/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p0p1c3x4-neon-2x1.c | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local 423 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() 424 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() 425 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() 426 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() 427 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
|
D | 3x3s2p0p1c3x4-neonfma-2x1.c | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local 421 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() 422 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() 423 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() 424 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() 425 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
|
D | 3x3s2p1c3x4-neonfma-2x1.c | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local 421 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() 422 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() 423 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() 424 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() 425 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
|
D | 3x3s2p1c3x4-neon-2x1.c | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local 423 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() 424 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() 425 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() 426 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() 427 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 584 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 585 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 586 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 587 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 588 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neon-2x1.c | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 586 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 587 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 588 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 589 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 590 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 586 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 587 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 588 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 589 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 590 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 584 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 585 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 586 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 587 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 588 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x4-neonfma-2x2.c | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local 723 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 724 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 725 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 726 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 727 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
|
D | 3x3s2p1c3x4-neon-2x2.c | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local 727 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 728 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 729 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 730 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 731 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
|
D | 3x3s2p0p1c3x4-neon-2x2.c | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local 727 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 728 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 729 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 730 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 731 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
|
D | 3x3s2p0p1c3x4-neonfma-2x2.c | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local 723 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 724 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 725 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 726 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 727 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
|
D | 3x3s2p1c3x8-neonfma-2x2.c | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 1048 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 1049 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 1050 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 1051 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 1052 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 1052 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 1053 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 1054 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 1055 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 1056 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 1048 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 1049 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 1050 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 1051 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 1052 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neon-2x2.c | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 1052 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 1053 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 1054 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 1055 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 1056 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
/external/XNNPACK/src/f32-conv-hwc2chw/ |
D | 3x3s2p1c3x4-sse-1x1.c | 35 …const size_t input_width_decrement = round_down_po2(input_width, 2) * 3 /* channels */ * sizeof(fl… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local 295 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() 296 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() 297 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
|
D | 3x3s2p1c3x4-scalar-1x1.c | 33 …const size_t input_width_decrement = round_down_po2(input_width, 2) * 3 /* channels */ * sizeof(fl… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1() local 655 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1() 656 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1() 657 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1()
|
/external/XNNPACK/src/f32-conv-hwc/ |
D | 3x3s2p1c3x4-scalar-1x1.c | 33 …const size_t input_width_decrement = round_down_po2(input_width, 2) * 3 /* channels */ * sizeof(fl… in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__scalar_1x1() local 669 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__scalar_1x1() 670 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__scalar_1x1() 671 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__scalar_1x1()
|
D | 3x3s2p0p1c3x4-scalar-1x1.c | 33 …const size_t input_width_decrement = round_down_po2(input_width - 1, 2) * 3 /* channels */ * sizeo… in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__scalar_1x1() local 669 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__scalar_1x1() 670 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__scalar_1x1() 671 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__scalar_1x1()
|
D | 3x3s2p0p1c3-neon-x1.c.in | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… 530 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); 531 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); 532 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); 533 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); 534 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement);
|
D | 3x3s2p1c3-neon-x1.c.in | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); 530 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); 531 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); 532 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); 533 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); 534 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement);
|
D | 3x3s2p0p1c3-neon-x2.c.in | 40 …const size_t input_width_decrement = (4 + ((input_width - 1) & 1) * 2 + (round_down_po2(input_widt… 925 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); 926 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); 927 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); 928 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); 929 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement);
|
D | 3x3s2p1c3-neon-x2.c.in | 40 const size_t input_width_decrement = input_width * 3 /* channels */ * sizeof(float); 925 i0 = (const float*) ((uintptr_t) i0 - input_width_decrement); 926 i1 = (const float*) ((uintptr_t) i1 - input_width_decrement); 927 i2 = (const float*) ((uintptr_t) i2 - input_width_decrement); 928 i3 = (const float*) ((uintptr_t) i3 - input_width_decrement); 929 i4 = (const float*) ((uintptr_t) i4 - input_width_decrement);
|