|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include <assert.h> |
|
|
|
#include <arm_neon.h> |
|
|
|
#include <xnnpack/gemm.h> |
|
|
|
|
|
void xnn_bf16_gemm_minmax_ukernel_2x4c8__neonfma_zip( |
|
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 <= 2); |
|
assert(nc != 0); |
|
assert(kc != 0); |
|
assert(kc % sizeof(uint16_t) == 0); |
|
assert(a != NULL); |
|
assert(w_ptr != NULL); |
|
assert(c != NULL); |
|
|
|
const uint16_t* a0 = (const uint16_t*) a; |
|
uint16_t* c0 = (uint16_t*) c; |
|
const uint16_t* a1 = (const uint16_t*) ((uintptr_t) a0 + a_stride); |
|
uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride); |
|
if XNN_UNPREDICTABLE(mr != 2) { |
|
a1 = a0; |
|
c1 = c0; |
|
} |
|
|
|
const uint16_t* w = (const uint16_t*) w_ptr; |
|
const uint16x8_t vzero = vmovq_n_u16(0); |
|
do { |
|
float32x4_t vacc0x0 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1; |
|
float32x4_t vacc0x1 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1; |
|
float32x4_t vacc0x2 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1; |
|
float32x4_t vacc0x3 = vreinterpretq_f32_u32(vshll_n_u16(vld1_lane_u16(w, vdup_n_u16(0), 0), 16)); w += 1; |
|
float32x4_t vacc1x0 = vacc0x0; |
|
float32x4_t vacc1x1 = vacc0x1; |
|
float32x4_t vacc1x2 = vacc0x2; |
|
float32x4_t vacc1x3 = vacc0x3; |
|
|
|
size_t k = kc; |
|
for (; k >= 8 * sizeof(uint16_t); k -= 8 * sizeof(uint16_t)) { |
|
const uint16x8_t va0h = vld1q_u16(a0); a0 += 8; |
|
const uint16x8_t va1h = vld1q_u16(a1); a1 += 8; |
|
|
|
const uint16x8_t vb0h = vld1q_u16(w); w += 8; |
|
const uint16x8_t vb1h = vld1q_u16(w); w += 8; |
|
const uint16x8_t vb2h = vld1q_u16(w); w += 8; |
|
const uint16x8_t vb3h = vld1q_u16(w); w += 8; |
|
|
|
const uint16x8x2_t va0f = vzipq_u16(vzero, va0h); |
|
const uint16x8x2_t va1f = vzipq_u16(vzero, va1h); |
|
|
|
const uint16x8x2_t vb0f = vzipq_u16(vzero, vb0h); |
|
const uint16x8x2_t vb1f = vzipq_u16(vzero, vb1h); |
|
const uint16x8x2_t vb2f = vzipq_u16(vzero, vb2h); |
|
const uint16x8x2_t vb3f = vzipq_u16(vzero, vb3h); |
|
|
|
vacc0x0 = vfmaq_f32(vacc0x0, vreinterpretq_f32_u16(va0f.val[0]), vreinterpretq_f32_u16(vb0f.val[0])); |
|
vacc1x0 = vfmaq_f32(vacc1x0, vreinterpretq_f32_u16(va1f.val[0]), vreinterpretq_f32_u16(vb0f.val[0])); |
|
vacc0x1 = vfmaq_f32(vacc0x1, vreinterpretq_f32_u16(va0f.val[0]), vreinterpretq_f32_u16(vb1f.val[0])); |
|
vacc1x1 = vfmaq_f32(vacc1x1, vreinterpretq_f32_u16(va1f.val[0]), vreinterpretq_f32_u16(vb1f.val[0])); |
|
