diff options
Diffstat (limited to 'arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h')
| -rw-r--r-- | arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h | 249 |
1 files changed, 38 insertions, 211 deletions
diff --git a/arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h b/arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h index 908fa13876..d56fd44700 100644 --- a/arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h +++ b/arm_compute/core/NEON/kernels/detail/NEDirectConvolutionDetail.h @@ -55,29 +55,6 @@ inline float32x4x3_t load_matrix_row(const float *ptr, int weights_offset = 0) return r; } -/** Loads a 3x3 matrix as a row (qint8_t). - * - * @param[in] ptr Pointer to a qint8 3x3 matrix. - * @param[in] weights_offset (Optional) Weights quantization offset. - * - * @return The loaded matrix. - */ -inline qint8x8x3_t load_matrix_row(const qint8_t *ptr, int weights_offset = 0) -{ - ARM_COMPUTE_UNUSED(weights_offset); - /* 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; -} - /** Loads a 3x3 matrix as a row (uint8_t). * * @param[in] ptr Pointer to a uint8_t 3x3 matrix. @@ -104,27 +81,25 @@ inline int32x4x3_t load_matrix_row(const uint8_t *ptr, int weights_offset = 0) /** Perform a convolve3x3 on float32. * - * @param[in] in_top Pointer to the first row of the input. - * @param[in] in_mid Pointer to the second row of the input. - * @param[in] in_low Pointer to the third row of the input. - * @param[in] m0 First row of the filter. - * @param[in] m1 Second row of the filter. - * @param[in] m2 Third row of the filter. - * @param[in] fixed_point_position (Optional) Fixed point position. - * @param[in] input_offset (Optional) Input quantization offset. + * @param[in] in_top Pointer to the first row of the input. + * @param[in] in_mid Pointer to the second row of the input. + * @param[in] in_low Pointer to the third row of the input. + * @param[in] m0 First row of the filter. + * @param[in] m1 Second row of the filter. + * @param[in] m2 Third row of the filter. + * @param[in] input_offset (Optional) Input quantization offset. * */ 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, int input_offset = 0); + int input_offset = 0); 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, int input_offset) + int input_offset) { - ARM_COMPUTE_UNUSED(fixed_point_position); ARM_COMPUTE_UNUSED(input_offset); const float32x4x3_t vtop = @@ -185,11 +160,11 @@ inline float32x4x2_t convolve_3x3<1>(const float *in_top, const float *in_mid, c 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, int input_offset) + int input_offset) { ARM_COMPUTE_UNUSED(input_offset); - float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position, input_offset); + float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, input_offset); 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); @@ -199,145 +174,35 @@ inline float32x4x2_t convolve_3x3<2>(const float *in_top, const float *in_mid, c 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, int input_offset) + int input_offset) { ARM_COMPUTE_UNUSED(input_offset); - float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position, input_offset); + float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, input_offset); out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 3), out.val[0], 1); return out; } -/** Perform a convolve3x3 on qint16. - * - * @param[in] in_top Pointer to the first row of the input. - * @param[in] in_mid Pointer to the second row of the input. - * @param[in] in_low Pointer to the third row of the input. - * @param[in] m0 First row of the filter. - * @param[in] m1 Second row of the filter. - * @param[in] m2 Third row of the filter. - * @param[in] fixed_point_position (Optional) Fixed point position. - * @param[in] input_offset (Optional) Input quantization offset. - * - */ -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, int input_offset = 0); - -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, int input_offset) -{ - ARM_COMPUTE_UNUSED(fixed_point_position); - ARM_COMPUTE_UNUSED(input_offset); - - 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, int input_offset) -{ - ARM_COMPUTE_UNUSED(input_offset); - - qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position, input_offset); - 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, int input_offset) -{ - ARM_COMPUTE_UNUSED(input_offset); - - qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position, input_offset); - 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; -} - /** Perform a convolve3x3 on uint8_t * - * @param[in] in_top Pointer to the first row of the input. - * @param[in] in_mid Pointer to the second row of the input. - * @param[in] in_low Pointer to the third row of the input. - * @param[in] m0 First row of the filter. - * @param[in] m1 Second row of the filter. - * @param[in] m2 Third row of the filter. - * @param[in] fixed_point_position (Optional) Fixed point position. - * @param[in] input_offset (Optional) Input quantization offset. + * @param[in] in_top Pointer to the first row of the input. + * @param[in] in_mid Pointer to the second row of the input. + * @param[in] in_low Pointer to the third row of the input. + * @param[in] m0 