diff --git a/extra/bench-all.sh b/extra/bench-all.sh index 246d0a33..db042673 100755 --- a/extra/bench-all.sh +++ b/extra/bench-all.sh @@ -18,11 +18,11 @@ else fi models=( \ - "tiny" "tiny-q5_0" "tiny-q5_1" "tiny-q8_0" \ - "base" "base-q5_0" "base-q5_1" "base-q8_0" \ - "small" "small-q5_0" "small-q5_1" "small-q8_0" \ - "medium" "medium-q5_0" "medium-q5_1" "medium-q8_0" \ - "large" "large-q5_0" "large-q5_1" "large-q8_0" \ + "tiny" "tiny-q4_0" "tiny-q4_1" "tiny-q5_0" "tiny-q5_1" "tiny-q8_0" \ + "base" "base-q4_0" "base-q4_1" "base-q5_0" "base-q5_1" "base-q8_0" \ + "small" "small-q4_0" "small-q4_1" "small-q5_0" "small-q5_1" "small-q8_0" \ + "medium" "medium-q4_0" "medium-q4_1" "medium-q5_0" "medium-q5_1" "medium-q8_0" \ + "large" "large-q4_0" "large-q4_1" "large-q5_0" "large-q5_1" "large-q8_0" \ ) if [ "$encoder_only" -eq 0 ]; then diff --git a/ggml-cuda.cu b/ggml-cuda.cu index f4a67955..309866b3 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -4476,6 +4476,13 @@ static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) { *dsti = __float2half(*xi); } +static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) { + const half * xi = (const half *) cxi; + half * dsti = (half *) cdsti; + + *dsti = *xi; +} + template static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int nb00, const int nb01, const int nb02, @@ -4729,6 +4736,25 @@ static __global__ void clamp_f32(const float * x, float * dst, const float min, dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); } +static __global__ void im2col_f32_f16( + const float * x, half * dst, + int ofs0, int ofs1, int IW, int IH, int CHW, + int s0, int s1, int p0, int p1, int d0, int d1) { + const int iiw = blockIdx.z * s0 + threadIdx.z * d0 - p0; + const int iih = blockIdx.y * s1 + threadIdx.y * d1 - p1; + + const int offset_dst = + (threadIdx.x * gridDim.y * gridDim.z + blockIdx.y * gridDim.z + blockIdx.z) * CHW + + (blockIdx.x * (blockDim.y * blockDim.z) + threadIdx.y * blockDim.z + threadIdx.z); + + if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { + const int offset_src = threadIdx.x * ofs0 + blockIdx.x * ofs1; + dst[offset_dst] = __float2half(x[offset_src + iih * IW + iiw]); + } else { + dst[offset_dst] = __float2half(0.0f); + } +} + template static void get_rows_cuda(const void * x, const int32_t * y, float * dst, const int nrows, const int ncols, cudaStream_t stream) { const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1); @@ -5618,6 +5644,16 @@ static void ggml_cpy_f32_f16_cuda( (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12); } +static void ggml_cpy_f16_f16_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int nb00, const int nb01, const int nb02, + const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) { + + const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; + cpy_f32_f16<<>> + (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12); +} + static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) { const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE; scale_f32<<>>(x, dst, scale, k); @@ -5701,6 +5737,15 @@ static void soft_max_f32_cuda(const float * x, float * dst, const int ncols_x, c soft_max_f32<<>>(x, dst, ncols_x); } +static void im2col_f32_f16_cuda(const float * x, half * dst, + int OH, int IW, int IH, int OW, int IC, + int KH, int KW, int N, int ofs0, int ofs1, + int s0, int s1, int p0, int p1, int d0, int d1, cudaStream_t stream) { + dim3 block_nums(IC, OH, OW); + dim3 block_dims(N, KH, KW); + im2col_f32_f16<<>>(x, dst, ofs0, ofs1, IW, IH, (IC * KH * KW), s0, s1, p0, p1, d0, d1); +} + // buffer pool for cuda #define MAX_CUDA_BUFFERS 256 @@ -6483,7 +6528,7 @@ inline void ggml_cuda_op_mul_mat_cublas( src1_as_f16 = (half *) ggml_cuda_pool_malloc_async(ne * sizeof(half), &src1_as, id, stream); to_fp16_cuda(src1_ddf_i, src1_as_f16, ne, stream); } - const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddq_i : src1_as_f16; + const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16; size_t dst_f16_as = 0; half * dst_f16 = (half *) ggml_cuda_pool_malloc_async(row_diff*src1_ncols * sizeof(half), &dst_f16_as, id, stream); @@ -6659,6 +6704,45 @@ inline void ggml_cuda_op_alibi( (void) src1_dd; } +inline void ggml_cuda_op_im2col( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16); + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + + const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1; + + const int64_t N = src1->ne[is_2D ? 