Home
last modified time | relevance | path

Searched refs:m0 (Results 1 – 25 of 1085) sorted by relevance

12345678910>>...44

/external/libaom/libaom/aom_dsp/x86/
Dintrapred_asm_sse2.asm35 movd m0, [aboveq]
37 punpckldq m0, m2
38 psadbw m0, m1
39 paddw m0, [GLOBAL(pw_4)]
40 psraw m0, 3
41 pshuflw m0, m0, 0x0
42 packuswb m0, m0
43 movd [dstq ], m0
44 movd [dstq+strideq], m0
46 movd [dstq ], m0
[all …]
Dhighbd_intrapred_asm_sse2.asm27 movq m0, [aboveq]
29 paddw m0, m2
30 pshuflw m1, m0, 0xe
31 paddw m0, m1
32 pshuflw m1, m0, 0x1
33 paddw m0, m1
34 paddw m0, [GLOBAL(pw_4)]
35 psraw m0, 3
36 pshuflw m0, m0, 0x0
37 movq [dstq ], m0
[all …]
Dsubpel_variance_sse2.asm390 movu m0, [srcq]
394 punpckhbw m2, m0, m4
395 punpcklbw m0, m4
397 pmaddubsw m0, filter_y_a
399 paddw m0, filter_rnd
401 punpckhbw m2, m0, m5
403 punpcklbw m0, m5
413 pmullw m0, filter_y_a
415 paddw m0, filter_rnd
417 paddw m0, m4
[all …]
Daom_subpixel_8t_ssse3.asm43 pshuflw m0, m4, 0b ;k0_k1
47 punpcklqdq m0, m0
51 mova k0k1, m0
125 palignr m0, m1, m4, 1
126 pmaddubsw m0, k0k1k4k5
133 punpckhqdq m4, m0, m2
134 punpcklqdq m0, m2
137 paddsw m0, m4
143 paddsw m0, m1
144 paddsw m0, krd
[all …]
Dhighbd_subpel_variance_impl_sse2.asm184 movu m0, [srcq]
189 pavgw m0, [secq]
192 SUM_SSE m0, m1, m2, m3, m6, m7
200 movu m0, [srcq]
205 pavgw m0, [secq]
209 SUM_SSE m0, m1, m2, m3, m6, m7
228 movu m0, [srcq]
234 pavgw m0, m4
237 pavgw m0, [secq]
240 SUM_SSE m0, m2, m1, m3, m6, m7
[all …]
Dhighbd_sad4d_sse2.asm20 movh m0, [srcq +%2*2]
26 movhps m0, [srcq +%4*2]
31 mova m3, m0
32 mova m2, m0
35 psubusw m4, m0
36 psubusw m5, m0
41 mova m3, m0
42 mova m2, m0
45 psubusw m6, m0
46 psubusw m7, m0
[all …]
Dsad_sse2.asm55 pxor m0, m0
75 paddd m0, m1
76 paddd m0, m3
99 paddd m0, m1
100 paddd m0, m3
105 movhlps m1, m0
106 paddd m0, m1
107 movd eax, m0
123 pxor m0, m0
143 paddd m0, m1
[all …]
/external/libvpx/libvpx/vpx_dsp/x86/
Dintrapred_sse2.asm48 movq m0, [aboveq]
50 psrldq m1, m0, 1
51 psrldq m2, m0, 2
52 X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3
63 psrlq m0, 56
64 movd tempd, m0
75 pslldq m0, m1, 1
79 X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3
80 punpckhbw m0, m0 ; 7 7
81 punpcklwd m0, m0 ; 7 7 7 7
[all …]
Dhighbd_intrapred_sse2.asm24 movq m0, [aboveq]
26 paddw m0, m2
27 pshuflw m1, m0, 0xe
28 paddw m0, m1
29 pshuflw m1, m0, 0x1
30 paddw m0, m1
31 paddw m0, [GLOBAL(pw_4)]
32 psraw m0, 3
33 pshuflw m0, m0, 0x0
34 movq [dstq ], m0
[all …]
Dsubpel_variance_sse2.asm387 movu m0, [srcq]