vacc0x2 = vfmaq_f32(vacc0x2, vreinterpretq_f32_u16(va0f.val[0]), vreinterpretq_f32_u16(vb2f.val[0])); |
|
vacc1x2 = vfmaq_f32(vacc1x2, vreinterpretq_f32_u16(va1f.val[0]), vreinterpretq_f32_u16(vb2f.val[0])); |
|
vacc0x3 = vfmaq_f32(vacc0x3, vreinterpretq_f32_u16(va0f.val[0]), vreinterpretq_f32_u16(vb3f.val[0])); |
|
vacc1x3 = vfmaq_f32(vacc1x3, vreinterpretq_f32_u16(va1f.val[0]), vreinterpretq_f32_u16(vb3f.val[0])); |
|
|
|
vacc0x0 = vfmaq_f32(vacc0x0, vreinterpretq_f32_u16(va0f.val[1]), vreinterpretq_f32_u16(vb0f.val[1])); |
|
vacc1x0 = vfmaq_f32(vacc1x0, vreinterpretq_f32_u16(va1f.val[1]), vreinterpretq_f32_u16(vb0f.val[1])); |
|
vacc0x1 = vfmaq_f32(vacc0x1, vreinterpretq_f32_u16(va0f.val[1]), vreinterpretq_f32_u16(vb1f.val[1])); |
|
vacc1x1 = vfmaq_f32(vacc1x1, vreinterpretq_f32_u16(va1f.val[1]), vreinterpretq_f32_u16(vb1f.val[1])); |
|
vacc0x2 = vfmaq_f32(vacc0x2, vreinterpretq_f32_u16(va0f.val[1]), vreinterpretq_f32_u16(vb2f.val[1])); |
|
vacc1x2 = vfmaq_f32(vacc1x2, vreinterpretq_f32_u16(va1f.val[1]), vreinterpretq_f32_u16(vb2f.val[1])); |
|
vacc0x3 = vfmaq_f32(vacc0x3, vreinterpretq_f32_u16(va0f.val[1]), vreinterpretq_f32_u16(vb3f.val[1])); |
|
vacc1x3 = vfmaq_f32(vacc1x3, vreinterpretq_f32_u16(va1f.val[1]), vreinterpretq_f32_u16(vb3f.val[1])); |
|
} |
|
if XNN_UNLIKELY(k != 0) { |
|
const uint16x8_t va0h = vld1q_u16(a0); a0 = (const uint16_t*) ((uintptr_t) a0 + k); |
|
const uint16x8_t va1h = vld1q_u16(a1); a1 = (const uint16_t*) ((uintptr_t) a1 + k); |
|
|
|
const uint16x8_t vb0h = vld1q_u16(w); w += 8; |
|
const uint16x8_t vb1h = vld1q_u16(w); w += 8; |
|
const uint16x8_t vb2h = vld1q_u16(w); w += 8; |
|
const uint16x8_t vb3h = vld1q_u16(w); w += 8; |
|
|
|
const uint16x8_t vm0h = vceqq_u16(vb0h, vmovq_n_u16(0)); |
|
const uint16x8_t vm1h = vceqq_u16(vb1h, vmovq_n_u16(0)); |
|
const uint16x8_t vm2h = vceqq_u16(vb2h, vmovq_n_u16(0)); |
|
const uint16x8_t vm3h = vceqq_u16(vb3h, vmovq_n_u16(0)); |
|
|
|
const uint16x8x2_t vb0f = vzipq_u16(vzero, vb0h); |
|
const uint16x8x2_t vb1f = vzipq_u16(vzero, vb1h); |
|
const uint16x8x2_t vb2f = vzipq_u16(vzero, vb2h); |
|
const uint16x8x2_t vb3f = vzipq_u16(vzero, vb3h); |
|
|
|
const uint16x8_t va0x0h = vbicq_u16(va0h, vm0h); |
|
const uint16x8_t va1x0h = vbicq_u16(va1h, vm0h); |
|
const uint16x8_t va0x1h = vbicq_u16(va0h, vm1h); |
|
const uint16x8_t va1x1h = vbicq_u16(va1h, vm1h); |
|
const uint16x8_t va0x2h = vbicq_u16(va0h, vm2h); |
|
const uint16x8_t va1x2h = vbicq_u16(va1h, vm2h); |
|