3 : 2]; + const int64_t IC = src1->ne[is_2D ? 2 : 1]; + const int64_t IH = is_2D ? src1->ne[1] : 1; + const int64_t IW = src1->ne[0]; + + const int64_t KH = is_2D ? src0->ne[1] : 1; + const int64_t KW = src0->ne[0]; + + const int64_t OH = is_2D ? dst->ne[2] : 1; + const int64_t OW = dst->ne[1]; + + const size_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32 + const size_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32 + + im2col_f32_f16_cuda(src1_dd, (half*) dst_dd, + OH, IW, IH, OW, IC, KH, KW, N, + ofs0, ofs1, s0, s1, p0, p1, d0, d1, main_stream); + + (void) src0; + (void) src0_dd; +} + inline void ggml_cuda_op_diag_mask_inf( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { @@ -7549,6 +7633,9 @@ static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, gg } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { ggml_cpy_f32_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f16_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, + ne10, ne11, nb10, nb11, nb12, main_stream); } else { fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type), ggml_type_name(src1->type)); @@ -7580,6 +7667,10 @@ static void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi); } +void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_im2col); +} + static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { (void) src0; (void) src1; @@ -7943,6 +8034,9 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_ case GGML_OP_ALIBI: func = ggml_cuda_alibi; break; + case GGML_OP_IM2COL: + func = ggml_cuda_im2col; + break; default: return false; } diff --git a/ggml-metal.m b/ggml-metal.m index 3bee8397..148c12b1 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -86,6 +86,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(rms_norm); GGML_METAL_DECL_KERNEL(norm); GGML_METAL_DECL_KERNEL(mul_mv_f32_f32); + GGML_METAL_DECL_KERNEL(mul_mv_f16_f16); GGML_METAL_DECL_KERNEL(mul_mv_f16_f32); GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_1row); GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_l4); @@ -114,6 +115,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(rope_f32); GGML_METAL_DECL_KERNEL(rope_f16); GGML_METAL_DECL_KERNEL(alibi_f32); + GGML_METAL_DECL_KERNEL(im2col_f16); GGML_METAL_DECL_KERNEL(cpy_f32_f16); GGML_METAL_DECL_KERNEL(cpy_f32_f32); GGML_METAL_DECL_KERNEL(cpy_f16_f16); @@ -287,6 +289,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(rms_norm); GGML_METAL_ADD_KERNEL(norm); GGML_METAL_ADD_KERNEL(mul_mv_f32_f32); + GGML_METAL_ADD_KERNEL(mul_mv_f16_f16); GGML_METAL_ADD_KERNEL(mul_mv_f16_f32); GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_1row); GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_l4); @@ -317,6 +320,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(rope_f32); GGML_METAL_ADD_KERNEL(rope_f16); GGML_METAL_ADD_KERNEL(alibi_f32); + GGML_METAL_ADD_KERNEL(im2col_f16); GGML_METAL_ADD_KERNEL(cpy_f32_f16); GGML_METAL_ADD_KERNEL(cpy_f32_f32); GGML_METAL_ADD_KERNEL(cpy_f16_f16); @@ -386,6 +390,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(rms_norm); GGML_METAL_DEL_KERNEL(norm); GGML_METAL_DEL_KERNEL(mul_mv_f32_f32); + GGML_METAL_DEL_KERNEL(mul_mv_f16_f16); GGML_METAL_DEL_KERNEL(mul_mv_f16_f32); GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_1row); GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_l4); @@ -416,6 +421,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(rope_f32); GGML_METAL_DEL_KERNEL(rope_f16); GGML_METAL_DEL_KERNEL(alibi_f32); + GGML_METAL_DEL_KERNEL(im2col_f16); GGML_METAL_DEL_KERNEL(cpy_f32_f16); GGML_METAL_DEL_KERNEL(cpy_f32_f32); GGML_METAL_DEL_KERNEL(cpy_f16_f16); @@ -1139,6 +1145,7 @@ void ggml_metal_graph_compute( switch (src0t) { case GGML_TYPE_F32: { + GGML_ASSERT(src1t == GGML_TYPE_F32); [encoder setComputePipelineState:ctx->pipeline_mul_mv_f32_f32]; nrows = 4; } break; @@ -1146,13 +1153,18 @@ void ggml_metal_graph_compute( { nth0 = 32; nth1 = 1; - if (ne11 * ne12 < 4) { - [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row]; - } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { - [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4]; - nrows = ne11; + if (src1t == GGML_TYPE_F32) { + if (ne11 * ne12 < 4) { + [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row]; + } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { + [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4]; + nrows = ne11; + } else { + [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32]; + nrows = 4; + } } else { - [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32]; + [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f16]; nrows = 4; } } break; @@ -1464,6 +1476,58 @@ void ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; + case GGML_OP_IM2COL: + { + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16); + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int32_t N = src1->ne[is_2D ? 3 : 2]; + const int32_t IC = src1->ne[is_2D ? 2 : 1]; + const int32_t IH = is_2D ? src1->ne[1] : 1; + const int32_t IW = src1->ne[0]; + + const int32_t KH = is_2D ? src0->ne[1] : 1; + const int32_t KW = src0->ne[0]; + + const int32_t OH = is_2D ? dst->ne[2] : 1; + const int32_t OW = dst->ne[1]; + + const int32_t CHW = IC * KH * KW; + + const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; + const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; + + switch (src0->type) { + case GGML_TYPE_F32: GGML_ASSERT(false && "not implemented"); break; + case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_im2col_f16]; break; + default: GGML_ASSERT(false); + }; + + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ofs0 length:sizeof( int32_t) atIndex:2]; + [encoder setBytes:&ofs1 length:sizeof( int32_t) atIndex:3]; + [encoder setBytes:&IW length:sizeof( int32_t) atIndex:4]; + [encoder setBytes:&IH length:sizeof( int32_t) atIndex:5]; + [encoder setBytes:&CHW length:sizeof( int32_t) atIndex:6]; + [encoder setBytes:&s0 length:sizeof( int32_t) atIndex:7]; + [encoder setBytes:&s1 length:sizeof( int32_t) atIndex:8]; + [encoder setBytes:&p0 length:sizeof( int32_t) atIndex:9]; + [encoder setBytes:&p1 length:sizeof( int32_t) atIndex:10]; + [encoder setBytes:&d0 length:sizeof( int32_t) atIndex:11]; + [encoder setBytes:&d1 length:sizeof( int32_t) atIndex:12]; + + [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)]; + } break; case GGML_OP_DUP: case GGML_OP_CPY: case GGML_OP_CONT: diff --git a/ggml-metal.metal b/ggml-metal.metal index 7c35f23a..4fdcaac9 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -792,7 +792,7 @@ kernel void kernel_mul_mv_f32_f32( constant int64_t & ne0, constant int64_t & ne1, uint3 tgpig[[threadgroup_position_in_grid]], - uint tiisg[[thread_index_in_simdgroup]]) { + uint tiisg[[thread_index_in_simdgroup]]) { const int64_t r0 = tgpig.x; const int64_t rb = tgpig.y*N_F32_F32; @@ -844,6 +844,79 @@ kernel void kernel_mul_mv_f32_f32( } } +#define N_F16_F16 4 + +kernel void kernel_mul_mv_f16_f16( + device const char * src0, + device const char * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiisg[[thread_index_in_simdgroup]]) { + + const int64_t r0 = tgpig.x; + const int64_t rb = tgpig.y*N_F16_F16; + const int64_t im = tgpig.z; + + device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); + + if (ne00 < 128) { + for (int row = 