Home
last modified time | relevance | path

Searched refs:vacc3x0123456789ABCDEF (Results 1 – 15 of 15) sorted by relevance

/external/XNNPACK/src/f32-gemm/gen-inc/
D4x16inc-minmax-avx512f-broadcast.c66 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast() local
77 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
91 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
97 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
100 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
120 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
D5x16inc-minmax-avx512f-broadcast.c72 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast() local
84 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
100 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
107 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
113 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
135 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
D6x16inc-minmax-avx512f-broadcast.c78 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast() local
91 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
109 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
117 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
126 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
150 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
D7x16inc-minmax-avx512f-broadcast.c84 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() local
98 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
118 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
127 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
139 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
165 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
D8x16inc-minmax-avx512f-broadcast.c90 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() local
105 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
127 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
137 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
152 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
180 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
/external/XNNPACK/src/f32-gemm/gen/
D4x16-minmax-avx512f-broadcast.c64 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast() local
75 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast()
89 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast()
95 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast()
98 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast()
118 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast()
D5x16-minmax-avx512f-broadcast.c70 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast() local
82 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast()
98 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast()
105 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast()
111 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast()
133 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast()
D6x16-minmax-avx512f-broadcast.c76 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast() local
89 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast()
107 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast()
115 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast()
124 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast()
148 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast()
D7x16-minmax-avx512f-broadcast.c82 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() local
96 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast()
116 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast()
125 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast()
137 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast()
163 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast()
D8x16-minmax-avx512f-broadcast.c88 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() local
103 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast()
125 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast()
135 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast()
150 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast()
178 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast()
/external/XNNPACK/src/f32-igemm/gen/
D4x16-minmax-avx512f-broadcast.c62 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast() local
97 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast()
113 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast()
119 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast()
122 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast()
138 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_4x16__avx512f_broadcast()
D5x16-minmax-avx512f-broadcast.c66 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast() local
107 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast()
125 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast()
132 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast()
138 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast()
155 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast()
D6x16-minmax-avx512f-broadcast.c70 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast() local
117 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast()
137 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast()
145 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast()
154 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast()
172 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast()
D7x16-minmax-avx512f-broadcast.c74 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() local
127 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast()
149 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast()
158 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast()
170 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast()
189 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast()
D8x16-minmax-avx512f-broadcast.c78 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() local
137 …vacc3x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a3), vb0123456789ABCDEF, vacc3x0123456789… in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast()
161 vacc3x0123456789ABCDEF = _mm512_min_ps(vacc3x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast()
171 vacc3x0123456789ABCDEF = _mm512_max_ps(vacc3x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast()
186 _mm512_storeu_ps(c3, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast()
206 _mm512_mask_storeu_ps(c3, vmask, vacc3x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast()