|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include <assert.h> |
|
|
|
#include <arm_neon.h> |
|
|
|
#include <xnnpack/gemm.h> |
|
|
|
|
|
void xnn_bf16_gemm_minmax_ukernel_6x8c2__neonbf16_bfdot_lane_ld128( |
|
size_t mr, |
|
size_t nc, |
|
size_t kc, |
|
const void* restrict a, |
|
size_t a_stride, |
|
const void* restrict w_ptr, |
|
void* restrict c, |
|
size_t cm_stride, |
|
size_t cn_stride, |
|
const union xnn_bf16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
|
{ |
|
assert(mr != 0); |
|
assert(mr <= 6); |
|
assert(nc != 0); |
|
assert(kc != 0); |
|
assert(kc % sizeof(bfloat16_t) == 0); |
|
assert(a != NULL); |
|
assert(w_ptr != NULL); |
|
assert(c != NULL); |
|
|
|
const bfloat16_t* a0 = (const bfloat16_t*) a; |
|
bfloat16_t* c0 = (bfloat16_t*) c; |
|
const bfloat16_t* a1 = (const bfloat16_t*) ((uintptr_t) a0 + a_stride); |
|
bfloat16_t* c1 = (bfloat16_t*) ((uintptr_t) c0 + cm_stride); |
|
if XNN_UNPREDICTABLE(mr < 2) { |
|
a1 = a0; |
|
c1 = c0; |
|
} |
|
const bfloat16_t* a2 = (const bfloat16_t*) ((uintptr_t) a1 + a_stride); |
|
bfloat16_t* c2 = (bfloat16_t*) ((uintptr_t) c1 + cm_stride); |
|
if XNN_UNPREDICTABLE(mr <= 2) { |
|
a2 = a1; |
|
c2 = c1; |
|
} |
|
const bfloat16_t* a3 = (const bfloat16_t*) ((uintptr_t) a2 + a_stride); |
|
bfloat16_t* c3 = (bfloat16_t*) ((uintptr_t) c2 + cm_stride); |
|
if XNN_UNPREDICTABLE(mr < 4) { |
|
a3 = a2; |
|
c3 = c2; |
|
} |
|
const bfloat16_t* a4 = (const bfloat16_t*) ((uintptr_t) a3 + a_stride); |
|
bfloat16_t* c4 = (bfloat16_t*) ((uintptr_t) c3 + cm_stride); |
|
if XNN_UNPREDICTABLE(mr <= 4) { |
|
a4 = a3; |
|
c4 = c3; |
|
} |
|
const bfloat16_t* a5 = (const bfloat16_t*) ((uintptr_t) a4 + a_stride); |
|
bfloat16_t* c5 = (bfloat16_t*) ((uintptr_t) c4 + cm_stride); |
|
if XNN_UNPREDICTABLE(mr != 6) { |
|
a5 = a4; |
|
c5 = c4; |
|
} |
|
|
|
const bfloat16_t* w = (const bfloat16_t*) w_ptr; |
|
do { |
|
float32x4_t vacc0x0123 = vcvt_f32_bf16(vld1_bf16(w)); w += 4; |
|
float32x4_t vacc0x4567 = vcvt_f32_bf16(vld1_bf16(w)); w += 4; |
|
float32x4_t vacc1x0123 = vacc0x0123; |
|
float32x4_t vacc1x4567 = vacc0x4567; |
|
float32x4_t vacc2x0123 = vacc0x0123; |
|
float32x4_t vacc2x4567 = vacc0x4567; |
|
float32x4_t vacc3x0123 = vacc0x0123; |
|
float32x4_t vacc3x4567 = vacc0x4567; |
|
float32x4_t vacc4x0123 = vacc0x0123; |
|
float32x4_t vacc4x4567 = vacc0x4567; |
|
float32x4_t vacc5x0123 = vacc0x0123; |
|
float32x4_t vacc5x4567 = vacc0x4567; |
|
|
|
size_t k = kc; |
|
for (; k >= 8 * sizeof(bfloat16_t); k -= 8 * sizeof(bfloat16_t)) { |
|
const bfloat16x8_t va0 = vld1q_bf16(a0); a0 += 8; |
|
const bfloat16x8_t va1 = vld1q_bf16(a1); a1 += 8; |
|
const bfloat16x8_t va2 = vld1q_bf16(a2); a2 += 8; |
|
const bfloat16x8_t va3 = vld1q_bf16(a3); a3 += 8; |
|
const bfloat16x8_t va4 = vld1q_bf16(a4); a4 += 8; |
|
const bfloat16x8_t va5 = vld1q_bf16(a5); a5 += 8; |
|
|
|
const bfloat16x8_t vb0123c01 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c01 = vld1q_bf16(w); w += 8; |
