|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include <assert.h> |
|
|
|
#include <arm_neon.h> |
|
|
|
#include <xnnpack/dwconv.h> |
|
#include <xnnpack/math.h> |
|
|
|
|
|
void xnn_f32_dwconv2d_chw_ukernel_3x3p1__aarch64_neonfma_1x4_acc3( |
|
size_t input_height, |
|
size_t input_width, |
|
const float* input, |
|
const float* weights, |
|
const float* zero, |
|
float* output, |
|
uint32_t padding_top, |
|
const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
|
{ |
|
assert(input_height != 0); |
|
assert(input_width != 0); |
|
assert(input_width % sizeof(float) == 0); |
|
assert(padding_top == 1); |
|
|
|
const uint32x4_t vmask = vld1q_u32(params->neon_stride1.mask); |
|
const float32x4_t vmax = vld1q_dup_f32(¶ms->neon_stride1.max); |
|
const float32x4_t vmin = vld1q_dup_f32(¶ms->neon_stride1.min); |
|
|
|
const float32x4_t vw0123 = vld1q_f32(weights); |
|
const float32x4_t vw4567 = vld1q_f32(weights + 4); |
|
const float32x2_t vw89 = vld1_f32(weights + 8); |
|
|
|
const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float)); |
|
|
|
const float* i0 = zero; |
|
const float* i1 = input; |
|
const float* i2 = (const float*) ((uintptr_t) i1 + input_width); |
|
|
|
float* o0 = output; |
|
|
|
size_t output_height = input_height; |
|
do { |
|
if XNN_UNPREDICTABLE(output_height < 2) { |
|
i2 = zero; |
|
} |
|
|
|
float32x4_t vi0x0123 = vmovq_n_f32(0.0f); |
|
float32x4_t vi1x0123 = vmovq_n_f32(0.0f); |
|
float32x4_t vi2x0123 = vmovq_n_f32(0.0f); |
|
|
|
float32x4_t vi0x4567 = vld1q_f32(i0); i0 += 4; |
|
float32x4_t vi1x4567 = vld1q_f32(i1); i1 += 4; |
|
float32x4_t vi2x4567 = vld1q_f32(i2); i2 += 4; |
|
|
|
size_t w = input_width; |
|
for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) { |
|
float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); |
|
|
|
const float32x4_t vi0x89AB = vld1q_f32(i0); i0 += 4; |
|
const float32x4_t vi1x89AB = vld1q_f32(i1); i1 += 4; |
|
const float32x4_t vi2x89AB = vld1q_f32(i2); i2 += 4; |
|
|
|
vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 0); |
|
|
|
float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); |
|
|
|
float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); |
|
|
|
const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3); |
|
const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3); |
|
const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); |
|
|
|
vo0p0 = vfmaq_lane_f32(vo0p0, vi0x3456, vget_low_f32(vw0123), 1); |
|
|
|
vo0p1 = vfmaq_lane_f32(vo0p1, vi1x3456, vget_low_f32(vw4567), 0); |
|
|
|
vo0p2 = vfmaq_lane_f32(vo0p2, vi2x3456, vget_high_f32(vw4567), 1); |
|
|
|
vi0x0123 = vi0x4567; |
|
vi1x0123 = vi1x4567; |
|
vi2x0123 = vi2x4567; |
|
|
|
const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); |
|
const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vi1x89AB, 1); |
|
const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vi2x89AB, 1); |
|
|
|
vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); |
|
|
|
vo0p1 = vfmaq_lane_f32(vo0p1, vi1x5678, vget_high_f32(vw4567), 0); |
|
|
|
vo0p2 = vfmaq_lane_f32(vo0p2, vi2x5678, vw89, 1); |
|
|
|
vi0x4567 = vi0x89AB; |
|
vi1x4567 = vi1x89AB; |
|
vi2x4567 = vi2x89AB; |
|
|
|
vo0p0 = vaddq_f32(vo0p0, vo0p1); |
|
vo0p0 = vaddq_f32(vo0p0, vo0p2); |
|
|
|
float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); |
|
|
|
vo0 = vminq_f32(vo0, vmax); |
|
|
|
vst1q_f32(o0, vo0); o0 += 4; |
|
} |
|
|
|
assert(w >= 1 * sizeof(float)); |
|
assert(w <= 4 * sizeof(float)); |
|
{ |
|
float32x4_t vo0p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); |
|
|
|
vi0x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi0x4567))); |
|
vi1x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi1x4567))); |
|
vi2x4567 = vreinterpretq_f32_u32(vandq_u32(vmask, vreinterpretq_u32_f32(vi2x4567))); |
|
|
|
vo0p0 = vfmaq_lane_f32(vo0p0, vi0x4567, vget_high_f32(vw0123), 0); |
|
|
|
float32x4_t vo0p1 = vmulq_lane_f32(vi1x4567, vget_low_f32(vw4567), 1); |
|
|
|
float32x4_t vo0p2 = vmulq_lane_f32(vi2x4567, vw89, 0); |
|
|
|
const float32x4_t vi0x3456 = vextq_f32(vi0x0123, vi0x4567, 3); |
|
const float32x4_t vi1x3456 = vextq_f32(vi1x0123, vi1x4567, 3); |
|
const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); |
|
|
|
vo0p0 = vfmaq_lane_f32(vo0p0, vi0x3456, vget_low_f32(vw0123), 1); |
|
|
|
vo0p1 = vfmaq_lane_f32(vo0p1, vi1x3456, vget_low_f32(vw4567), 0); |
|
|
|
vo0p2 = vfmaq_lane_f32(vo0p2, vi2x3456, vget_high_f32(vw4567), 1); |
|
|
|
const float32x4_t vzero = vmovq_n_f32(0.0f); |
|
const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); |
|
const float32x4_t vi1x5678 = vextq_f32(vi1x4567, vzero, 1); |
|
const float32x4_t vi2x5678 = vextq_f32(vi2x4567, vzero, 1); |
|
|
|
vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); |
|
|
|
vo0p1 = vfmaq_lane_f32(vo0p1, vi1x5678, vget_high_f32(vw4567), 0); |
|
|
|
vo0p2 = vfmaq_lane_f32(vo0p2, vi2x5678, vw89, 1); |
|
|
|
vo0p0 = vaddq_f32(vo0p0, vo0p1); |
|
vo0p0 = vaddq_f32(vo0p0, vo0p2); |
|
|
|
float32x4_t vo0 = vmaxq_f32(vo0p0, vmin); |
|
|
|
vo0 = vminq_f32(vo0, vmax); |
|
|
|
if XNN_LIKELY(w == 4 * sizeof(float)) { |
|
vst1q_f32(o0, vo0); o0 += 4; |
|
} else { |
|
float32x2_t vo0_lo = vget_low_f32(vo0); |
|
if (w & (2 * sizeof(float))) { |
|
vst1_f32(o0, vo0_lo); o0 += 2; |
|
|
|
vo0_lo = vget_high_f32(vo0); |
|
} |
|
if (w & (1 * sizeof(float))) { |
|
vst1_lane_f32(o0, vo0_lo, 0); o0 += 1; |
|
} |
|
} |
|
} |
|
|
|
i0 = (const float*) ((uintptr_t) i1 - input_decrement); |
|
i1 = (const float*) ((uintptr_t) i2 - input_decrement); |
|
i2 = (const float*) ((uintptr_t) i1 + input_width); |
|
|
|
|
|
} while (--output_height != 0); |
|
} |
|
|