const uint16x8_t va0x3h = vbicq_u16(va0h, vm3h); |
|
const uint16x8_t va1x3h = vbicq_u16(va1h, vm3h); |
|
|
|
const uint16x8x2_t va0x0f = vzipq_u16(vzero, va0x0h); |
|
const uint16x8x2_t va1x0f = vzipq_u16(vzero, va1x0h); |
|
const uint16x8x2_t va0x1f = vzipq_u16(vzero, va0x1h); |
|
const uint16x8x2_t va1x1f = vzipq_u16(vzero, va1x1h); |
|
const uint16x8x2_t va0x2f = vzipq_u16(vzero, va0x2h); |
|
const uint16x8x2_t va1x2f = vzipq_u16(vzero, va1x2h); |
|
const uint16x8x2_t va0x3f = vzipq_u16(vzero, va0x3h); |
|
const uint16x8x2_t va1x3f = vzipq_u16(vzero, va1x3h); |
|
|
|
vacc0x0 = vfmaq_f32(vacc0x0, vreinterpretq_f32_u16(va0x0f.val[0]), vreinterpretq_f32_u16(vb0f.val[0])); |
|
vacc1x0 = vfmaq_f32(vacc1x0, vreinterpretq_f32_u16(va1x0f.val[0]), vreinterpretq_f32_u16(vb0f.val[0])); |
|
vacc0x1 = vfmaq_f32(vacc0x1, vreinterpretq_f32_u16(va0x1f.val[0]), vreinterpretq_f32_u16(vb1f.val[0])); |
|
vacc1x1 = vfmaq_f32(vacc1x1, vreinterpretq_f32_u16(va1x1f.val[0]), vreinterpretq_f32_u16(vb1f.val[0])); |
|
vacc0x2 = vfmaq_f32(vacc0x2, vreinterpretq_f32_u16(va0x2f.val[0]), vreinterpretq_f32_u16(vb2f.val[0])); |
|
vacc1x2 = vfmaq_f32(vacc1x2, vreinterpretq_f32_u16(va1x2f.val[0]), vreinterpretq_f32_u16(vb2f.val[0])); |
|
vacc0x3 = vfmaq_f32(vacc0x3, vreinterpretq_f32_u16(va0x3f.val[0]), vreinterpretq_f32_u16(vb3f.val[0])); |
|
vacc1x3 = vfmaq_f32(vacc1x3, vreinterpretq_f32_u16(va1x3f.val[0]), vreinterpretq_f32_u16(vb3f.val[0])); |
|
|
|
vacc0x0 = vfmaq_f32(vacc0x0, vreinterpretq_f32_u16(va0x0f.val[1]), vreinterpretq_f32_u16(vb0f.val[1])); |
|
vacc1x0 = vfmaq_f32(vacc1x0, vreinterpretq_f32_u16(va1x0f.val[1]), vreinterpretq_f32_u16(vb0f.val[1])); |
|
vacc0x1 = vfmaq_f32(vacc0x1, vreinterpretq_f32_u16(va0x1f.val[1]), vreinterpretq_f32_u16(vb1f.val[1])); |
|
vacc1x1 = vfmaq_f32(vacc1x1, vreinterpretq_f32_u16(va1x1f.val[1]), vreinterpretq_f32_u16(vb1f.val[1])); |
|
vacc0x2 = vfmaq_f32(vacc0x2, vreinterpretq_f32_u16(va0x2f.val[1]), vreinterpretq_f32_u16(vb2f.val[1])); |
|
vacc1x2 = vfmaq_f32(vacc1x2, vreinterpretq_f32_u16(va1x2f.val[1]), vreinterpretq_f32_u16(vb2f.val[1])); |
|
vacc0x3 = vfmaq_f32(vacc0x3, vreinterpretq_f32_u16(va0x3f.val[1]), vreinterpretq_f32_u16(vb3f.val[1])); |
|
vacc1x3 = vfmaq_f32(vacc1x3, vreinterpretq_f32_u16(va1x3f.val[1]), vreinterpretq_f32_u16(vb3f.val[1])); |
|
} |
|
|
|
#if XNN_ARCH_ARM64 |
|
const float32x4_t vacc0x01 = vpaddq_f32(vacc0x0, vacc0x1); |
|
const float32x4_t vacc1x01 = vpaddq_f32(vacc1x0, vacc1x1); |
|
const float32x4_t vacc0x23 = vpaddq_f32(vacc0x2, vacc0x3); |