391 punpckhbw m2, m0, m4
392 punpcklbw m0, m4
394 pmaddubsw m0, filter_y_a
396 paddw m0, filter_rnd
398 punpckhbw m2, m0, m5
400 punpcklbw m0, m5
410 pmullw m0, filter_y_a
412 paddw m0, filter_rnd
414 paddw m0, m4
[all …]
Dintrapred_ssse3.asm37 mova m0, [aboveq]
42 pshufb m2, m0, [GLOBAL(sh_b23456789abcdefff)]
43 pavgb m3, m2, m0
44 pxor m2, m0
45 pshufb m0, m1
48 pavgb m0, m3
53 mova [dstq ], m0
54 movhps [dst8q ], m0
55 pshufb m0, m1
56 mova [dstq +strideq ], m0
[all …]
Dhighbd_subpel_variance_impl_sse2.asm181 movu m0, [srcq]
186 pavgw m0, [second_predq]
189 SUM_SSE m0, m1, m2, m3, m6, m7
197 movu m0, [srcq]
202 pavgw m0, [second_predq]
206 SUM_SSE m0, m1, m2, m3, m6, m7
225 movu m0, [srcq]
231 pavgw m0, m4
234 pavgw m0, [second_predq]
237 SUM_SSE m0, m2, m1, m3, m6, m7
[all …]
Dhighbd_sad4d_sse2.asm17 movh m0, [srcq +%2*2]
23 movhps m0, [srcq +%4*2]
28 mova m3, m0
29 mova m2, m0
32 psubusw m4, m0
33 psubusw m5, m0
38 mova m3, m0
39 mova m2, m0
42 psubusw m6, m0
43 psubusw m7, m0
[all …]
Dvpx_subpixel_8t_ssse3.asm39 pshuflw m0, m4, 0b ;k0_k1
43 punpcklqdq m0, m0
47 mova k0k1, m0
121 palignr m0, m1, m4, 1
122 pmaddubsw m0, k0k1k4k5
129 punpckhqdq m4, m0, m2
130 punpcklqdq m0, m2
133 paddsw m0, m4
139 paddsw m0, m1
140 paddsw m0, krd
[all …]
Dsad4d_sse2.asm17 movd m0, [srcq +%2]
25 punpckldq m0, m1
33 movlhps m0, m0
36 psadbw m6, m0
37 psadbw m7, m0
52 punpckldq m0, m5
53 movlhps m0, m0
56 psadbw m1, m0
57 psadbw m3, m0
72 movh m0, [srcq +%2]
[all …]
/external/tensorflow/tensorflow/lite/kernels/
Dconcatenation_test.cc103 ConcatenationOpModel m0({TensorType_FLOAT32, {2, 1, 2}}, /*axis=*/1, in TEST() local
105 m0.SetInput(0, {1.0f, 3.0f, 4.0f, 7.0f}); in TEST()
106 m0.Invoke(); in TEST()
107 EXPECT_THAT(m0.GetOutput(), ElementsAreArray({1, 3, 4, 7})); in TEST()
111 ConcatenationOpModel m0({TensorType_FLOAT32, {2, 1, 2, 1, 3}}, /*axis=*/2, in TEST() local
113 m0.SetInput(0, {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f, in TEST()
115 m0.Invoke(); in TEST()
116 EXPECT_THAT(m0.GetOutput(), in TEST()
121 ConcatenationOpModel m0({TensorType_FLOAT32, {2, 1, 2, 1, 3}}, /*axis=*/0, in TEST() local
123 m0.SetInput(0, {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f, in TEST()
[all …]
/external/llvm-project/llvm/test/CodeGen/AMDGPU/
Dmerge-m0.mir1 # RUN: llc -march=amdgcn -amdgpu-enable-merge-m0 -verify-machineinstrs -run-pass si-fix-sgpr-copies…
3 # GCN-LABEL: name: merge-m0-many-init
52 name: merge-m0-many-init
