            vld1q_f32(in_low + 4),
-            vld1q_f32(in_low + 8)
-        }
-    };
-    float32x4x2_t out =
-    {
-        {
-            vmulq_f32(vtop.val[0], m0.val[0]),
-            vmulq_f32(vtop.val[1], m0.val[0])
-        }
-    };
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 1), m0.val[1]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 2), m0.val[2]);
-
-    out.val[0] = vmlaq_f32(out.val[0], vmid.val[0], m1.val[0]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 1), m1.val[1]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 2), m1.val[2]);
-
-    out.val[0] = vmlaq_f32(out.val[0], vlow.val[0], m2.val[0]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 1), m2.val[1]);
-    out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 2), m2.val[2]);
-
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 1), m0.val[1]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 2), m0.val[2]);
-
-    out.val[1] = vmlaq_f32(out.val[1], vmid.val[1], m1.val[0]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 1), m1.val[1]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 2), m1.val[2]);
-
-    out.val[1] = vmlaq_f32(out.val[1], vlow.val[1], m2.val[0]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 1), m2.val[1]);
-    out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 2), m2.val[2]);
-    return out;
-}
-
-template <>
-inline float32x4x2_t convolve_3x3<2>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
-{
-    float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 2), out.val[0], 1);
-    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 0), out.val[0], 2);
-    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 2), out.val[0], 3);
-    return out;
-}
-
-template <>
-inline float32x4x2_t convolve_3x3<3>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position)
-{
-    float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]        = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 3), out.val[0], 1);
-    return out;
-}
-
-template <unsigned int stridex>
-qint16x8x2_t convolve_3x3(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position);
-
-template <>
-inline qint16x8x2_t convolve_3x3<1>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
-{
-    ARM_COMPUTE_UNUSED(fixed_point_position);
-
-    const qint8x8x3_t vtop =
-    {
-        {
-            vld1_qs8(in_top),
-            vld1_qs8(in_top + 8),
-            vld1_qs8(in_top + 16)
-        }
-    };
-    const qint8x8x3_t vmid =
-    {
-        {
-            vld1_qs8(in_mid),
-            vld1_qs8(in_mid + 8),
-            vld1_qs8(in_mid + 16)
-        }
-    };
-    const qint8x8x3_t vlow =
-    {
-        {
-            vld1_qs8(in_low),
-            vld1_qs8(in_low + 8),
-            vld1_qs8(in_low + 16)
-        }
-    };
-    qint16x8x2_t out =
-    {
-        {
-            vmull_qs8(vtop.val[0], m0.val[0], fixed_point_position),
-            vmull_qs8(vtop.val[1], m0.val[0], fixed_point_position)
-        }
-    };
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vtop.val[0], vtop.val[1], 1), m0.val[1], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vtop.val[0], vtop.val[1], 2), m0.val[2], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vmid.val[0], m1.val[0], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vmid.val[0], vmid.val[1], 1), m1.val[1], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vmid.val[0], vmid.val[1], 2), m1.val[2], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vlow.val[0], m2.val[0], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vlow.val[0], vlow.val[1], 1), m2.val[1], fixed_point_position);
-    out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vlow.val[0], vlow.val[1], 2), m2.val[2], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vtop.val[1], vtop.val[2], 1), m0.val[1], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vtop.val[1], vtop.val[2], 2), m0.val[2], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vmid.val[1], m1.val[0], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vmid.val[1], vmid.val[2], 1), m1.val[1], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vmid.val[1], vmid.val[2], 2), m1.val[2], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vlow.val[1], m2.val[0], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vlow.val[1], vlow.val[2], 1), m2.val[1], fixed_point_position);
-    out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vlow.val[1], vlow.val[2], 2), m2.val[2], fixed_point_position);
-    return out;
-}
-
-template <>
-inline qint16x8x2_t convolve_3x3<2>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
-{
-    qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 2), out.val[0], 1);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 4), out.val[0], 2);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 6), out.val[0], 3);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 0), out.val[0], 4);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 2), out.val[0], 5);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 4), out.val[0], 6);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 6), out.val[0], 7);
-    return out;
-}
-
-template <>
-inline qint16x8x2_t convolve_3x3<3>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position)
-{
-    qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 3), out.val[0], 1);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 6), out.val[0], 2);
-    out.val[0]       = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 1), out.val[0], 3);
-    return out;
-}
-
-template <unsigned int stridex>
-void store_results(float *buffer, const float32x4x2_t &values);
-
-template <>
-void store_results<1>(float *buffer, const float32x4x2_t &values)
-{
-    vst1q_f32(buffer, values.val[0]);
-    vst1q_f32(buffer + 4, values.val[1]);
-}
-
-template <>
-void store_results<2>(float *buffer, const float32x4x2_t &values)
-{
-    vst1q_f32(buffer, values.val[0]);
-}
-
-template <>
-void store_results<3>(float *buffer, const float32x4x2_t &values)
-{
-    vst1_f32(buffer, vget_low_f32(values.val[0]));
-}
-
-template <unsigned int stridex>
-void store_results(qint16_t *buffer, const qint16x8x2_t &values);
-
-template <>
-void store_results<1>(qint16_t *buffer, const qint16x8x2_t &values)
-{
-    vst1q_qs16(buffer, values.val[0]);
-    vst1q_qs16(buffer + 8, values.val[1]);
-}
-
-template <>
-void store_results<2>(qint16_t *buffer, const qint16x8x2_t &values)
-{
-    vst1q_qs16(buffer, values.val[0]);
-}
-
-template <>
-void store_results<3>(qint16_t *buffer, const qint16x8x2_t &values)
-{
-    vst1_qs16(buffer, vget_low_s16(values.val[0]));
-}
-
-template <unsigned int stridex>
 void accumulate_results(float *buffer, const float32x4x2_t &values);
 
 template <>
@@ -1000,27 +652,6 @@ void accumulate_results<3>(qint16_t *buffer, const qint16x8x2_t &values)
     vst1_qs16(buffer, vqadd_qs16(vld1_qs16(buffer), vget_low_s16(values.val[0])));
 }
 
-template <unsigned int stridex>
-int get_input_num_elems_processed(unsigned int num_elems_written_per_iteration);
-
-template <>
-int get_input_num_elems_processed<1>(unsigned int num_elems_written_per_iteration)
-{
-    return num_elems_written_per_iteration;
-}
-
-template <>
-int get_input_num_elems_processed<2>(unsigned int num_elems_written_per_iteration)
-{
-    return num_elems_written_per_iteration << 1;
-}
-
-template <>
-int get_input_num_elems_processed<3>(unsigned int num_elems_written_per_iteration)
-{
-    return num_elems_written_per_iteration * 3;
-}
-
 template <typename T1, typename T2, unsigned int stridex>
 class convolver_3x3
 {