|
const float32x4_t vacc1x23 = vpaddq_f32(vacc1x2, vacc1x3); |
|
|
|
float32x4_t vacc0x0123 = vpaddq_f32(vacc0x01, vacc0x23); |
|
float32x4_t vacc1x0123 = vpaddq_f32(vacc1x01, vacc1x23); |
|
#else |
|
const float32x2_t vsum0x0 = vadd_f32(vget_low_f32(vacc0x0), vget_high_f32(vacc0x0)); |
|
const float32x2_t vsum1x0 = vadd_f32(vget_low_f32(vacc1x0), vget_high_f32(vacc1x0)); |
|
const float32x2_t vsum0x1 = vadd_f32(vget_low_f32(vacc0x1), vget_high_f32(vacc0x1)); |
|
const float32x2_t vsum1x1 = vadd_f32(vget_low_f32(vacc1x1), vget_high_f32(vacc1x1)); |
|
const float32x2_t vsum0x2 = vadd_f32(vget_low_f32(vacc0x2), vget_high_f32(vacc0x2)); |
|
const float32x2_t vsum1x2 = vadd_f32(vget_low_f32(vacc1x2), vget_high_f32(vacc1x2)); |
|
const float32x2_t vsum0x3 = vadd_f32(vget_low_f32(vacc0x3), vget_high_f32(vacc0x3)); |
|
const float32x2_t vsum1x3 = vadd_f32(vget_low_f32(vacc1x3), vget_high_f32(vacc1x3)); |
|
|
|
float32x4_t vacc0x0123 = vcombine_f32(vpadd_f32(vsum0x0, vsum0x1), vpadd_f32(vsum0x2, vsum0x3)); |
|
float32x4_t vacc1x0123 = vcombine_f32(vpadd_f32(vsum1x0, vsum1x1), vpadd_f32(vsum1x2, vsum1x3)); |
|
#endif |
|
|
|
const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); |
|
vacc0x0123 = vminq_f32(vacc0x0123, vmax); |
|
vacc1x0123 = vminq_f32(vacc1x0123, vmax); |
|
|
|
const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); |
|
vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); |
|
vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); |
|
|
|
uint16x4_t vout0x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc0x0123), 16); |
|
uint16x4_t vout1x0123 = vshrn_n_u32(vreinterpretq_u32_f32(vacc1x0123), 16); |
|
|
|
if XNN_LIKELY(nc >= 4) { |
|
vst1_u16(c0, vout0x0123); |
|
c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride); |
|
vst1_u16(c1, vout1x0123); |
|
c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride); |
|
|
|
a0 = (const uint16_t*) ((uintptr_t) a0 - kc); |
|
a1 = (const uint16_t*) ((uintptr_t) a1 - kc); |
|
|
|
nc -= 4; |
|
} else { |
|
if (nc & 2) { |
|
vst1_lane_u32((void*) c0, vreinterpret_u32_u16(vout0x0123), 0); c0 += 2; |
|
vst1_lane_u32((void*) c1, vreinterpret_u32_u16(vout1x0123), 0); c1 += 2; |
|
|
|
vout0x0123 = vext_u16(vout0x0123, vout0x0123, 2); |
|
vout1x0123 = vext_u16(vout1x0123, vout1x0123, 2); |
|
} |
|
if (nc & 1) { |
|
vst1_lane_u16(c0, vout0x0123, 0); |
|
vst1_lane_u16(c1, vout1x0123, 0); |
|
} |
|
|
|
nc = 0; |
|
} |
|
} while (nc != 0); |
|
} |
|
|