|
|
|
vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c01, va0, 0); |
|
vacc1x0123 = vbfdotq_laneq_f32(vacc1x0123, vb0123c01, va1, 0); |
|
vacc2x0123 = vbfdotq_laneq_f32(vacc2x0123, vb0123c01, va2, 0); |
|
vacc3x0123 = vbfdotq_laneq_f32(vacc3x0123, vb0123c01, va3, 0); |
|
vacc4x0123 = vbfdotq_laneq_f32(vacc4x0123, vb0123c01, va4, 0); |
|
vacc5x0123 = vbfdotq_laneq_f32(vacc5x0123, vb0123c01, va5, 0); |
|
vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c01, va0, 0); |
|
vacc1x4567 = vbfdotq_laneq_f32(vacc1x4567, vb4567c01, va1, 0); |
|
vacc2x4567 = vbfdotq_laneq_f32(vacc2x4567, vb4567c01, va2, 0); |
|
vacc3x4567 = vbfdotq_laneq_f32(vacc3x4567, vb4567c01, va3, 0); |
|
vacc4x4567 = vbfdotq_laneq_f32(vacc4x4567, vb4567c01, va4, 0); |
|
vacc5x4567 = vbfdotq_laneq_f32(vacc5x4567, vb4567c01, va5, 0); |
|
const bfloat16x8_t vb0123c23 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c23 = vld1q_bf16(w); w += 8; |
|
|
|
vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c23, va0, 1); |
|
vacc1x0123 = vbfdotq_laneq_f32(vacc1x0123, vb0123c23, va1, 1); |
|
vacc2x0123 = vbfdotq_laneq_f32(vacc2x0123, vb0123c23, va2, 1); |
|
vacc3x0123 = vbfdotq_laneq_f32(vacc3x0123, vb0123c23, va3, 1); |
|
vacc4x0123 = vbfdotq_laneq_f32(vacc4x0123, vb0123c23, va4, 1); |
|
vacc5x0123 = vbfdotq_laneq_f32(vacc5x0123, vb0123c23, va5, 1); |
|
vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c23, va0, 1); |
|
vacc1x4567 = vbfdotq_laneq_f32(vacc1x4567, vb4567c23, va1, 1); |
|
vacc2x4567 = vbfdotq_laneq_f32(vacc2x4567, vb4567c23, va2, 1); |
|
vacc3x4567 = vbfdotq_laneq_f32(vacc3x4567, vb4567c23, va3, 1); |
|
vacc4x4567 = vbfdotq_laneq_f32(vacc4x4567, vb4567c23, va4, 1); |
|
vacc5x4567 = vbfdotq_laneq_f32(vacc5x4567, vb4567c23, va5, 1); |
|
const bfloat16x8_t vb0123c45 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c45 = vld1q_bf16(w); w += 8; |
|
|
|
vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c45, va0, 2); |
|
vacc1x0123 = vbfdotq_laneq_f32(vacc1x0123, vb0123c45, va1, 2); |
|
vacc2x0123 = vbfdotq_laneq_f32(vacc2x0123, vb0123c45, va2, 2); |
|
vacc3x0123 = vbfdotq_laneq_f32(vacc3x0123, vb0123c45, va3, 2); |
|
vacc4x0123 = vbfdotq_laneq_f32(vacc4x0123, vb0123c45, va4, 2); |
|
vacc5x0123 = vbfdotq_laneq_f32(vacc5x0123, vb0123c45, va5, 2); |
|
vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c45, va0, 2); |
|
vacc1x4567 = vbfdotq_laneq_f32(vacc1x4567, vb4567c45, va1, 2); |
|
vacc2x4567 = vbfdotq_laneq_f32(vacc2x4567, vb4567c45, va2, 2); |
|
vacc3x4567 = vbfdotq_laneq_f32(vacc3x4567, vb4567c45, va3, 2); |
|
vacc4x4567 = vbfdotq_laneq_f32(vacc4x4567, vb4567c45, va4, 2); |
|
vacc5x4567 = vbfdotq_laneq_f32(vacc5x4567, vb4567c45, va5, 2); |
|
const bfloat16x8_t vb0123c67 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c67 = vld1q_bf16(w); w += 8; |
|
|
|
vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c67, va0, 3); |
|
