src/f32-raddstoreexpminusmax/neon-p5.c.in - platform/external/XNNPACK - Gitiles

 // Copyright 2020 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.

 $assert ELEMENTS_TILE % 4 == 0
 $assert ELEMENTS_TILE >= 4
 $SIMD_TILE = ELEMENTS_TILE // 4
 $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
 $VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
 #include <assert.h>

 #include <arm_neon.h>

 #include <xnnpack/common.h>
 #include <xnnpack/raddstoreexpminusmax.h>


 $PARAMS_STRUCT = "neonfma_rr1_p5" if FMA else "neon_rr2_p5"
 void xnn_f32_raddstoreexpminusmax_ukernel__${"neonfma" if FMA else "neon"}_rr${1 if FMA else 2}_p5_x${ELEMENTS_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
     size_t elements,
     const float* input,
     const float* max,
     float* output,
     float* sum,
     const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
 {
   assert(elements % sizeof(float) == 0);

   const float32x4_t vi_max = vld1q_dup_f32(max);
   const float32x4_t vlog2e = vld1q_dup_f32(&params->${PARAMS_STRUCT}.log2e);
   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->${PARAMS_STRUCT}.magic_bias);
   $if FMA:
     const float32x4_t vminus_ln2 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.minus_ln2);
   $else:
     const float32x4_t vminus_ln2_hi = vld1q_dup_f32(&params->${PARAMS_STRUCT}.minus_ln2_hi);
     const float32x4_t vminus_ln2_lo = vld1q_dup_f32(&params->${PARAMS_STRUCT}.minus_ln2_lo);
   const float32x4_t vc5 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c5);
   const float32x4_t vc4 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c4);
   const float32x4_t vc3 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c3);
   const float32x4_t vc2 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c2);
   const float32x4_t vc1 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c1);
   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->${PARAMS_STRUCT}.denorm_cutoff);

   $if ELEMENTS_TILE > 4:
     $for K in range(ACCUMULATORS):
       float32x4_t vacc${K} = vmovq_n_f32(0.0f);
     for (; elements >= ${ELEMENTS_TILE} * sizeof(float); elements -= ${ELEMENTS_TILE} * sizeof(float)) {
       $for N in range(0, ELEMENTS_TILE, 4):
         const float32x4_t vi${ABC[N:N+4]} = vld1q_f32(input); input += 4;

       $for N in range(0, ELEMENTS_TILE, 4):
         const float32x4_t vx${ABC[N:N+4]} = vsubq_f32(vi${ABC[N:N+4]}, vi_max);

       $for N in range(0, ELEMENTS_TILE, 4):
         float32x4_t vn${ABC[N:N+4]} = ${VMULADDQ_F32}(vmagic_bias, vx${ABC[N:N+4]}, vlog2e);

       $for N in range(0, ELEMENTS_TILE, 4):
         const float32x4_t vs${ABC[N:N+4]} = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), 23));

       $for N in range(0, ELEMENTS_TILE, 4):
         vn${ABC[N:N+4]} = vsubq_f32(vn${ABC[N:N+4]}, vmagic_bias);

       $if FMA:
         $for N in range(0, ELEMENTS_TILE, 4):
           float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vx${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2);
       $else:
         $for N in range(0, ELEMENTS_TILE, 4):
           float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vx${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_hi);

         $for N in range(0, ELEMENTS_TILE, 4):
           vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vt${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_lo);

       $for N in range(0, ELEMENTS_TILE, 4):
         float32x4_t vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc4, vc5, vt${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc3, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc2, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc1, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         vt${ABC[N:N+4]} = vmulq_f32(vt${ABC[N:N+4]}, vs${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         float32x4_t vf${ABC[N:N+4]} = ${VMULADDQ_F32}(vs${ABC[N:N+4]}, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

       $for N in range(0, ELEMENTS_TILE, 4):
         vf${ABC[N:N+4]} = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf${ABC[N:N+4]}), vcltq_f32(vx${ABC[N:N+4]}, vdenorm_cutoff)));

       $for N in range(0, ELEMENTS_TILE, 4):
         vst1q_f32(output, vf${ABC[N:N+4]}); output += 4;

       $for N in range(0, ELEMENTS_TILE, 4):
         vacc${N % ACCUMULATORS} = vaddq_f32(vacc${N % ACCUMULATORS}, vf${ABC[N:N+4]});
     }
     $if ACCUMULATORS > 1:
       $ACC_SLICE = 1
       $while ACC_SLICE < ACCUMULATORS:
         $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
           $if A + ACC_SLICE < ACCUMULATORS:
             vacc${A} = vaddq_f32(vacc${A}, vacc${A + ACC_SLICE});
         $ACC_SLICE *= 2

     float32x4_t vacc = vacc0;
   $else:
     float32x4_t vacc = vmovq_n_f32(0.0f);
   for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
     const float32x4_t vi = vld1q_f32(input); input += 4;

     const float32x4_t vx = vsubq_f32(vi, vi_max);

     float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

     vn = vsubq_f32(vn, vmagic_bias);

