|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include <assert.h> |
|
|
|
#include <arm_neon.h> |
|
|
|
#include <xnnpack/gemm.h> |
|
|
|
|
|
void xnn_bf16_gemm_minmax_ukernel_1x8c2__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 <= 1); |
|
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* 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; |
|
|
|
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 vb0123c01 = vld1q_bf16(w); w += 8; |
|
const bfloat16x8_t vb4567c01 = vld1q_bf16(w); w += 8; |
|
|
|
vacc0x0123 = vbfdotq_laneq_f32(vacc0x0123, vb0123c01, va0, 0); |
|
vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c01, va0, 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); |
|
vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c23, va0, 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); |
|
vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c45, va0, 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); |
|
vacc0x4567 = vbfdotq_laneq_f32(vacc0x4567, vb4567c67, va0, 3); |
|
} |
|
if XNN_UNLIKELY(k != 0) { |
|
const bfloat16x8_t va0 = vld1q_bf16(a0); a0 = (const bfloat16_t*) ((uintptr_t) a0 + 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 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 va0x4567c01 = vbicq_u32(va0c01, vm4567c01); |
|
vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c01, vreinterpretq_bf16_u32(va0x4567c01)); |
|
|
|
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 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 va0x4567c23 = vbicq_u32(va0c23, vm4567c23); |
|
vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c23, vreinterpretq_bf16_u32(va0x4567c23)); |
|
|
|
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 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 va0x4567c45 = vbicq_u32(va0c45, vm4567c45); |
|
vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c45, vreinterpretq_bf16_u32(va0x4567c45)); |
|
|
|
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 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 va0x4567c67 = vbicq_u32(va0c67, vm4567c67); |
|
vacc0x4567 = vbfdotq_f32(vacc0x4567, vb4567c67, vreinterpretq_bf16_u32(va0x4567c67)); |
|
} |
|
} |
|
} |
|
} |
|
|
|
const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); |
|
vacc0x0123 = vminq_f32(vacc0x0123, vmax); |
|
vacc0x4567 = vminq_f32(vacc0x4567, vmax); |
|
|
|
const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); |
|
vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); |
|
vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); |
|
|
|
bfloat16x4_t vout0x0123 = vcvt_bf16_f32(vacc0x0123); |
|
bfloat16x4_t vout0x4567 = vcvt_bf16_f32(vacc0x4567); |
|
|
|
if XNN_LIKELY(nc >= 8) { |
|
vst1_bf16(c0, vout0x0123); |
|
vst1_bf16(c0 + 4, vout0x4567); |
|
c0 = (bfloat16_t*) ((uintptr_t) c0 + cn_stride); |
|
|
|
a0 = (const bfloat16_t*) ((uintptr_t) a0 - kc); |
|
|
|
nc -= 8; |
|
} else { |
|
if (nc & 4) { |
|
vst1_bf16(c0, vout0x0123); c0 += 4; |
|
|
|
vout0x0123 = vout0x4567; |
|
} |
|
if (nc & 2) { |
|
vst1_lane_u32((void*) c0, vreinterpret_u32_bf16(vout0x0123), 0); c0 += 2; |
|
|
|
vout0x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout0x0123), vreinterpret_u16_bf16(vout0x0123), 2)); |
|
} |
|
if (nc & 1) { |
|
vst1_lane_bf16(c0, vout0x0123, 0); |
|
} |
|
|
|
nc = 0; |
|
} |
|
} while (nc != 0); |
|
} |
|
|