vacc1x0123 = vbfdotq_laneq_f32(vacc1x0123, vb0123c67, va1, 3); |
|
vacc2x0123 = vbfdotq_laneq_f32(vacc2x0123, vb0123c67, va2, 3); |
|
vacc3x0123 = vbfdotq_laneq_f32(vacc3x0123, vb0123c67, va3, 3); |
|
vacc4x0123 = vbfdotq_laneq_f32(vacc4x0123, vb0123c67, va4, 3); |
|
vacc5x0123 = vbfdotq_laneq_f32(vacc5x0123, vb0123c67, va5, 3); |
|
vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c67, va0, 3); |
|
vacc1x4567 = vbfdotq_laneq_f32(vacc1x4567, vb4567c67, va1, 3); |
|
vacc2x4567 = vbfdotq_laneq_f32(vacc2x4567, vb4567c67, va2, 3); |
|
vacc3x4567 = vbfdotq_laneq_f32(vacc3x4567, vb4567c67, va3, 3); |
|
vacc4x4567 = vbfdotq_laneq_f32(vacc4x4567, vb4567c67, va4, 3); |
|
vacc5x4567 = vbfdotq_laneq_f32(vacc5x4567, vb4567c67, va5, 3); |
|
} |
|
if XNN_UNLIKELY(k != 0) { |
|
const bfloat16x8_t va0 = vld1q_bf16(a0); a0 = (const bfloat16_t*) ((uintptr_t) a0 + k); |
|
const bfloat16x8_t va1 = vld1q_bf16(a1); a1 = (const bfloat16_t*) ((uintptr_t) a1 + k); |
|
const bfloat16x8_t va2 = vld1q_bf16(a2); a2 = (const bfloat16_t*) ((uintptr_t) a2 + k); |
|
const bfloat16x8_t va3 = vld1q_bf16(a3); a3 = (const bfloat16_t*) ((uintptr_t) a3 + k); |
|
const bfloat16x8_t va4 = vld1q_bf16(a4); a4 = (const bfloat16_t*) ((uintptr_t) a4 + k); |
|
const bfloat16x8_t va5 = vld1q_bf16(a5); a5 = (const bfloat16_t*) ((uintptr_t) a5 + k); |
|
|
|
const bfloat16x8_t vb0123c01 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c01 = vld1q_bf16(w); w += 8; |
|
|
|
const uint32x4_t va0c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va0)), 0); |
|
const uint32x4_t va1c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va1)), 0); |
|
const uint32x4_t va2c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va2)), 0); |
|
const uint32x4_t va3c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va3)), 0); |
|
const uint32x4_t va4c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va4)), 0); |
|
const uint32x4_t va5c01 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va5)), 0); |
|
|
|
const uint32x4_t vm0123c01 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c01), vmovq_n_u16(0))); |
|
const uint32x4_t vm4567c01 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c01), vmovq_n_u16(0))); |
|
|
|
const uint32x4_t va0x0123c01 = vbicq_u32(va0c01, vm0123c01); |
|
vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c01, vreinterpretq_bf16_u32(va0x0123c01)); |
|
const uint32x4_t va1x0123c01 = vbicq_u32(va1c01, vm0123c01); |
|
vacc1x0123 = vbfdotq_f32(vacc1x0123, vb0123c01, vreinterpretq_bf16_u32(va1x0123c01)); |
|
const uint32x4_t va2x0123c01 = vbicq_u32(va2c01, vm0123c01); |
|
vacc2x0123 = vbfdotq_f32(vacc2x0123, vb0123c01, vreinterpretq_bf16_u32(va2x0123c01)); |
|
const uint32x4_t va3x0123c01 = vbicq_u32(va3c01, vm0123c01); |
|
vacc3x0123 = vbfdotq_f32(vacc3x0123, vb0123c01, vreinterpretq_bf16_u32(va3x0123c01)); |
|
const uint32x4_t va4x0123c01 = vbicq_u32(va4c01, vm0123c01); |
|
vacc4x0123 = vbfdotq_f32(vacc4x0123, vb0123c01, vreinterpretq_bf16_u32(va4x0123c01)); |