63 SI_INIT_M0 -1, implicit-def $m0
64 DS_WRITE_B32 %0, %1, 0, 0, implicit $m0, implicit $exec
65 SI_INIT_M0 -1, implicit-def $m0
66 DS_WRITE_B32 %0, %1, 0, 0, implicit $m0, implicit $exec
67 SI_INIT_M0 65536, implicit-def $m0
68 DS_WRITE_B32 %0, %1, 0, 0, implicit $m0, implicit $exec
69 SI_INIT_M0 65536, implicit-def $m0
[all …]
Dset-gpr-idx-peephole.mir10 …_SET_GPR_IDX_ON $sgpr2, 1, implicit-def $mode, implicit-def $m0, implicit $mode, implicit undef $m0
11 …2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $m0
12 …2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $m0
14 …_SET_GPR_IDX_ON $sgpr2, 1, implicit-def $mode, implicit-def $m0, implicit $mode, implicit undef $m0
15 …2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $m0
17 …_IDX_ON killed $sgpr2, 1, implicit-def $mode, implicit-def $m0, implicit $mode, implicit undef $m0
18 …2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $m0
28 …; GCN: S_SET_GPR_IDX_ON $sgpr2, 1, implicit-def $m0, implicit-def $mode, implicit undef $m0, impli…
29 …2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $m0
31 …2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, implicit $m0
[all …]
Dfold-operands-remove-m0-redef.mir34 ; GCN: $m0 = COPY [[COPY1]]
35 …; GCN: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec…
36 …; GCN: [[DS_READ_B32_1:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 64, 0, implicit $m0, implicit $ex…
39 $m0 = COPY %1
40 %2:vgpr_32 = DS_READ_B32 %0, 0, 0, implicit $m0, implicit $exec :: (load 4)
41 $m0 = COPY %1
42 %3:vgpr_32 = DS_READ_B32 %0, 64, 0, implicit $m0, implicit $exec :: (load 4)
58 ; GCN: $m0 = COPY [[COPY1]]
59 …; GCN: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec…
60 …; GCN: [[DS_READ_B32_1:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 64, 0, implicit $m0, implicit $ex…
[all …]
Dspill-m0.ll11 ; GCN-NEXT: s_mov_b32 m0, 0
13 ; GCN-DAG: s_mov_b32 [[M0_COPY:s[0-9]+]], m0
27 ; TOVGPR: s_mov_b32 m0, [[M0_RESTORE]]
32 ; TOVMEM: s_mov_b32 m0, [[M0_RESTORE]]
34 ; GCN: s_add_i32 s{{[0-9]+}}, m0, 1
37 %m0 = call i32 asm sideeffect "s_mov_b32 m0, 0", "={m0}"() #0
46 %foo = call i32 asm sideeffect "s_add_i32 $0, $1, 1", "=s,{m0}"(i32 %m0) #0
53 ; m0 is killed, so it isn't necessary during the entry block spill to preserve it
56 ; GCN-NOT: v_readlane_b32 m0
57 ; GCN-NOT: s_buffer_store_dword m0
[all …]
Dgws-hazards.mir16 ; GFX9: $m0 = S_MOV_B32 -1
18 ; GFX9: DS_GWS_INIT $vgpr0, 0, implicit $m0, implicit $exec
21 ; VI: $m0 = S_MOV_B32 -1
