// Auto-generated file. Do not edit! // Template: src/f32-igemm/neon-ld128.c.in // Generator: tools/xngen // // Copyright 2019 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include "xnnpack/igemm.h" void xnn_f32_igemm_minmax_ukernel_2x16__neon_lane_ld128( size_t mr, size_t nc, size_t kc, size_t ks, const float** restrict a, const float* restrict w, float* restrict c, size_t cm_stride, size_t cn_stride, size_t a_offset, const float* zero, const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) { assert(mr != 0); assert(mr <= 2); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(float) == 0); assert(ks != 0); assert(ks % (2 * sizeof(void*)) == 0); assert(a_offset % sizeof(float) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); float* c0 = c; float* c1 = (float*) ((uintptr_t) c0 + cm_stride); if XNN_UNPREDICTABLE(mr != 2) { c1 = c0; } do { float32x4_t vacc0x0 = vld1q_f32(w); w += 4; float32x4_t vacc0x1 = vld1q_f32(w); w += 4; float32x4_t vacc0x2 = vld1q_f32(w); w += 4; float32x4_t vacc0x3 = vld1q_f32(w); w += 4; float32x4_t vacc1x0 = vacc0x0; float32x4_t vacc1x1 = vacc0x1; float32x4_t vacc1x2 = vacc0x2; float32x4_t vacc1x3 = vacc0x3; size_t p = ks; do { const float* restrict a0 = a[0]; assert(a0 != NULL); if XNN_UNPREDICTABLE(a0 != zero) { a0 = (const float*) ((uintptr_t) a0 + a_offset); } const float* restrict a1 = a[1]; assert(a1 != NULL); if XNN_UNPREDICTABLE(a1 != zero) { a1 = (const float*) ((uintptr_t) a1 + a_offset); } a += 2; size_t k = kc; for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) { const float32x4_t va0 = vld1q_f32(a0); a0 += 4; const float32x4_t va1 = vld1q_f32(a1); a1 += 4; const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; const float32x4_t vb89ABc0 = vld1q_f32(w); w += 4; const float32x4_t vbCDEFc0 = vld1q_f32(w); w += 4; vacc0x0 = vmlaq_lane_f32(vacc0x0, vb0123c0, vget_low_f32(va0), 0); vacc1x0 = vmlaq_lane_f32(vacc1x0, vb0123c0, vget_low_f32(va1), 0); vacc0x1 = vmlaq_lane_f32(vacc0x1, vb4567c0, vget_low_f32(va0), 0); vacc1x1 = vmlaq_lane_f32(vacc1x1, vb4567c0, vget_low_f32(va1), 0); vacc0x2 = vmlaq_lane_f32(vacc0x2, vb89ABc0, vget_low_f32(va0), 0); vacc1x2 = vmlaq_lane_f32(vacc1x2, vb89ABc0, vget_low_f32(va1), 0); vacc0x3 = vmlaq_lane_f32(vacc0x3, vbCDEFc0, vget_low_f32(va0), 0); vacc1x3 = vmlaq_lane_f32(vacc1x3, vbCDEFc0, vget_low_f32(va1), 0); const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; const float32x4_t vb4567c1 = vld1q_f32(w); w += 4; const float32x4_t vb89ABc1 = vld1q_f32(w); w += 4; const float32x4_t vbCDEFc1 = vld1q_f32(w); w += 4; vacc0x0 = vmlaq_lane_f32(vacc0x0, vb0123c1, vget_low_f32(va0), 1); vacc1x0 = vmlaq_lane_f32(vacc1x0, vb0123c1, vget_low_f32(va1), 1); vacc0x1 = vmlaq_lane_f32(vacc0x1, vb4567c1, vget_low_f32(va0), 1); vacc1x1 = vmlaq_lane_f32(vacc1x1, vb4567c1, vget_low_f32(va1), 1); vacc0x2 = vmlaq_lane_f32(vacc0x2, vb89ABc1, vget_low_f32(va0), 1); vacc1x2 = vmlaq_lane_f32(vacc1x2, vb89ABc1, vget_low_f32(va1), 1); vacc0x3 = vmlaq_lane_f32(vacc0x3, vbCDEFc1, vget_low_f32(va0), 1); vacc1x3 = vmlaq_lane_f32(vacc1x3, vbCDEFc1, vget_low_f32(va1), 1); const float32x4_t vb0123c2 = vld1q_f32(w); w += 4; const float32x4_t vb4567c2 = vld1q_f32(w); w += 4; const float32x4_t vb89ABc2 = vld1q_f32(w); w += 4; const float32x4_t vbCDEFc2 = vld1q_f32(w); w += 4; vacc0x0 = vmlaq_lane_f32(vacc0x0, vb0123c2, vget_high_f32(va0), 0); vacc1x0 = vmlaq_lane_f32(vacc1x0, vb0123c2, vget_high_f32(va1), 0); vacc0x1 = vmlaq_lane_f32(vacc0x1, vb4567c2, vget_high_f32(va0), 0); vacc1x1 = vmlaq_lane_f32(vacc1x1, vb4567c2, vget_high_f32(va1), 0); vacc0x2 = vmlaq_lane_f32(vacc0x2, vb89ABc2, vget_high_f32(va0), 0); vacc1x2 = vmlaq_lane_f32(vacc1x2, vb89ABc2, vget_high_f32(va1), 0); vacc0x3 = vmlaq_lane_f32(vacc0x3, vbCDEFc2, vget_high_f32(va0), 0); vacc1x3 = vmlaq_lane_f32(vacc1x3, vbCDEFc2, vget_high_f32(va1), 0); const float32x4_t vb0123c3 = vld1q_f32(w); w += 4; const float32x4_t vb4567c3 = vld1q_f32(w); w += 4; const float32x4_t vb89ABc3 = vld1q_f32(w); w += 4; const float32x4_t vbCDEFc3 = vld1q_f32(w); w += 4; vacc0x0 = vmlaq_lane_f32(vacc0x0, vb0123c3, vget_high_f32(va0), 1); vacc1x0 = vmlaq_lane_f32(vacc1x0, vb0123c3, vget_high_f32(va1), 1); vacc0x1 = vmlaq_lane_f32(vacc0x1, vb4567c3, vget_high_f32(va0), 1); vacc1x1 = vmlaq_lane_f32(vacc1x1, vb4567c3, vget_high_f32(va1), 1); vacc0x2 = vmlaq_lane_f32(vacc0x2, vb89ABc3, vget_high_f32(va0), 1); vacc1x2 = vmlaq_lane_f32(vacc1x2, vb89ABc3, vget_high_f32(va1), 1); vacc0x3 = vmlaq_lane_f32(vacc0x3, vbCDEFc3, vget_high_f32(va0), 1); vacc1x3 = vmlaq_lane_f32(vacc1x3, vbCDEFc3, vget_high_f32(va1), 1); } if XNN_UNLIKELY(k != 0) { do { const float32x4_t va0 = vld1q_dup_f32(a0); a0 += 1; const float32x4_t va1 = vld1q_dup_f32(a1); a1 += 1; const float32x4_t vb0123 = vld1q_f32(w); w += 4; const float32x4_t vb4567 = vld1q_f32(w); w += 4; const float32x4_t vb89AB = vld1q_f32(w); w += 4; const float32x4_t vbCDEF = vld1q_f32(w); w += 4; vacc0x0 = vmlaq_f32(vacc0x0, va0, vb0123); vacc1x0 = vmlaq_f32(vacc1x0, va1, vb0123); vacc0x1 = vmlaq_f32(vacc0x1, va0, vb4567); vacc1x1 = vmlaq_f32(vacc1x1, va1, vb4567); vacc0x2 = vmlaq_f32(vacc0x2, va0, vb89AB); vacc1x2 = vmlaq_f32(vacc1x2, va1, vb89AB); vacc0x3 = vmlaq_f32(vacc0x3, va0, vbCDEF); vacc1x3 = vmlaq_f32(vacc1x3, va1, vbCDEF); k -= sizeof(float); } while (k != 0); } p -= 2 * sizeof(void*); } while (p != 0); const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); vacc0x0 = vminq_f32(vacc0x0, vmax); vacc1x0 = vminq_f32(vacc1x0, vmax); vacc0x1 = vminq_f32(vacc0x1, vmax); vacc1x1 = vminq_f32(vacc1x1, vmax); vacc0x2 = vminq_f32(vacc0x2, vmax); vacc1x2 = vminq_f32(vacc1x2, vmax); vacc0x3 = vminq_f32(vacc0x3, vmax); vacc1x3 = vminq_f32(vacc1x3, vmax); const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); vacc0x0 = vmaxq_f32(vacc0x0, vmin); vacc1x0 = vmaxq_f32(vacc1x0, vmin); vacc0x1 = vmaxq_f32(vacc0x1, vmin); vacc1x1 = vmaxq_f32(vacc1x1, vmin); vacc0x2 = vmaxq_f32(vacc0x2, vmin); vacc1x2 = vmaxq_f32(vacc1x2, vmin); vacc0x3 = vmaxq_f32(vacc0x3, vmin); vacc1x3 = vmaxq_f32(vacc1x3, vmin); if XNN_LIKELY(nc >= 16) { vst1q_f32(c1, vacc1x0); vst1q_f32(c1 + 4, vacc1x1); vst1q_f32(c1 + 8, vacc1x2); vst1q_f32(c1 + 12, vacc1x3); c1 = (float*) ((uintptr_t) c1 + cn_stride); vst1q_f32(c0, vacc0x0); vst1q_f32(c0 + 4, vacc0x1); vst1q_f32(c0 + 8, vacc0x2); vst1q_f32(c0 + 12, vacc0x3); c0 = (float*) ((uintptr_t) c0 + cn_stride); a = (const float**restrict) ((uintptr_t) a - ks); nc -= 16; } else { if (nc & 8) { vst1q_f32(c1, vacc1x0); c1 += 4; vst1q_f32(c0, vacc0x0); c0 += 4; vst1q_f32(c1, vacc1x1); c1 += 4; vst1q_f32(c0, vacc0x1); c0 += 4; vacc1x0 = vacc1x2; vacc1x1 = vacc1x3; vacc0x0 = vacc0x2; vacc0x1 = vacc0x3; } if (nc & 4) { vst1q_f32(c1, vacc1x0); c1 += 4; vst1q_f32(c0, vacc0x0); c0 += 4; vacc1x0 = vacc1x1; vacc1x1 = vacc1x2; vacc1x2 = vacc1x3; vacc0x0 = vacc0x1; vacc0x1 = vacc0x2; vacc0x2 = vacc0x3; } float32x2_t vacc1 = vget_low_f32(vacc1x0); float32x2_t vacc0 = vget_low_f32(vacc0x0); if (nc & 2) { vst1_f32(c1, vacc1); c1 += 2; vst1_f32(c0, vacc0); c0 += 2; vacc1 = vget_high_f32(vacc1x0); vacc0 = vget_high_f32(vacc0x0); } if (nc & 1) { vst1_lane_f32(c1, vacc1, 0); vst1_lane_f32(c0, vacc0, 0); } nc = 0; } } while (nc != 0); }