|
const uint32x4_t va5x0123c01 = vbicq_u32(va5c01, vm0123c01); |
|
vacc5x0123 = vbfdotq_f32(vacc5x0123, vb0123c01, vreinterpretq_bf16_u32(va5x0123c01)); |
|
const uint32x4_t va0x4567c01 = vbicq_u32(va0c01, vm4567c01); |
|
vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c01, vreinterpretq_bf16_u32(va0x4567c01)); |
|
const uint32x4_t va1x4567c01 = vbicq_u32(va1c01, vm4567c01); |
|
vacc1x4567 = vbfdotq_f32(vacc1x4567, vb4567c01, vreinterpretq_bf16_u32(va1x4567c01)); |
|
const uint32x4_t va2x4567c01 = vbicq_u32(va2c01, vm4567c01); |
|
vacc2x4567 = vbfdotq_f32(vacc2x4567, vb4567c01, vreinterpretq_bf16_u32(va2x4567c01)); |
|
const uint32x4_t va3x4567c01 = vbicq_u32(va3c01, vm4567c01); |
|
vacc3x4567 = vbfdotq_f32(vacc3x4567, vb4567c01, vreinterpretq_bf16_u32(va3x4567c01)); |
|
const uint32x4_t va4x4567c01 = vbicq_u32(va4c01, vm4567c01); |
|
vacc4x4567 = vbfdotq_f32(vacc4x4567, vb4567c01, vreinterpretq_bf16_u32(va4x4567c01)); |
|
const uint32x4_t va5x4567c01 = vbicq_u32(va5c01, vm4567c01); |
|
vacc5x4567 = vbfdotq_f32(vacc5x4567, vb4567c01, vreinterpretq_bf16_u32(va5x4567c01)); |
|
|
|
if (k > 2 * sizeof(bfloat16_t)) { |
|
const bfloat16x8_t vb0123c23 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c23 = vld1q_bf16(w); w += 8; |
|
|
|
const uint32x4_t va0c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va0)), 1); |
|
const uint32x4_t va1c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va1)), 1); |
|
const uint32x4_t va2c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va2)), 1); |
|
const uint32x4_t va3c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va3)), 1); |
|
const uint32x4_t va4c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va4)), 1); |
|
const uint32x4_t va5c23 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_low_bf16(va5)), 1); |
|
|
|
const uint32x4_t vm0123c23 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c23), vmovq_n_u16(0))); |
|
const uint32x4_t vm4567c23 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c23), vmovq_n_u16(0))); |
|
|
|
const uint32x4_t va0x0123c23 = vbicq_u32(va0c23, vm0123c23); |
|
vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c23, vreinterpretq_bf16_u32(va0x0123c23)); |
|
const uint32x4_t va1x0123c23 = vbicq_u32(va1c23, vm0123c23); |
|
vacc1x0123 = vbfdotq_f32(vacc1x0123, vb0123c23, vreinterpretq_bf16_u32(va1x0123c23)); |
|
const uint32x4_t va2x0123c23 = vbicq_u32(va2c23, vm0123c23); |
|
vacc2x0123 = vbfdotq_f32(vacc2x0123, vb0123c23, vreinterpretq_bf16_u32(va2x0123c23)); |
|
const uint32x4_t va3x0123c23 = vbicq_u32(va3c23, vm0123c23); |
|
vacc3x0123 = vbfdotq_f32(vacc3x0123, vb0123c23, vreinterpretq_bf16_u32(va3x0123c23)); |
|
const uint32x4_t va4x0123c23 = vbicq_u32(va4c23, vm0123c23); |
|
vacc4x0123 = vbfdotq_f32(vacc4x0123, vb0123c23, vreinterpretq_bf16_u32(va4x0123c23)); |
|
const uint32x4_t va5x0123c23 = vbicq_u32(va5c23, vm0123c23); |
|
