|
| 1 | +#include "ggml-dsp.h" |
| 2 | + |
| 3 | +inline static void ggmlhexagon_dsp_add_f32 (const int n, float * z, const float * x, const float * y) { |
| 4 | + HVX_Vector * va; |
| 5 | + HVX_Vector * vb; |
| 6 | + HVX_Vector * vc; |
| 7 | + HVX_Vector qf32; |
| 8 | + const int FLOATS_PER_VECTOR = 128 / sizeof(float); |
| 9 | + const int block = n / FLOATS_PER_VECTOR; |
| 10 | + const int left = n % FLOATS_PER_VECTOR; |
| 11 | + const int blocks = block * FLOATS_PER_VECTOR; |
| 12 | + |
| 13 | + if (0 == block) { |
| 14 | + for (size_t i = 0; i < n; ++i) |
| 15 | + z[i] = x[i] + y[i]; |
| 16 | + |
| 17 | + return; |
| 18 | + } |
| 19 | + |
| 20 | + if ((((uintptr_t)z | (uintptr_t)x | (uintptr_t)y) % ALIGN_128_BYTE) != 0) { |
| 21 | + GGMLHEXAGON_LOG_DEBUG("memaddress mismatch alignment 128 bytes z:%p x:%p y:%p", z, x, y); |
| 22 | + for (size_t i = 0; i < n; ++i) |
| 23 | + z[i] = x[i] + y[i]; |
| 24 | + |
| 25 | + return; |
| 26 | + } |
| 27 | + |
| 28 | + va = (HVX_Vector *)x; |
| 29 | + vb = (HVX_Vector *)y; |
| 30 | + vc = (HVX_Vector *)z; |
| 31 | + for (size_t i = 0; i < block; ++i) { |
| 32 | + qf32 = Q6_Vqf32_vadd_VsfVsf(*va++, *vb++); |
| 33 | + *vc = Q6_Vsf_equals_Vqf32(qf32); |
| 34 | + vc++; |
| 35 | + } |
| 36 | + |
| 37 | + if (left > 0) { |
| 38 | + for (size_t i = 0; i < left; ++i) |
| 39 | + z[i + blocks] = x[i + blocks] + y[i + blocks]; |
| 40 | + } |
| 41 | +} |
| 42 | + |
| 43 | +static void ggml_compute_forward_add_f32( |
| 44 | + const struct ggml_tensor * src0, |
| 45 | + const struct ggml_tensor * src1, |
| 46 | + struct ggml_tensor * dst) { |
| 47 | + GGMLHEXAGON_LOG_DEBUG("enter %s", __func__ ); |
| 48 | + uint64_t start_time = ggml_time_us(); |
| 49 | + |
| 50 | + memcpy(dst->ne, src1->ne, 16); |
| 51 | + memcpy(dst->nb, src1->nb, 16); |
| 52 | + ggmlhexagon_dump_tensor(src0, 1); |
| 53 | + ggmlhexagon_dump_tensor(src1, 1); |
| 54 | + ggmlhexagon_dump_tensor(dst, 1); |
| 55 | + |
| 56 | + GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); |
| 57 | + |
| 58 | + const int ith = 0; |
| 59 | + const int nth = 1; |
| 60 | + |
| 61 | + const int nr = ggml_nrows(src0); |
| 62 | + GGML_TENSOR_BINARY_OP_LOCALS |
| 63 | + |
| 64 | + GGML_ASSERT( nb0 == sizeof(float)); |
| 65 | + GGML_ASSERT(nb00 == sizeof(float)); |
| 66 | + |
| 67 | + const int dr = (nr + nth - 1)/nth; |
| 68 | + const int ir0 = dr*ith; |
| 69 | + const int ir1 = MIN(ir0 + dr, nr); |
| 70 | + if (nb10 == sizeof(float)) { |
| 71 | + for (int ir = ir0; ir < ir1; ++ir) { |
| 72 | + // src1 is broadcastable across src0 and dst in i1, i2, i3 |
| 73 | + const int32_t i03 = ir/(ne02*ne01); |
| 74 | + const int32_t i02 = (ir - i03*ne02*ne01)/ne01; |
| 75 | + const int32_t i01 = (ir - i03*ne02*ne01 - i02*ne01); |
| 76 | + |
| 77 | + const int32_t i13 = i03 % ne13; |
| 78 | + const int32_t i12 = i02 % ne12; |
| 79 | + const int32_t i11 = i01 % ne11; |
| 80 | + const int32_t nr0 = ne00 / ne10; |
| 81 | + |
| 82 | + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); |
| 83 | + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); |
| 84 | + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); |
| 85 | + for (int32_t r = 0; r < nr0; ++r) { |
| 86 | + ggmlhexagon_dsp_add_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); |
| 87 | + } |
| 88 | + } |
| 89 | + } else { |
| 90 | + // src1 is not contiguous |
| 91 | + for (int ir = ir0; ir < ir1; ++ir) { |
| 92 | + // src1 is broadcastable across src0 and dst in i1, i2, i3 |
| 93 | + const int32_t i03 = ir/(ne02*ne01); |
| 94 | + const int32_t i02 = (ir - i03*ne02*ne01)/ne01; |
| 95 | + const int32_t i01 = (ir - i03*ne02*ne01 - i02*ne01); |
| 96 | + |
| 97 | + const int32_t i13 = i03 % ne13; |
| 98 | + const int32_t i12 = i02 % ne12; |
| 99 | + const int32_t i11 = i01 % ne11; |
| 100 | + |
| 101 | + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); |
| 102 | + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); |
| 103 | + |
| 104 | + for (int32_t i0 = 0; i0 < ne0; ++i0) { |
| 105 | + const int32_t i10 = i0 % ne10; |
| 106 | + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); |
| 107 | + |
| 108 | + dst_ptr[i0] = src0_ptr[i0] + *src1_ptr; |
| 109 | + } |
| 110 | + } |
| 111 | + } |
| 112 | + |
| 113 | + uint64_t end_time = ggml_time_us(); |
| 114 | + uint64_t duration = (end_time - start_time); |
| 115 | + GGMLHEXAGON_LOG_DEBUG("duration %llu us", duration); |
| 116 | +#if !GGMLHEXAGON_DEBUG |
| 117 | + UNUSED(duration); |
| 118 | +#endif |
| 119 | + |
| 120 | + GGMLHEXAGON_LOG_DEBUG("leave %s", __func__ ); |
| 121 | +} |
| 122 | + |
| 123 | +//FIXME: why failed with test-backend-ops when disable ion rpc mempool |
| 124 | +int ggmlop_dsp_add(remote_handle64 h, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) |
| 125 | +{ |
| 126 | + GGMLHEXAGON_LOG_DEBUG("enter %s\n", __func__); |
| 127 | + switch (src0->type) { |
| 128 | + case GGML_TYPE_F32: |
| 129 | + { |
| 130 | + if (src1->type == GGML_TYPE_F32) { |
| 131 | + ggml_compute_forward_add_f32(src0, src1, dst); |
| 132 | + } else { |
| 133 | + GGML_ABORT("fatal error"); |
| 134 | + } |
| 135 | + break; |
| 136 | + } |
| 137 | + default: |
| 138 | + { |
| 139 | + GGML_ABORT("fatal error"); |
| 140 | + } |
| 141 | + } |
| 142 | + GGMLHEXAGON_LOG_DEBUG("leave %s\n", __func__); |
| 143 | + return 0; |
| 144 | +} |
0 commit comments