| // 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 NR == 2 |
| $assert MR % 2 == 0 |
| #include <assert.h> |
| |
| #include <psimd.h> |
| |
| #include <xnnpack/igemm.h> |
| |
| |
| void xnn_f32_igemm_ukernel_${MR}x${NR}c4__psimd( |
| 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_output_params params[restrict static 1]) |
| { |
| assert(mr != 0); |
| assert(mr <= ${MR}); |
| assert(nc != 0); |
| assert(kc != 0); |
| assert(kc % sizeof(float) == 0); |
| assert(ks != 0); |
| assert(ks % (${MR} * sizeof(void*)) == 0); |
| assert(a_offset % sizeof(float) == 0); |
| assert(a != NULL); |
| assert(w != NULL); |
| assert(c != NULL); |
| |
| float* c0 = c; |
| $for M in range(1, MR): |
| float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride); |
| $if M % 2 == 0: |
| if XNN_UNPREDICTABLE(mr <= ${M}) { |
| c${M} = c${M-1}; |
| } |
| $elif M + 1 == MR: |
| if XNN_UNPREDICTABLE(mr != ${M+1}) { |
| c${M} = c${M-1}; |
| } |
| $else: |
| if XNN_UNPREDICTABLE(mr < ${M+1}) { |
| c${M} = c${M-1}; |
| } |
| |
| do { |
| psimd_f32 vacc0x0c4 = psimd_load1_f32(w); |
| $for N in range(1, NR): |
| psimd_f32 vacc0x${N}c4 = psimd_load1_f32(w + ${N}); |
| $for M in range(1, MR): |
| $for N in range(NR): |
| psimd_f32 vacc${M}x${N}c4 = vacc0x${N}c4; |
| w += ${NR}; |
| |
| size_t p = ks; |
| do { |
| $for M in range(MR): |
| const float* restrict a${M} = a[${M}]; |
| assert(a${M} != NULL); |
| if XNN_UNPREDICTABLE(a${M} != zero) { |
| a${M} = (const float*) ((uintptr_t) a${M} + a_offset); |
| } |
| a += ${MR}; |
| |
| size_t k = kc; |
| for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) { |
| $for M in range(MR): |
| const psimd_f32 va${M} = psimd_load_f32(a${M}); |
| a${M} += 4; |
| |
| const psimd_f32 vb0 = psimd_load_f32(w); |
| $for N in range(1, NR): |
| const psimd_f32 vb${N} = psimd_load_f32(w + ${N * 4}); |
| w += ${NR * 4}; |
| |
| $for M in range(MR): |
| $for N in range(NR): |
| vacc${M}x${N}c4 = psimd_qfma_f32(vacc${M}x${N}c4, va${M}, vb${N}); |
| } |
| if XNN_UNLIKELY(k != 0) { |
| $for M in range(MR): |
| const psimd_f32 va${M} = psimd_load_f32(a${M}); |
| |
| const psimd_f32 vb0 = psimd_load_f32(w); |
| $for N in range(1, NR): |
| const psimd_f32 vb${N} = psimd_load_f32(w + ${N * 4}); |
| w += ${NR * 4}; |
| |
| const psimd_f32 vzero = psimd_splat_f32(0.0f); |
| $for N in range(NR): |
| const psimd_s32 vmask${N} = vb${N} != vzero; |
| |
| $for M in range(MR): |
| $for N in range(NR): |
| vacc${M}x${N}c4 = psimd_qfma_f32(vacc${M}x${N}c4, psimd_andmask_f32(vmask${N}, va${M}), vb${N}); |
| } |
| p -= ${MR} * sizeof(void*); |
| } while (p != 0); |
| |
| $for M in range(MR): |
| const psimd_f32 vacc${M}x01c2 = psimd_add_f32(psimd_interleave_lo_f32(vacc${M}x0c4, vacc${M}x1c4), psimd_interleave_hi_f32(vacc${M}x0c4, vacc${M}x1c4)); |
| |
| $for M in range(0, MR, 2): |
| psimd_f32 vacc${M}${M+1}x01 = psimd_add_f32(psimd_concat_lo_f32(vacc${M}x01c2, vacc${M+1}x01c2), psimd_concat_hi_f32(vacc${M}x01c2, vacc${M+1}x01c2)); |
| |
| const psimd_f32 vmax = psimd_load_splat_f32(¶ms->scalar.max); |
| $for M in range(0, MR, 2): |
| vacc${M}${M+1}x01 = psimd_min_f32(vacc${M}${M+1}x01, vmax); |
| |
| const psimd_f32 vmin = psimd_load_splat_f32(¶ms->scalar.min); |
| $for M in range(0, MR, 2): |
| vacc${M}${M+1}x01 = psimd_max_f32(vacc${M}${M+1}x01, vmin); |
| |
| if XNN_LIKELY(nc >= ${NR}) { |
| $for M in reversed(range(0, MR, 2)): |
| psimd_store2_f32(c${M+1}, psimd_concat_hi_f32(vacc${M}${M+1}x01, vacc${M}${M+1}x01)); |
| c${M+1} = (float*) ((uintptr_t) c${M+1} + cn_stride); |
| psimd_store2_f32(c${M}, vacc${M}${M+1}x01); |
| c${M} = (float*) ((uintptr_t) c${M} + cn_stride); |
| |
| a = (const float**restrict) ((uintptr_t) a - ks); |
| nc -= ${NR}; |
| } else { |
| assert(nc == 1); |
| $for M in reversed(range(0, MR, 2)): |
| psimd_store1_f32(c${M+1}, psimd_concat_hi_f32(vacc${M}${M+1}x01, vacc${M}${M+1}x01)); |
| psimd_store1_f32(c${M}, vacc${M}${M+1}x01); |
| |
| nc = 0; |
| } |
| } while (nc != 0); |
| } |