vacc5x0123 = vbfdotq_f32(vacc5x0123, vb0123c23, vreinterpretq_bf16_u32(va5x0123c23)); |
|
const uint32x4_t va0x4567c23 = vbicq_u32(va0c23, vm4567c23); |
|
vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c23, vreinterpretq_bf16_u32(va0x4567c23)); |
|
const uint32x4_t va1x4567c23 = vbicq_u32(va1c23, vm4567c23); |
|
vacc1x4567 = vbfdotq_f32(vacc1x4567, vb4567c23, vreinterpretq_bf16_u32(va1x4567c23)); |
|
const uint32x4_t va2x4567c23 = vbicq_u32(va2c23, vm4567c23); |
|
vacc2x4567 = vbfdotq_f32(vacc2x4567, vb4567c23, vreinterpretq_bf16_u32(va2x4567c23)); |
|
const uint32x4_t va3x4567c23 = vbicq_u32(va3c23, vm4567c23); |
|
vacc3x4567 = vbfdotq_f32(vacc3x4567, vb4567c23, vreinterpretq_bf16_u32(va3x4567c23)); |
|
const uint32x4_t va4x4567c23 = vbicq_u32(va4c23, vm4567c23); |
|
vacc4x4567 = vbfdotq_f32(vacc4x4567, vb4567c23, vreinterpretq_bf16_u32(va4x4567c23)); |
|
const uint32x4_t va5x4567c23 = vbicq_u32(va5c23, vm4567c23); |
|
vacc5x4567 = vbfdotq_f32(vacc5x4567, vb4567c23, vreinterpretq_bf16_u32(va5x4567c23)); |
|
|
|
if (k > 4 * sizeof(bfloat16_t)) { |
|
const bfloat16x8_t vb0123c45 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c45 = vld1q_bf16(w); w += 8; |
|
|
|
const uint32x4_t va0c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va0)), 0); |
|
const uint32x4_t va1c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va1)), 0); |
|
const uint32x4_t va2c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va2)), 0); |
|
const uint32x4_t va3c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va3)), 0); |
|
const uint32x4_t va4c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va4)), 0); |
|
const uint32x4_t va5c45 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va5)), 0); |
|
|
|
const uint32x4_t vm0123c45 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c45), vmovq_n_u16(0))); |
|
const uint32x4_t vm4567c45 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c45), vmovq_n_u16(0))); |
|
|
|
const uint32x4_t va0x0123c45 = vbicq_u32(va0c45, vm0123c45); |
|
vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c45, vreinterpretq_bf16_u32(va0x0123c45)); |
|
const uint32x4_t va1x0123c45 = vbicq_u32(va1c45, vm0123c45); |
|
vacc1x0123 = vbfdotq_f32(vacc1x0123, vb0123c45, vreinterpretq_bf16_u32(va1x0123c45)); |
|
const uint32x4_t va2x0123c45 = vbicq_u32(va2c45, vm0123c45); |
|
vacc2x0123 = vbfdotq_f32(vacc2x0123, vb0123c45, vreinterpretq_bf16_u32(va2x0123c45)); |
|
const uint32x4_t va3x0123c45 = vbicq_u32(va3c45, vm0123c45); |
|
vacc3x0123 = vbfdotq_f32(vacc3x0123, vb0123c45, vreinterpretq_bf16_u32(va3x0123c45)); |
|
const uint32x4_t va4x0123c45 = vbicq_u32(va4c45, vm0123c45); |
|
vacc4x0123 = vbfdotq_f32(vacc4x0123, vb0123c45, vreinterpretq_bf16_u32(va4x0123c45)); |
|
const uint32x4_t va5x0123c45 = vbicq_u32(va5c45, vm0123c45); |
|
vacc5x0123 = vbfdotq_f32(vacc5x0123, vb0123c45, vreinterpretq_bf16_u32(va5x0123c45)); |
|
const uint32x4_t va0x4567c45 = vbicq_u32(va0c45, vm4567c45); |
|
vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c45, vreinterpretq_bf16_u32(va0x4567c45)); |
|
const uint32x4_t va1x4567c45 = vbicq_u32(va1c45, vm4567c45); |
|
vacc1x4567 = vbfdotq_f32(vacc1x4567, vb4567c45, vreinterpretq_bf16_u32(va1x4567c45)); |
|
const uint32x4_t va2x4567c45 = vbicq_u32(va2c45, vm4567c45); |
|
vacc2x4567 = vbfdotq_f32(vacc2x4567, vb4567c45, vreinterpretq_bf16_u32(va2x4567c45)); |
|
const uint32x4_t va3x4567c45 = vbicq_u32(va3c45, vm4567c45); |
|
vacc3x4567 = vbfdotq_f32(vacc3x4567, vb4567c45, vreinterpretq_bf16_u32(va3x4567c45)); |
|
const uint32x4_t va4x4567c45 = vbicq_u32(va4c45, vm4567c45); |
|
vacc4x4567 = vbfdotq_f32(vacc4x4567, vb4567c45, vreinterpretq_bf16_u32(va4x4567c45)); |
|
const uint32x4_t va5x4567c45 = vbicq_u32(va5c45, vm4567c45); |
|
vacc5x4567 = vbfdotq_f32(vacc5x4567, vb4567c45, vreinterpretq_bf16_u32(va5x4567c45)); |
|
|
|
if (k > 6 * sizeof(bfloat16_t)) { |
|
const bfloat16x8_t vb0123c67 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c67 = vld1q_bf16(w); w += 8; |
|
|
|
const uint32x4_t va0c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va0)), 1); |
|
const uint32x4_t va1c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va1)), 1); |
|
const uint32x4_t va2c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va2)), 1); |
|
const uint32x4_t va3c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va3)), 1); |
|
const uint32x4_t va4c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va4)), 1); |
|
const uint32x4_t va5c67 = vdupq_lane_u32(vreinterpret_u32_bf16(vget_high_bf16(va5)), 1); |
|
|
|
const uint32x4_t vm0123c67 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb0123c67), vmovq_n_u16(0))); |
|
const uint32x4_t vm4567c67 = vreinterpretq_u32_u16(vceqq_u16(vreinterpretq_u16_bf16(vb4567c67), vmovq_n_u16(0))); |
|
|
|
const uint32x4_t va0x0123c67 = vbicq_u32(va0c67, vm0123c67); |
|
vacc0x0123 = vbfdotq_f32(vacc0x0123, vb0123c67, vreinterpretq_bf16_u32(va0x0123c67)); |
|
const uint32x4_t va1x0123c67 = vbicq_u32(va1c67, vm0123c67); |
|
vacc1x0123 = vbfdotq_f32(vacc1x0123, vb0123c67, vreinterpretq_bf16_u32(va1x0123c67)); |
|
const uint32x4_t va2x0123c67 = vbicq_u32(va2c67, vm0123c67); |
|
vacc2x0123 = vbfdotq_f32(vacc2x0123, vb0123c67, vreinterpretq_bf16_u32(va2x0123c67)); |
|
const uint32x4_t va3x0123c67 = vbicq_u32(va3c67, vm0123c67); |
|
vacc3x0123 = vbfdotq_f32(vacc3x0123, vb0123c67, vreinterpretq_bf16_u32(va3x0123c67)); |
|
const uint32x4_t va4x0123c67 = vbicq_u32(va4c67, vm0123c67); |
|
vacc4x0123 = vbfdotq_f32(vacc4x0123, vb0123c67, vreinterpretq_bf16_u32(va4x0123c67)); |
|
const uint32x4_t va5x0123c67 = vbicq_u32(va5c67, vm0123c67); |
|
vacc5x0123 = vbfdotq_f32(vacc5x0123, vb0123c67, vreinterpretq_bf16_u32(va5x0123c67)); |
|
const uint32x4_t va0x4567c67 = vbicq_u32(va0c67, vm4567c67); |
|
vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c67, vreinterpretq_bf16_u32(va0x4567c67)); |
|
const uint32x4_t va1x4567c67 = vbicq_u32(va1c67, vm4567c67); |