0; row < N_F16_F16; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + device const half * y = (device const half *) (src1 + r1*nb11 + im*nb12); + + float sumf = 0; + for (int i = tiisg; i < ne00; i += 32) { + sumf += (half) x[i] * (half) y[i]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } else { + device const half4 * x4 = (device const half4 *)x; + for (int row = 0; row < N_F16_F16; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + device const half * y = (device const half *) (src1 + r1*nb11 + im*nb12); + device const half4 * y4 = (device const half4 *) y; + + float sumf = 0; + for (int i = tiisg; i < ne00/4; i += 32) { + for (int k = 0; k < 4; ++k) sumf += (half) x4[i][k] * y4[i][k]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (half) x[i] * y[i]; + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } +} + kernel void kernel_mul_mv_f16_f32_1row( device const char * src0, device const char * src1, @@ -1229,6 +1302,39 @@ kernel void kernel_rope( template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope; template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope; +kernel void kernel_im2col_f16( + device const float * x, + device half * dst, + constant int32_t & ofs0, + constant int32_t & ofs1, + constant int32_t & IW, + constant int32_t & IH, + constant int32_t & CHW, + constant int32_t & s0, + constant int32_t & s1, + constant int32_t & p0, + constant int32_t & p1, + constant int32_t & d0, + constant int32_t & d1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tgpg[[threadgroups_per_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int32_t iiw = tgpig[2] * s0 + tpitg[2] * d0 - p0; + const int32_t iih = tgpig[1] * s1 + tpitg[1] * d1 - p1; + + const int32_t offset_dst = + (tpitg[0] * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW + + (tgpig[0] * (ntg[1] * ntg[2]) + tpitg[1] * ntg[2] + tpitg[2]); + + if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { + const int32_t offset_src = tpitg[0] * ofs0 + tgpig[0] * ofs1; + dst[offset_dst] = x[offset_src + iih * IW + iiw]; + } else { + dst[offset_dst] = 0.0f; + } +} + kernel void kernel_cpy_f16_f16( device const half * src0, device half * dst, diff --git a/ggml.c b/ggml.c index d1b7e94d..2723c5be 100644 --- a/ggml.c +++ b/ggml.c @@ -1634,13 +1634,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "ROPE_BACK", "ALIBI", "CLAMP", - "CONV_1D", - "CONV_1D_STAGE_0", - "CONV_1D_STAGE_1", "CONV_TRANSPOSE_1D", - "CONV_2D", - "CONV_2D_STAGE_0", - "CONV_2D_STAGE_1", + "IM2COL", "CONV_TRANSPOSE_2D", "POOL_1D", "POOL_2D", @@ -1671,7 +1666,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73"); +static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1721,13 +1716,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "rope_back(x)", "alibi(x)", "clamp(x)", - "conv_1d(x)", - "conv_1d_stage_0(x)", - "conv_1d_stage_1(x)", "conv_transpose_1d(x)", - "conv_2d(x)", - "conv_2d_stage_0(x)", - "conv_2d_stage_1(x)", + "im2col(x)", "conv_transpose_2d(x)", "pool_1d(x)", "pool_2d(x)", @@ -1758,7 +1748,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73"); +static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -1786,13 +1776,8 @@ static void ggml_setup_op_has_task_pass(void) { p[GGML_OP_GET_ROWS_BACK ] = true; p[GGML_OP_DIAG_MASK_INF ] = true; p[GGML_OP_DIAG_MASK_ZERO ] = true; - p[GGML_OP_CONV_1D ] = true; - p[GGML_OP_CONV_1D_STAGE_0 ] = true; - p[GGML_OP_CONV_1D_STAGE_1 ] = true; p[GGML_OP_CONV_TRANSPOSE_1D ] = true; - p[GGML_OP_CONV_2D ] = true; - p[GGML_OP_CONV_2D_STAGE_0 ] = true; - p[GGML_OP_CONV_2D_STAGE_1 ] = true; + p[GGML_OP_IM2COL ] = true; p[GGML_OP_CONV_TRANSPOSE_2D ] = true; p[GGML_OP_FLASH_ATTN_BACK ] = true; p[GGML_OP_CROSS_ENTROPY_LOSS ] = true; @@ -5137,80 +5122,6 @@ static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, return (ins + 2 * p - d * (ks - 1) - 1) / s + 1; } -// im2col: [N, IC, IL] => [N, OL, IC*K] -// a: [OC，IC, K] -// b: [N, IC, IL] -// result: [N, OL, IC*K] -static struct ggml_tensor * ggml_conv_1d_stage_0( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int p0, - int d0) { - GGML_ASSERT(a->ne[1] == b->ne[1]); - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t OL = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); - - const int64_t ne[4] = { - a->ne[1] * a->ne[0], - OL, - b->ne[2], - 1, - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); - - int32_t params[] = { s0, p0, d0 }; - ggml_set_op_params(result, params, sizeof(params)); - - result->op = GGML_OP_CONV_1D_STAGE_0; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; -} - -// ggml_conv_1d_stage_1 - -// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] -// a: [OC, IC, K] -// b: [N, OL, IC * K] -// result: [N, OC, OL] -static struct ggml_tensor * ggml_conv_1d_stage_1( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { - - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t ne[4] = { - b->ne[1], - a->ne[2], - b->ne[2], - 1, - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - - result->op = GGML_OP_CONV_1D_STAGE_1; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; -} - // ggml_conv_1d GGML_API struct ggml_tensor * ggml_conv_1d( @@ -5220,44 +5131,18 @@ GGML_API struct ggml_tensor * ggml_conv_1d( int s0, int p0, int d0) { - struct ggml_tensor * result = ggml_conv_1d_stage_0(ctx, a, b, s0, p0, d0); - result = ggml_conv_1d_stage_1(ctx, a, result); + struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false); // [N, OL, IC * K] + + struct ggml_tensor * result = + ggml_mul_mat(ctx, + ggml_reshape_2d(ctx, im2col, im2col->ne[0], (im2col->ne[2] * im2col->ne[1])), // [N, OL, IC * K] => [N*OL, IC * K] + ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1]), a->ne[2])); // [OC，IC, K] => [OC, IC * K] + + result = ggml_reshape_3d(ctx, result, im2col->ne[1], a->ne[2], im2col->ne[2]); // [N, OC, OL] + return result; } -// GGML_API struct ggml_tensor * ggml_conv_1d( -// struct ggml_context * ctx, -// struct ggml_tensor * a, -// struct ggml_tensor * b, -// int s0, -// int p0, -// int d0) { -// GGML_ASSERT(ggml_is_matrix(b)); -// GGML_ASSERT(a->ne[1] == b->ne[1]); -// bool is_node = false; - -// if (a->grad || b->grad) { -// GGML_ASSERT(false); // TODO: implement backward -// is_node = true; -// } - -// const int64_t ne[4] = { -// ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0), -// a->ne[2], 1, 1, -// }; -// struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne); - -// int32_t params[] = { s0, p0, d0 }; -// ggml_set_op_params(result, params, sizeof(params)); - -// result->op = GGML_OP_CONV_1D; -// result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; -// result->src[0] = a; -// result->src[1] = b; - -// return result; -// } - // ggml_conv_1d_ph struct ggml_tensor* ggml_conv_1d_ph( @@ -5319,7 +5204,7 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( // a: [OC，IC, KH, KW] // b: [N, IC, IH, IW] // result: [N, OH, OW, IC*KH*KW] -static struct ggml_tensor * ggml_conv_2d_stage_0( +struct ggml_tensor * ggml_im2col( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5328,9 +5213,14 @@ static struct ggml_tensor * ggml_conv_2d_stage_0( int p0, int p1, int d0, - int d1) { + int d1, + bool is_2D) { - GGML_ASSERT(a->ne[2] == b->ne[2]); + if(is_2D) { + GGML_ASSERT(a->ne[2] == b->ne[2]); + } else { + GGML_ASSERT(a->ne[1] == b->ne[1]); + } bool is_node = false; if (a->grad || b->grad) { @@ -5338,81 +5228,51 @@ static struct ggml_tensor * ggml_conv_2d_stage_0( is_node = true; } - const int64_t OH = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1); - const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); + const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0; + const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); const int64_t ne[4] = { - a->ne[2] * a->ne[1] * a->ne[0], + is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0], OW, - OH, - b->ne[3], + is_2D ? OH : b->ne[2], + is_2D ? b->ne[3] : 1, }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); - int32_t params[] = { s0, s1, p0, p1, d0, d1 }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); + int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_CONV_2D_STAGE_0; + result->op = GGML_OP_IM2COL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; return result; - -} - -// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] -// a: [OC, IC, KH, KW] -// b: [N, OH, OW, IC * KH * KW] -// result: [N, OC, OH, OW] -static struct ggml_tensor * ggml_conv_2d_stage_1( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { - - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t ne[4] = { - b->ne[1], - b->ne[2], - a->ne[3], - b->ne[3], - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - - result->op = GGML_OP_CONV_2D_STAGE_1; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; - } // a: [OC，IC, KH, KW] // b: [N, IC, IH, IW] // result: [N, OC, OH, OW] struct ggml_tensor * ggml_conv_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1) { + struct ggml_tensor * result = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true); // [N, OH, OW, IC * KH * KW] - struct ggml_tensor * result = ggml_conv_2d_stage_0(ctx, a, b, s0, s1, p0, p1, d0, d1); // [N, OH, OW, IC * KH * KW] - result = ggml_conv_2d_stage_1(ctx, a, result); + result = + ggml_reshape_4d(ctx, + ggml_mul_mat(ctx, + ggml_reshape_2d(ctx, result, result->ne[0], result->ne[3] * result->ne[2] * result->ne[1]), // [N, OH, OW, IC * KH * KW] => [N*OH*OW, IC * KH * KW] + ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1] * a->ne[2]), a->ne[3])), // [OC，IC, KH, KW] => [OC, IC * KH * KW] + result->ne[1], result->ne[2], a->ne[3], result->ne[3]); // [N, OC, OH, OW] return result; - } // ggml_conv_2d_sk_p0 @@ -9507,6 +9367,8 @@ static bool ggml_compute_forward_mul_mat_use_blas( // TODO: find the optimal values for these if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && + src0->type == GGML_TYPE_F32 && + src1->type == GGML_TYPE_F32 && (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) { /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/ @@ -9517,6 +9379,7 @@ static bool ggml_compute_forward_mul_mat_use_blas( } #endif + static void ggml_compute_forward_mul_mat( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -9545,7 +9408,7 @@ static void ggml_compute_forward_mul_mat( // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == ggml_type_size(type)); - GGML_ASSERT(nb10 == sizeof(float)); + GGML_ASSERT(nb10 == ggml_type_size(src1->type)); // dst cannot be transposed or permuted GGML_ASSERT(nb0 == sizeof(float)); @@ -11637,416 +11500,6 @@ static void ggml_compute_forward_rope_back( } } -// ggml_compute_forward_conv_1d - -static void ggml_compute_forward_conv_1d_f16_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00; - - // size of the convolution row - the kernel size unrolled across all input channels - const int ew0 = nk*ne01; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int64_t i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i11*nb11); - ggml_fp16_t * dst_data = wdata; - - for (int64_t i0 = 0; i0 < ne0; i0++) { - for (int64_t ik = 0; ik < nk; ik++) { - const int idx0 = i0*s0 + ik*d0 - p0; - - if(!(idx0 < 0 || idx0 >= ne10)) { - dst_data[i0*ew0 + i11*nk + ik] = GGML_FP32_TO_FP16(src[idx0]); - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // total rows in dst - const int nr = ne2; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int i2 = 0; i2 < ne2; i2++) { - for (int i1 = ir0; i1 < ir1; i1++) { - float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1); - - for (int i0 = 0; i0 < ne0; i0++) { - ggml_vec_dot_f16(ew0, dst_data + i0, - (ggml_fp16_t *) ((char *) src0->data + i1*nb02), - (ggml_fp16_t *) wdata + i2*nb2 + i0*ew0); - } - } - } -} - -static void ggml_compute_forward_conv_1d_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00; - - const int ew0 = nk*ne01; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(float)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - float * const wdata = (float *) params->wdata + 0; - - for (int64_t i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i11*nb11); - float * dst_data = wdata; - - for (int64_t i0 = 0; i0 < ne0; i0++) { - for (int64_t ik = 0; ik < nk; ik++) { - const int idx0 = i0*s0 + ik*d0 - p0; - - if(!