First row of the filter. + * @param[in] m1 Second row of the filter. + * @param[in] m2 Third row of the filter. + * @param[in] input_offset (Optional) Input quantization offset. * */ template <unsigned int stridex> int32x4x2_t convolve_3x3(const uint8_t *in_top, const uint8_t *in_mid, const uint8_t *in_low, const int32x4x3_t &m0, const int32x4x3_t &m1, const int32x4x3_t &m2, - int fixed_point_position, int input_offset); + int input_offset); template <> inline int32x4x2_t convolve_3x3<1>(const uint8_t *in_top, const uint8_t *in_mid, const uint8_t *in_low, const int32x4x3_t &m0, const int32x4x3_t &m1, const int32x4x3_t &m2, - int fixed_point_position, int input_offset) + int input_offset) { - ARM_COMPUTE_UNUSED(fixed_point_position); - const int32x4_t v_input_offset = vdupq_n_s32(input_offset); const uint8x8x2_t vtop = @@ -427,11 +292,9 @@ inline int32x4x2_t convolve_3x3<1>(const uint8_t *in_top, const uint8_t *in_mid, template <> inline int32x4x2_t convolve_3x3<2>(const uint8_t *in_top, const uint8_t *in_mid, const uint8_t *in_low, const int32x4x3_t &m0, const int32x4x3_t &m1, const int32x4x3_t &m2, - int fixed_point_position, int input_offset) + int input_offset) { - ARM_COMPUTE_UNUSED(fixed_point_position); - - int32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position, input_offset); + int32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, input_offset); out.val[0] = vsetq_lane_s32(vgetq_lane_s32(out.val[0], 2), out.val[0], 1); out.val[0] = vsetq_lane_s32(vgetq_lane_s32(out.val[1], 0), out.val[0], 2); out.val[0] = vsetq_lane_s32(vgetq_lane_s32(out.val[1], 2), out.val[0], 3); @@ -441,10 +304,9 @@ inline int32x4x2_t convolve_3x3<2>(const uint8_t *in_top, const uint8_t *in_mid, template <> inline int32x4x2_t convolve_3x3<3>(const uint8_t *in_top, const uint8_t *in_mid, const uint8_t *in_low, const int32x4x3_t &m0, const int32x4x3_t &m1, const int32x4x3_t &m2, - int fixed_point_position, int input_offset) + int input_offset) { - ARM_COMPUTE_UNUSED(fixed_point_position); - int32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position, input_offset); + int32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, input_offset); out.val[0] = vsetq_lane_s32(vgetq_lane_s32(out.val[0], 3), out.val[0], 1); return out; } @@ -477,34 +339,6 @@ inline void store_results<3>(float *buffer, const float32x4x2_t &values) vst1_f32(buffer, vget_low_f32(values.val[0])); } -/** Stores a qint16_t array into a memory location. - * - * @param[in] buffer Pointer to the memory location where the values will be stored. - * @param[in] values Values that will be stored. - * - */ -template <unsigned int stridex> -void store_results(qint16_t *buffer, const qint16x8x2_t &values); - -template <> -inline 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 <> -inline void store_results<2>(qint16_t *buffer, const qint16x8x2_t &values) -{ - vst1q_qs16(buffer, values.val[0]); -} - -template <> -inline void store_results<3>(qint16_t *buffer, const qint16x8x2_t &values) -{ - vst1_qs16(buffer, vget_low_s16(values.val[0])); -} - /** Stores a uint32_t array into a memory location. * * @param[in] buffer Pointer to the memory location where the values will be stored. @@ -557,25 +391,20 @@ inline float16x8x3_t load_matrix_row(const float16_t *ptr) /** Perform a convolve3x3 on float16. * - * @param[in] in_top Pointer to the first row of the input. - * @param[in] in_mid Pointer to the second row of the input. - * @param[in] in_low Pointer to the third row of the input. - * @param[in] m0 First row of the filter. - * @param[in] m1 Second row of the filter. - * @param[in] m2 Third row of the filter. - * @param[in] fixed_point_position (Optional) Fixed point position. + * @param[in] in_top Pointer to the first row of the input. + * @param[in] in_mid Pointer to the second row of the input. + * @param[in] in_low Pointer to the third row of the input. + * @param[in] m0 First row of the filter. + * @param[in] m1 Second row of the filter. + * @param[in] m2 Third row of the filter. * */ 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); +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); template <> -inline 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) +inline 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) { - ARM_COMPUTE_UNUSED(fixed_point_position); - const float16x8x3_t vtop = { { @@ -627,10 +456,9 @@ inline float16x8x2_t convolve_3x3<1>(const float16_t *in_top, const float16_t *i } 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) +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) { - float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position); + float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2); 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); @@ -638,10 +466,9 @@ inline float16x8x2_t convolve_3x3<2>(const float16_t *in_top, const float16_t *i } 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) +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) { - float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position); + float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2); out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 3), out.val[0], 1); return out; } |