|
vacc1x4567 = vbfdotq_f32(vacc1x4567, vb4567c67, vreinterpretq_bf16_u32(va1x4567c67)); |
|
const uint32x4_t va2x4567c67 = vbicq_u32(va2c67, vm4567c67); |
|
vacc2x4567 = vbfdotq_f32(vacc2x4567, vb4567c67, vreinterpretq_bf16_u32(va2x4567c67)); |
|
const uint32x4_t va3x4567c67 = vbicq_u32(va3c67, vm4567c67); |
|
vacc3x4567 = vbfdotq_f32(vacc3x4567, vb4567c67, vreinterpretq_bf16_u32(va3x4567c67)); |
|
const uint32x4_t va4x4567c67 = vbicq_u32(va4c67, vm4567c67); |
|
vacc4x4567 = vbfdotq_f32(vacc4x4567, vb4567c67, vreinterpretq_bf16_u32(va4x4567c67)); |
|
const uint32x4_t va5x4567c67 = vbicq_u32(va5c67, vm4567c67); |
|
vacc5x4567 = vbfdotq_f32(vacc5x4567, vb4567c67, vreinterpretq_bf16_u32(va5x4567c67)); |
|
} |
|
} |
|
} |
|
} |
|
|
|
const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); |
|
vacc0x0123 = vminq_f32(vacc0x0123, vmax); |
|
vacc1x0123 = vminq_f32(vacc1x0123, vmax); |
|
vacc2x0123 = vminq_f32(vacc2x0123, vmax); |
|
vacc3x0123 = vminq_f32(vacc3x0123, vmax); |
|
vacc4x0123 = vminq_f32(vacc4x0123, vmax); |
|
vacc5x0123 = vminq_f32(vacc5x0123, vmax); |
|
vacc0x4567 = vminq_f32(vacc0x4567, vmax); |
|
vacc1x4567 = vminq_f32(vacc1x4567, vmax); |
|
vacc2x4567 = vminq_f32(vacc2x4567, vmax); |
|
vacc3x4567 = vminq_f32(vacc3x4567, vmax); |
|
vacc4x4567 = vminq_f32(vacc4x4567, vmax); |
|
vacc5x4567 = vminq_f32(vacc5x4567, vmax); |
|
|
|
const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); |
|
vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); |
|
vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); |
|
vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); |
|
vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); |
|
vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); |
|
vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); |
|
vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); |
|
vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); |
|
vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); |
|
vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); |
|
vacc4x4567 = vmaxq_f32(vacc4x4567, vmin); |
|
vacc5x4567 = vmaxq_f32(vacc5x4567, vmin); |
|
|
|
bfloat16x4_t vout0x0123 = vcvt_bf16_f32(vacc0x0123); |
|
bfloat16x4_t vout1x0123 = vcvt_bf16_f32(vacc1x0123); |
|
bfloat16x4_t vout2x0123 = vcvt_bf16_f32(vacc2x0123); |
|
bfloat16x4_t vout3x0123 = vcvt_bf16_f32(vacc3x0123); |
|
bfloat16x4_t vout4x0123 = vcvt_bf16_f32(vacc4x0123); |
|
bfloat16x4_t vout5x0123 = vcvt_bf16_f32(vacc5x0123); |
|
bfloat16x4_t vout0x4567 = vcvt_bf16_f32(vacc0x4567); |
|
bfloat16x4_t vout1x4567 = vcvt_bf16_f32(vacc1x4567); |
|
bfloat16x4_t vout2x4567 = vcvt_bf16_f32(vacc2x4567); |
|
bfloat16x4_t vout3x4567 = vcvt_bf16_f32(vacc3x4567); |
|
bfloat16x4_t vout4x4567 = vcvt_bf16_f32(vacc4x4567); |
|
bfloat16x4_t vout5x4567 = vcvt_bf16_f32(vacc5x4567); |
|
|
|
if XNN_LIKELY(nc >= 8) { |
|
vst1_bf16(c0, vout0x0123); |
|
vst1_bf16(c0 + 4, vout0x4567); |
