/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 44 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 45 const v128_t vxi0x89ABCDEF = wasm_i16x8_load_8x8(i0 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 46 const v128_t vxi0xGHIJKLMN = wasm_i16x8_load_8x8(i0 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 48 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 49 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 50 const v128_t vxi1xGHIJKLMN = wasm_i16x8_load_8x8(i1 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 52 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 53 const v128_t vxi2x89ABCDEF = wasm_i16x8_load_8x8(i2 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 54 const v128_t vxi2xGHIJKLMN = wasm_i16x8_load_8x8(i2 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 56 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() [all …]
|
D | 7p7x-minmax-wasmsimd-c16-acc2.c | 44 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 45 const v128_t vxi0x89ABCDEF = wasm_i16x8_load_8x8(i0 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 47 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 48 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 50 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 51 const v128_t vxi2x89ABCDEF = wasm_i16x8_load_8x8(i2 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 53 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 54 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 56 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 57 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() [all …]
|
D | 7p7x-minmax-wasmsimd-c8-acc2.c | 44 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 46 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 48 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 50 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 52 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 54 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 56 const v128_t vxi6x01234567 = wasm_i16x8_load_8x8(i6); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 89 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 91 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 93 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() [all …]
|
D | 7x-minmax-wasmsimd-c24-acc2.c | 62 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 63 const v128_t vxi0x89ABCDEF = wasm_i16x8_load_8x8(i0 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 64 const v128_t vxi0xGHIJKLMN = wasm_i16x8_load_8x8(i0 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 66 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 67 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 68 const v128_t vxi1xGHIJKLMN = wasm_i16x8_load_8x8(i1 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 70 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 71 const v128_t vxi2x89ABCDEF = wasm_i16x8_load_8x8(i2 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 72 const v128_t vxi2xGHIJKLMN = wasm_i16x8_load_8x8(i2 + 16); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 74 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() [all …]
|
D | 7x-minmax-wasmsimd-c16-acc2.c | 62 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 63 const v128_t vxi0x89ABCDEF = wasm_i16x8_load_8x8(i0 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 65 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 66 const v128_t vxi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 68 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 69 const v128_t vxi2x89ABCDEF = wasm_i16x8_load_8x8(i2 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 71 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 72 const v128_t vxi3x89ABCDEF = wasm_i16x8_load_8x8(i3 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 74 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 75 const v128_t vxi4x89ABCDEF = wasm_i16x8_load_8x8(i4 + 8); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() [all …]
|
D | 7x-minmax-wasmsimd-c8-acc2.c | 62 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 64 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 66 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 68 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 70 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 72 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 74 const v128_t vxi6x01234567 = wasm_i16x8_load_8x8(i6); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 131 const v128_t vxi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 133 const v128_t vxi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 135 const v128_t vxi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() [all …]
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up24x9-minmax-wasmsimd-mul16.c | 91 const v128_t vi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 92 …const v128_t vk0x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 93 const v128_t vi0x89ABCDEF = wasm_i16x8_load_8x8(i0 + 8); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 94 …const v128_t vk0x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 95 const v128_t vi0xGHIJKLMN = wasm_i16x8_load_8x8(i0 + 16); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 96 …const v128_t vk0xGHIJKLMN = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 110 const v128_t vi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 111 …const v128_t vk1x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 112 const v128_t vi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() 113 …const v128_t vk1x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__wasmsimd_mul16() [all …]
|
