| // 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. |
| |
| $assert MR % 4 == 0 |
| $ABC = "0123456789ABCDEFGHIJK" |
| #include <assert.h> |
| |
| #include <arm_neon.h> |
| |
| #include <xnnpack/spmm.h> |
| |
| |
| void xnn_f32_spmm_ukernel_${MR}x${NR}__${"neonfma" if FMA else "neon"}_pipelined( |
| uint32_t m, |
| uint32_t n, |
| const float*restrict a, |
| const float*restrict weights, |
| const int32_t*restrict widx_dmap, |
| const uint32_t*restrict nidx_nnzmap, |
| float*restrict c, |
| const union xnn_f32_output_params params[restrict static 1]) |
| { |
| assert(m != 0); |
| |
| const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); |
| const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); |
| size_t i = m; |
| while XNN_LIKELY(i >= ${MR}) { |
| const float*restrict w = weights; |
| const int32_t* dmap = widx_dmap; |
| const uint32_t* nnzmap = nidx_nnzmap; |
| float32x4_t vw = vld1q_dup_f32(w); w += 1; |
| intptr_t diff = *dmap++; |
| float32x4_t va0123 = vld1q_f32(a); |
| $for M in range(4, MR, 4): |
| float32x4_t va${ABC[M:M+4]} = vld1q_f32(a + ${M}); |
| $if MR == 16: |
| __builtin_prefetch(a + 16); |
| size_t j = n; |
| do { |
| uint32_t nnz = *nnzmap++; |
| $for M in range(0, MR, 4): |
| float32x4_t vacc${ABC[M:M+4]} = vw; |
| vw = vld1q_dup_f32(w); w += 1; |
| if XNN_LIKELY(nnz != 0) { |
| do { |
| $for M in range(0, MR, 4): |
| vacc${ABC[M:M+4]} = vfmaq_f32(vacc${ABC[M:M+4]}, va${ABC[M:M+4]}, vw); |
| a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff); |
| |
| diff = *dmap++; |
| vw = vld1q_dup_f32(w); w += 1; |
| va0123 = vld1q_f32(a); |
| $for M in range(4, MR, 4): |
| va${ABC[M:M+4]} = vld1q_f32(a + ${M}); |
| $if MR == 16: |
| __builtin_prefetch(a + 16); |
| } while (--nnz != 0); |
| } |
| $for M in range(0, MR, 4): |
| float32x4_t vout${ABC[M:M+4]} = vminq_f32(vacc${ABC[M:M+4]}, vmax); |
| $for M in range(0, MR, 4): |
| vout${ABC[M:M+4]} = vmaxq_f32(vout${ABC[M:M+4]}, vmin); |
| vst1q_f32(c, vout0123); |
| $for M in range(4, MR, 4): |
| vst1q_f32(c + ${M}, vout${ABC[M:M+4]}); |
| c += m; |
| } while (--j != 0); |
| c -= m * n; |
| c += ${MR}; |
| a += ${MR}; |
| i -= ${MR}; |
| } |
| if XNN_UNLIKELY(i != 0) { |
| $for LOG2M in reversed(range((MR - 1).bit_length())): |
| $SUBMR = 1 << LOG2M |
| if (i & ${SUBMR}) { |
| const float*restrict w = weights; |
| const int32_t* dmap = widx_dmap; |
| const uint32_t* nnzmap = nidx_nnzmap; |
| size_t j = n; |
| do { |
| uint32_t nnz = *nnzmap++; |
| $if SUBMR <= 2: |
| float32x2_t vacc${ABC[0:SUBMR]} = vld1_dup_f32(w); w += 1; |
| $else: |
| float32x4_t vacc0123 = vld1q_dup_f32(w); w += 1; |
| $for M in range(4, SUBMR, 4): |
| float32x4_t vacc${ABC[M:M+4]} = vacc0123; |
| if XNN_LIKELY(nnz != 0) { |
| do { |
| const intptr_t diff = *dmap++; |
| $if SUBMR == 1: |
| const float32x2_t va0 = vld1_dup_f32(a); |
| $elif SUBMR == 2: |
| const float32x2_t va01 = vld1_f32(a); |
| $else: |
| const float32x4_t va0123 = vld1q_f32(a); |
| $for M in range(4, SUBMR, 4): |
| const float32x4_t va${ABC[M:M+4]} = vld1q_f32(a + ${M}); |
| a = (const float*restrict) ((uintptr_t) a + (uintptr_t) diff); |
| $if SUBMR <= 2: |
| const float32x2_t vb = vld1_dup_f32(w); w += 1; |
| $else: |
| const float32x4_t vb = vld1q_dup_f32(w); w += 1; |
| $if SUBMR <= 2: |
| vacc${ABC[0:SUBMR]} = vfma_f32(vacc${ABC[0:SUBMR]}, va${ABC[0:SUBMR]}, vb); |
| $else: |
| $for M in range(0, SUBMR, 4): |
| vacc${ABC[M:M+4]} = vfmaq_f32(vacc${ABC[M:M+4]}, va${ABC[M:M+4]}, vb); |
| } while (--nnz != 0); |
| } |
| $if SUBMR <= 2: |
| float32x2_t vout${ABC[0:SUBMR]} = vmin_f32(vacc${ABC[0:SUBMR]}, vget_low_f32(vmax)); |
| vout${ABC[0:SUBMR]} = vmax_f32(vout${ABC[0:SUBMR]}, vget_low_f32(vmin)); |
| $if SUBMR == 1: |
| vst1_lane_f32(c, vout${ABC[0]}, 0); |
| $else: |
| vst1_f32(c, vout${ABC[0:SUBMR]}); |
| $else: |
| $for M in range(0, SUBMR, 4): |
| float32x4_t vout${ABC[M:M+4]} = vminq_f32(vacc${ABC[M:M+4]}, vmax); |
| $for M in range(0, SUBMR, 4): |
| vout${ABC[M:M+4]} = vmaxq_f32(vout${ABC[M:M+4]}, vmin); |
| vst1q_f32(c, vout0123); |
| $for M in range(4, SUBMR, 4): |
| vst1q_f32(c + ${M}, vout${ABC[M:M+4]}); |
| c += m; |
| } while (--j != 0); |
| c -= m * n; |
| c += ${SUBMR}; |
| a += ${SUBMR}; |
| } |
| } |
| } |