|
c0 = (bfloat16_t*) ((uintptr_t) c0 + cn_stride); |
|
vst1_bf16(c1, vout1x0123); |
|
vst1_bf16(c1 + 4, vout1x4567); |
|
c1 = (bfloat16_t*) ((uintptr_t) c1 + cn_stride); |
|
vst1_bf16(c2, vout2x0123); |
|
vst1_bf16(c2 + 4, vout2x4567); |
|
c2 = (bfloat16_t*) ((uintptr_t) c2 + cn_stride); |
|
vst1_bf16(c3, vout3x0123); |
|
vst1_bf16(c3 + 4, vout3x4567); |
|
c3 = (bfloat16_t*) ((uintptr_t) c3 + cn_stride); |
|
vst1_bf16(c4, vout4x0123); |
|
vst1_bf16(c4 + 4, vout4x4567); |
|
c4 = (bfloat16_t*) ((uintptr_t) c4 + cn_stride); |
|
vst1_bf16(c5, vout5x0123); |
|
vst1_bf16(c5 + 4, vout5x4567); |
|
c5 = (bfloat16_t*) ((uintptr_t) c5 + cn_stride); |
|
|
|
a0 = (const bfloat16_t*) ((uintptr_t) a0 - kc); |
|
a1 = (const bfloat16_t*) ((uintptr_t) a1 - kc); |
|
a2 = (const bfloat16_t*) ((uintptr_t) a2 - kc); |
|
a3 = (const bfloat16_t*) ((uintptr_t) a3 - kc); |
|
a4 = (const bfloat16_t*) ((uintptr_t) a4 - kc); |
|
a5 = (const bfloat16_t*) ((uintptr_t) a5 - kc); |
|
|
|
nc -= 8; |
|
} else { |
|
if (nc & 4) { |
|
vst1_bf16(c0, vout0x0123); c0 += 4; |
|
vst1_bf16(c1, vout1x0123); c1 += 4; |
|
vst1_bf16(c2, vout2x0123); c2 += 4; |
|
vst1_bf16(c3, vout3x0123); c3 += 4; |
|
vst1_bf16(c4, vout4x0123); c4 += 4; |
|
vst1_bf16(c5, vout5x0123); c5 += 4; |
|
|
|
vout0x0123 = vout0x4567; |
|
vout1x0123 = vout1x4567; |
|
vout2x0123 = vout2x4567; |
|
vout3x0123 = vout3x4567; |
|
vout4x0123 = vout4x4567; |
|
vout5x0123 = vout5x4567; |
|
} |
|
if (nc & 2) { |
|
vst1_lane_u32((void*) c0, vreinterpret_u32_bf16(vout0x0123), 0); c0 += 2; |
|
vst1_lane_u32((void*) c1, vreinterpret_u32_bf16(vout1x0123), 0); c1 += 2; |
|
vst1_lane_u32((void*) c2, vreinterpret_u32_bf16(vout2x0123), 0); c2 += 2; |
|
vst1_lane_u32((void*) c3, vreinterpret_u32_bf16(vout3x0123), 0); c3 += 2; |
|
vst1_lane_u32((void*) c4, vreinterpret_u32_bf16(vout4x0123), 0); c4 += 2; |
|
vst1_lane_u32((void*) c5, vreinterpret_u32_bf16(vout5x0123), 0); c5 += 2; |
|
|
|
vout0x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout0x0123), vreinterpret_u16_bf16(vout0x0123), 2)); |
|
vout1x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout1x0123), vreinterpret_u16_bf16(vout1x0123), 2)); |
|
vout2x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout2x0123), vreinterpret_u16_bf16(vout2x0123), 2)); |
|
vout3x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout3x0123), vreinterpret_u16_bf16(vout3x0123), 2)); |
|
vout4x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout4x0123), vreinterpret_u16_bf16(vout4x0123), 2)); |
|
vout5x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout5x0123), vreinterpret_u16_bf16(vout5x0123), 2)); |
|
} |
|
if (nc & 1) { |
|
vst1_lane_bf16(c0, vout0x0123, 0); |
|
vst1_lane_bf16(c1, vout1x0123, 0); |
|
vst1_lane_bf16(c2, vout2x0123, 0); |
|
vst1_lane_bf16(c3, vout3x0123, 0); |
|
vst1_lane_bf16(c4, vout4x0123, 0); |
|
vst1_lane_bf16(c5, vout5x0123, 0); |
|
} |
|
|
|
nc = 0; |
|
} |
|
} while (nc != 0); |
|
} |
|
|