23 ; VI: DS_GWS_INIT $vgpr0, 0, implicit $m0, implicit $exec
26 ; CI: $m0 = S_MOV_B32 -1
27 ; CI: DS_GWS_INIT $vgpr0, 0, implicit $m0, implicit $exec
30 ; SI: $m0 = S_MOV_B32 -1
31 ; SI: DS_GWS_INIT $vgpr0, 0, implicit $m0, implicit $exec
32 $m0 = S_MOV_B32 -1
33 DS_GWS_INIT $vgpr0, 0, implicit $m0, implicit $exec
[all …]
Dlocal-atomics64.ll6 ; SICIVI: s_mov_b32 m0
7 ; GFX9-NOT: m0
18 ; SICIVI: s_mov_b32 m0
19 ; GFX9-NOT: m0
31 ; SICIVI: s_mov_b32 m0
32 ; GFX9-NOT: m0
44 ; SICIVI: s_mov_b32 m0
45 ; GFX9-NOT: m0
56 ; SICIVI-DAG: s_mov_b32 m0
57 ; GFX9-NOT: m0
[all …]
/external/tensorflow/tensorflow/lite/delegates/hexagon/builders/tests/
Dconcat_test.cc87 QuantizedConcatenationOpModel m0({{tensor_dtype, {2, 1, 1, 2}, -12.7, 12.8}, in FourInputsQuantizedSameRangeImpl() local
93 m0.SetInput<integer_type>(0, {1.0f, 3.0f, 4.0f, 7.0f}); in FourInputsQuantizedSameRangeImpl()
94 m0.SetInput<integer_type>(1, {1.1f, 3.1f, 4.1f, 7.1f}); in FourInputsQuantizedSameRangeImpl()
95 m0.SetInput<integer_type>(2, {1.2f, 3.2f, 4.2f, 7.2f}); in FourInputsQuantizedSameRangeImpl()
96 m0.SetInput<integer_type>(3, {1.3f, 3.3f, 4.3f, 7.3f}); in FourInputsQuantizedSameRangeImpl()
97 m0.ApplyDelegateAndInvoke(); in FourInputsQuantizedSameRangeImpl()
98 EXPECT_THAT(m0.GetDequantizedOutput<integer_type>(), in FourInputsQuantizedSameRangeImpl()
119 QuantizedConcatenationOpModel m0({{tensor_dtype, in TwoInputsNegativeAxisImpl() local
133 m0.SetInput<integer_type>(0, tensor0); in TwoInputsNegativeAxisImpl()
134 m0.SetInput<integer_type>(1, tensor1); in TwoInputsNegativeAxisImpl()
[all …]
/external/python/cpython3/Modules/_blake2/impl/
Dblake2s-load-sse41.h18 buf = TOI(_mm_shuffle_ps(TOF(m0), TOF(m1), _MM_SHUFFLE(2,0,2,0)));
21 buf = TOI(_mm_shuffle_ps(TOF(m0), TOF(m1), _MM_SHUFFLE(3,1,3,1)));
44 t2 = _mm_blend_epi16(m0, t1, 0xF0); \
48 t0 = _mm_unpackhi_epi32(m0,m1); \
60 t0 = _mm_unpacklo_epi32(m2,m0); \
61 t1 = _mm_blend_epi16(t0, m0, 0xF0); \
66 t0 = _mm_blend_epi16(m0, m2, 0x3C); \
73 t1 = _mm_blend_epi16(m0, m1, 0x33); \
78 t0 = _mm_unpackhi_epi32(m0,m1); \
85 t1 = _mm_blend_epi16(m3,m0,0x0C); \
[all …]
/external/libaom/libaom/third_party/libyuv/source/
Drow_x86.asm33 mov%2 m0, [src_yuy2q]
37 pand m0, m0, m2 ; YUY2 even bytes are Y
40 psrlw m0, m0, 8 ; UYVY odd bytes are Y
43 packuswb m0, m0, m1
45 vpermq m0, m0, 0xd8
48 mov%2 [dst_yq], m0
79 mov%1 m0, [src_uvq]
82 psrlw m2, m0, 8 ; odd bytes
84 pand m0, m0, m4 ; even bytes
86 packuswb m0, m0, m1
[all …]

12345678910>>...44