     $if FMA:
       float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2);
     $else:
       float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
       vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

     float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
     vp = ${VMULADDQ_F32}(vc3, vp, vt);
     vp = ${VMULADDQ_F32}(vc2, vp, vt);
     vp = ${VMULADDQ_F32}(vc1, vp, vt);

     vt = vmulq_f32(vt, vs);
     float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

     vst1q_f32(output, vf); output += 4;

     vacc = vaddq_f32(vacc, vf);
   }
 #if XNN_ARCH_ARM64
   float vacc_lo = vaddvq_f32(vacc);
 #else
   float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
 #endif
   if (elements != 0) {
     assert(elements >= 1 * sizeof(float));
     assert(elements <= 3 * sizeof(float));
     const float32x4_t vi = vld1q_f32(input); input += 4;

     const float32x4_t vx = vsubq_f32(vi, vi_max);

     float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

     vn = vsubq_f32(vn, vmagic_bias);

     $if FMA:
       float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2);
     $else:
       float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
       vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

     float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
     vp = ${VMULADDQ_F32}(vc3, vp, vt);
     vp = ${VMULADDQ_F32}(vc2, vp, vt);
     vp = ${VMULADDQ_F32}(vc1, vp, vt);

     vt = vmulq_f32(vt, vs);
     float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

     float32x2_t vf_lo = vget_low_f32(vf);
     if (elements & (2 * sizeof(float))) {
       vst1_f32(output, vf_lo); output += 2;

       #if XNN_ARCH_ARM64
         vacc_lo += vaddv_f32(vf_lo);
       #else
         vacc_lo = vadd_f32(vacc_lo, vf_lo);
       #endif

       vf_lo = vget_high_f32(vf);
     }
     if (elements & (1 * sizeof(float))) {
       vst1_lane_f32(output, vf_lo, 0);

       #if XNN_ARCH_ARM64
         vacc_lo += vget_lane_f32(vf_lo, 0);
       #else
         vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
       #endif
     }
   }
 #if XNN_ARCH_ARM64
   *sum = vacc_lo;
 #else
   vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
 #endif
 }
	// Copyright 2020 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.

	$assert ELEMENTS_TILE % 4 == 0
	$assert ELEMENTS_TILE >= 4
	$SIMD_TILE = ELEMENTS_TILE // 4
	$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
	$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
	#include <assert.h>

	#include <arm_neon.h>

	#include <xnnpack/common.h>
	#include <xnnpack/raddstoreexpminusmax.h>


	$PARAMS_STRUCT = "neonfma_rr1_p5" if FMA else "neon_rr2_p5"
	void xnn_f32_raddstoreexpminusmax_ukernel__${"neonfma" if FMA else "neon"}_rr${1 if FMA else 2}_p5_x${ELEMENTS_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
	size_t elements,
	const float* input,
	const float* max,
	float* output,
	float* sum,
	const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
	{
	assert(elements % sizeof(float) == 0);

	const float32x4_t vi_max = vld1q_dup_f32(max);
	const float32x4_t vlog2e = vld1q_dup_f32(&params->${PARAMS_STRUCT}.log2e);
	const float32x4_t vmagic_bias = vld1q_dup_f32(&params->${PARAMS_STRUCT}.magic_bias);
	$if FMA:
	const float32x4_t vminus_ln2 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.minus_ln2);
	$else:
	const float32x4_t vminus_ln2_hi = vld1q_dup_f32(&params->${PARAMS_STRUCT}.minus_ln2_hi);
	const float32x4_t vminus_ln2_lo = vld1q_dup_f32(&params->${PARAMS_STRUCT}.minus_ln2_lo);
	const float32x4_t vc5 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c5);
	const float32x4_t vc4 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c4);
	const float32x4_t vc3 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c3);
	const float32x4_t vc2 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c2);
	const float32x4_t vc1 = vld1q_dup_f32(&params->${PARAMS_STRUCT}.c1);
	const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->${PARAMS_STRUCT}.denorm_cutoff);

	$if ELEMENTS_TILE > 4:
	$for K in range(ACCUMULATORS):
	float32x4_t vacc${K} = vmovq_n_f32(0.0f);
	for (; elements >= ${ELEMENTS_TILE} * sizeof(float); elements -= ${ELEMENTS_TILE} * sizeof(float)) {
	$for N in range(0, ELEMENTS_TILE, 4):
	const float32x4_t vi${ABC[N:N+4]} = vld1q_f32(input); input += 4;

	$for N in range(0, ELEMENTS_TILE, 4):
	const float32x4_t vx${ABC[N:N+4]} = vsubq_f32(vi${ABC[N:N+4]}, vi_max);

	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vn${ABC[N:N+4]} = ${VMULADDQ_F32}(vmagic_bias, vx${ABC[N:N+4]}, vlog2e);

	$for N in range(0, ELEMENTS_TILE, 4):
	const float32x4_t vs${ABC[N:N+4]} = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn${ABC[N:N+4]}), 23));