D | up16x9-minmax-wasmsimd-mul16.c | 89 const v128_t vi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 90 …const v128_t vk0x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 91 const v128_t vi0x89ABCDEF = wasm_i16x8_load_8x8(i0 + 8); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 92 …const v128_t vk0x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 103 const v128_t vi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 104 …const v128_t vk1x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 105 const v128_t vi1x89ABCDEF = wasm_i16x8_load_8x8(i1 + 8); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 106 …const v128_t vk1x89ABCDEF = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 117 const v128_t vi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() 118 …const v128_t vk2x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__wasmsimd_mul16() [all …]
|
D | up8x9-minmax-wasmsimd-mul16.c | 87 const v128_t vi0x01234567 = wasm_i16x8_load_8x8(i0); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 88 …const v128_t vk0x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 96 const v128_t vi1x01234567 = wasm_i16x8_load_8x8(i1); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 97 …const v128_t vk1x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 105 const v128_t vi2x01234567 = wasm_i16x8_load_8x8(i2); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 106 …const v128_t vk2x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 114 const v128_t vi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 115 …const v128_t vk3x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 123 const v128_t vi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() 124 …const v128_t vk4x01234567 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__wasmsimd_mul16() [all …]
|
/external/XNNPACK/src/qs8-vadd/gen/ |
D | minmax-wasmsimd-x32.c | 35 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 36 const v128_t vy01234567 = wasm_i16x8_load_8x8(input_y); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 37 const v128_t vx89ABCDEF = wasm_i16x8_load_8x8(input_x + 8); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 38 const v128_t vy89ABCDEF = wasm_i16x8_load_8x8(input_y + 8); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 39 const v128_t vxGHIJKLMN = wasm_i16x8_load_8x8(input_x + 16); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 40 const v128_t vyGHIJKLMN = wasm_i16x8_load_8x8(input_y + 16); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 41 const v128_t vxOPQRSTUV = wasm_i16x8_load_8x8(input_x + 24); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 42 const v128_t vyOPQRSTUV = wasm_i16x8_load_8x8(input_y + 24); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 102 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32() 103 const v128_t vy01234567 = wasm_i16x8_load_8x8(input_y); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32()
|
D | minmax-wasmsimd-x24.c | 35 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24() 36 const v128_t vy01234567 = wasm_i16x8_load_8x8(input_y); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24() 37 const v128_t vx89ABCDEF = wasm_i16x8_load_8x8(input_x + 8); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24() 38 const v128_t vy89ABCDEF = wasm_i16x8_load_8x8(input_y + 8); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24() 39 const v128_t vxGHIJKLMN = wasm_i16x8_load_8x8(input_x + 16); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24() 40 const v128_t vyGHIJKLMN = wasm_i16x8_load_8x8(input_y + 16); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24() 91 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24() 92 const v128_t vy01234567 = wasm_i16x8_load_8x8(input_y); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x24()
|
D | minmax-wasmsimd-x16.c | 35 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16() 36 const v128_t vy01234567 = wasm_i16x8_load_8x8(input_y); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16() 37 const v128_t vx89ABCDEF = wasm_i16x8_load_8x8(input_x + 8); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16() 38 const v128_t vy89ABCDEF = wasm_i16x8_load_8x8(input_y + 8); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16() 76 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16() 77 const v128_t vy01234567 = wasm_i16x8_load_8x8(input_y); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x16()
|
D | minmax-wasmsimd-x8.c | 35 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8() 36 const v128_t vy01234567 = wasm_i16x8_load_8x8(input_y); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8() 65 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8() 66 const v128_t vy01234567 = wasm_i16x8_load_8x8(input_y); in xnn_qs8_vadd_minmax_ukernel__wasmsimd_x8()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x4c8-minmax-wasmsimd-ld64.c | 53 const v128_t vxa0 = wasm_i16x8_load_8x8(a0); in xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64() 56 const v128_t vxb0 = wasm_i16x8_load_8x8(w); in xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64() 61 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64() 66 const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64() 71 const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_1x4c8__wasmsimd_ld64()
|
D | 2x4c8-minmax-wasmsimd-ld64.c | 63 const v128_t vxa0 = wasm_i16x8_load_8x8(a0); in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 65 const v128_t vxa1 = wasm_i16x8_load_8x8(a1); in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 68 const v128_t vxb0 = wasm_i16x8_load_8x8(w); in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 76 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 84 const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 92 const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_2x4c8__wasmsimd_ld64()