	$for N in range(0, ELEMENTS_TILE, 4):
	vn${ABC[N:N+4]} = vsubq_f32(vn${ABC[N:N+4]}, vmagic_bias);

	$if FMA:
	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vx${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2);
	$else:
	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vx${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_hi);

	$for N in range(0, ELEMENTS_TILE, 4):
	vt${ABC[N:N+4]} = ${VMULADDQ_F32}(vt${ABC[N:N+4]}, vn${ABC[N:N+4]}, vminus_ln2_lo);

	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc4, vc5, vt${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc3, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc2, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	vp${ABC[N:N+4]} = ${VMULADDQ_F32}(vc1, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	vt${ABC[N:N+4]} = vmulq_f32(vt${ABC[N:N+4]}, vs${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	float32x4_t vf${ABC[N:N+4]} = ${VMULADDQ_F32}(vs${ABC[N:N+4]}, vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

	$for N in range(0, ELEMENTS_TILE, 4):
	vf${ABC[N:N+4]} = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf${ABC[N:N+4]}), vcltq_f32(vx${ABC[N:N+4]}, vdenorm_cutoff)));

	$for N in range(0, ELEMENTS_TILE, 4):
	vst1q_f32(output, vf${ABC[N:N+4]}); output += 4;

	$for N in range(0, ELEMENTS_TILE, 4):
	vacc${N % ACCUMULATORS} = vaddq_f32(vacc${N % ACCUMULATORS}, vf${ABC[N:N+4]});
	}
	$if ACCUMULATORS > 1:
	$ACC_SLICE = 1
	$while ACC_SLICE < ACCUMULATORS:
	$for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
	$if A + ACC_SLICE < ACCUMULATORS:
	vacc${A} = vaddq_f32(vacc${A}, vacc${A + ACC_SLICE});
	$ACC_SLICE *= 2

	float32x4_t vacc = vacc0;
	$else:
	float32x4_t vacc = vmovq_n_f32(0.0f);
	for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
	const float32x4_t vi = vld1q_f32(input); input += 4;

	const float32x4_t vx = vsubq_f32(vi, vi_max);

	float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

	const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

	vn = vsubq_f32(vn, vmagic_bias);

	$if FMA:
	float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2);
	$else:
	float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
	vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

	float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
	vp = ${VMULADDQ_F32}(vc3, vp, vt);
	vp = ${VMULADDQ_F32}(vc2, vp, vt);
	vp = ${VMULADDQ_F32}(vc1, vp, vt);

	vt = vmulq_f32(vt, vs);
	float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

	vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

	vst1q_f32(output, vf); output += 4;

	vacc = vaddq_f32(vacc, vf);
	}
	#if XNN_ARCH_ARM64
	float vacc_lo = vaddvq_f32(vacc);
	#else
	float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc));
	#endif
	if (elements != 0) {
	assert(elements >= 1 * sizeof(float));
	assert(elements <= 3 * sizeof(float));
	const float32x4_t vi = vld1q_f32(input); input += 4;

	const float32x4_t vx = vsubq_f32(vi, vi_max);

	float32x4_t vn = ${VMULADDQ_F32}(vmagic_bias, vx, vlog2e);

	const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));

	vn = vsubq_f32(vn, vmagic_bias);

	$if FMA:
	float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2);
	$else:
	float32x4_t vt = ${VMULADDQ_F32}(vx, vn, vminus_ln2_hi);
	vt = ${VMULADDQ_F32}(vt, vn, vminus_ln2_lo);

	float32x4_t vp = ${VMULADDQ_F32}(vc4, vc5, vt);
	vp = ${VMULADDQ_F32}(vc3, vp, vt);
	vp = ${VMULADDQ_F32}(vc2, vp, vt);
	vp = ${VMULADDQ_F32}(vc1, vp, vt);

	vt = vmulq_f32(vt, vs);
	float32x4_t vf = ${VMULADDQ_F32}(vs, vp, vt);

	vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcltq_f32(vx, vdenorm_cutoff)));

	float32x2_t vf_lo = vget_low_f32(vf);
	if (elements & (2 * sizeof(float))) {
	vst1_f32(output, vf_lo); output += 2;

	#if XNN_ARCH_ARM64
	vacc_lo += vaddv_f32(vf_lo);
	#else
	vacc_lo = vadd_f32(vacc_lo, vf_lo);
	#endif

	vf_lo = vget_high_f32(vf);
	}
	if (elements & (1 * sizeof(float))) {
	vst1_lane_f32(output, vf_lo, 0);

	#if XNN_ARCH_ARM64
	vacc_lo += vget_lane_f32(vf_lo, 0);
	#else
	vacc_lo = vadd_f32(vacc_lo, vreinterpret_f32_u64(vshl_n_u64(vreinterpret_u64_f32(vf_lo), 32)));
	#endif
	}
	}
	#if XNN_ARCH_ARM64
	*sum = vacc_lo;
	#else
	vst1_lane_f32(sum, vpadd_f32(vacc_lo, vacc_lo), 0);
	#endif
	}