|
D | 3x4c8-minmax-wasmsimd-ld64.c | 73 const v128_t vxa0 = wasm_i16x8_load_8x8(a0); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 75 const v128_t vxa1 = wasm_i16x8_load_8x8(a1); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 77 const v128_t vxa2 = wasm_i16x8_load_8x8(a2); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 80 const v128_t vxb0 = wasm_i16x8_load_8x8(w); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 91 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 102 const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 113 const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t))); in xnn_qs8_gemm_minmax_ukernel_3x4c8__wasmsimd_ld64()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x4c8-minmax-wasmsimd-ld64.c | 64 const v128_t vxa0 = wasm_i16x8_load_8x8(a0); in xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64() 67 const v128_t vxb0 = wasm_i16x8_load_8x8(w); in xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64() 72 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64() 77 … const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64() 82 … const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_1x4c8__wasmsimd_ld64()
|
D | 2x4c8-minmax-wasmsimd-ld64.c | 76 const v128_t vxa0 = wasm_i16x8_load_8x8(a0); in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 78 const v128_t vxa1 = wasm_i16x8_load_8x8(a1); in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 81 const v128_t vxb0 = wasm_i16x8_load_8x8(w); in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 89 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 97 … const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld64() 105 … const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_2x4c8__wasmsimd_ld64()
|
D | 3x4c8-minmax-wasmsimd-ld64.c | 88 const v128_t vxa0 = wasm_i16x8_load_8x8(a0); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 90 const v128_t vxa1 = wasm_i16x8_load_8x8(a1); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 92 const v128_t vxa2 = wasm_i16x8_load_8x8(a2); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 95 const v128_t vxb0 = wasm_i16x8_load_8x8(w); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 106 const v128_t vxb1 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 8 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 117 … const v128_t vxb2 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 16 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64() 128 … const v128_t vxb3 = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + 24 * sizeof(int8_t))); in xnn_qs8_igemm_minmax_ukernel_3x4c8__wasmsimd_ld64()
|
/external/XNNPACK/src/qs8-vaddc/gen/ |
D | minmax-wasmsimd-x32.c | 36 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32() 37 const v128_t vx89ABCDEF = wasm_i16x8_load_8x8(input_x + 8); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32() 38 const v128_t vxGHIJKLMN = wasm_i16x8_load_8x8(input_x + 16); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32() 39 const v128_t vxOPQRSTUV = wasm_i16x8_load_8x8(input_x + 24); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32() 89 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32()
|
D | minmax-wasmsimd-x24.c | 36 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24() 37 const v128_t vx89ABCDEF = wasm_i16x8_load_8x8(input_x + 8); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24() 38 const v128_t vxGHIJKLMN = wasm_i16x8_load_8x8(input_x + 16); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24() 81 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x24()
|
D | minmax-wasmsimd-x16.c | 36 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16() 37 const v128_t vx89ABCDEF = wasm_i16x8_load_8x8(input_x + 8); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16() 69 const v128_t vx01234567 = wasm_i16x8_load_8x8(input_x); in xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x16()
|
/external/XNNPACK/src/qs8-vadd/ |
D | wasmsimd.c.in | 34 const v128_t vx${ABC[0:8]} = wasm_i16x8_load_8x8(input_x); 35 const v128_t vy${ABC[0:8]} = wasm_i16x8_load_8x8(input_y); 37 const v128_t vx${ABC[N:N+8]} = wasm_i16x8_load_8x8(input_x + ${N}); 38 const v128_t vy${ABC[N:N+8]} = wasm_i16x8_load_8x8(input_y + ${N}); 90 const v128_t vx${ABC[0:8]} = wasm_i16x8_load_8x8(input_x); 91 const v128_t vy${ABC[0:8]} = wasm_i16x8_load_8x8(input_y);
|
/external/XNNPACK/src/qs8-dwconv/ |
D | unipass-wasmsimd-mul16.c.in | 52 const v128_t vi${K}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${K}); 54 const v128_t vi${K}x${ABC[C:C+8]} = wasm_i16x8_load_8x8(i${K} + ${C}); 55 …const v128_t vk${K}x${ABC[C:C+8]} = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + ${CHANNEL_T… 123 const v128_t vi${K}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${K}); 126 const v128_t vk${K}x${ABC[0:8]} = wasm_i16x8_load_8x8(k); 128 … const v128_t vk${K}x${ABC[0:8]} = wasm_i16x8_load_8x8((const void*) (k + ${K * CHANNEL_TILE})); 130 …const v128_t vk${K}x${ABC[0:8]} = wasm_i16x8_load_8x8((const void*) ((uintptr_t) w + ${CHANNEL_TIL…
|
/external/XNNPACK/src/qs8-gavgpool/ |
D | multipass-wasmsimd.c.in | 49 const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M}); 51 const v128_t vxi${M}x${ABC[C:C+8]} = wasm_i16x8_load_8x8(i${M} + ${C}); 85 const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M}); 122 const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M}); 124 const v128_t vxi${M}x${ABC[C:C+8]} = wasm_i16x8_load_8x8(i${M} + ${C}); 158 const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M}); 206 const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M}); 208 const v128_t vxi${M}x${ABC[C:C+8]} = wasm_i16x8_load_8x8(i${M} + ${C}); 284 const v128_t vxi${M}x${ABC[0:8]} = wasm_i16x8_load_8x8(i${M});
|