(idx0 < 0 || idx0 >= ne10)) { - dst_data[i0*ew0 + i11*nk + ik] = src[idx0]; - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // total rows in dst - const int nr = ne02; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - float * const wdata = (float *) params->wdata + 0; - - for (int i2 = 0; i2 < ne2; i2++) { - for (int i1 = ir0; i1 < ir1; i1++) { - float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1); - - for (int i0 = 0; i0 < ne0; i0++) { - ggml_vec_dot_f32(ew0, dst_data + i0, - (float *) ((char *) src0->data + i1*nb02), - (float *) wdata + i2*nb2 + i0*ew0); - } - } - } -} - -// TODO: reuse ggml_mul_mat or implement ggml_im2col and remove stage_0 and stage_1 -static void gemm_f16_out_f32(int64_t m, int64_t n, int64_t k, - ggml_fp16_t * A, - ggml_fp16_t * B, - float * C, - const int ith, const int nth) { - // does not seem to make a difference - int64_t m0, m1, n0, n1; - // patches per thread - if (m > n) { - n0 = 0; - n1 = n; - - // total patches in dst - const int np = m; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - m0 = dp*ith; - m1 = MIN(m0 + dp, np); - } else { - m0 = 0; - m1 = m; - - // total patches in dst - const int np = n; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - n0 = dp*ith; - n1 = MIN(n0 + dp, np); - } - - // block-tiling attempt - int64_t blck_n = 16; - int64_t blck_m = 16; - - // int64_t CACHE_SIZE = 2 * 1024 * 1024; // 2MB - // int64_t blck_size = CACHE_SIZE / (sizeof(float) + 2 * sizeof(ggml_fp16_t) * K); - // if (blck_size > 0) { - // blck_0 = 4; - // blck_1 = blck_size / blck_0; - // if (blck_1 < 0) { - // blck_1 = 1; - // } - // // blck_0 = (int64_t)sqrt(blck_size); - // // blck_1 = blck_0; - // } - // // printf("%zd %zd %zd %zd\n", blck_size, K, blck_0, blck_1); - - for (int j = n0; j < n1; j+=blck_n) { - for (int i = m0; i < m1; i+=blck_m) { - // printf("i j k => %d %d %d\n", i, j, K); - for (int ii = i; ii < i + blck_m && ii < m1; ii++) { - for (int jj = j; jj < j + blck_n && jj < n1; jj++) { - ggml_vec_dot_f16(k, - C + ii*n + jj, - A + ii * k, - B + jj * k); - } - } - } - } -} - -// src0: kernel [OC, IC, K] -// src1: signal [N, IC, IL] -// dst: result [N, OL, IC*K] -static void ggml_compute_forward_conv_1d_stage_0_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F16); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS; - - const int64_t N = ne12; - const int64_t IC = ne11; - const int64_t IL = ne10; - - const int64_t K = ne00; - - const int64_t OL = ne1; - - const int ith = params->ith; - const int nth = params->nth; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(dst->data, 0, ggml_nbytes(dst)); - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // im2col: [N, IC, IL] => [N, OL, IC*K] - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; - - for (int64_t in = 0; in < N; in++) { - for (int64_t iol = 0; iol < OL; iol++) { - for (int64_t iic = ith; iic < IC; iic+=nth) { - - // micro kernel - ggml_fp16_t * dst_data = wdata + (in*OL + iol)*(IC*K); // [IC, K] - const float * const src_data = (float *)((char *) src1->data + in*nb12 + iic*nb11); // [IL] - - for (int64_t ik = 0; ik < K; ik++) { - const int64_t iil = iol*s0 + ik*d0 - p0; - - if (!(iil < 0 || iil >= IL)) { - dst_data[iic*K + ik] = GGML_FP32_TO_FP16(src_data[iil]); - } - } - } - } - } - } -} - -// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] -// src0: [OC, IC, K] -// src1: [N, OL, IC * K] -// result: [N, OC, OL] -static void ggml_compute_forward_conv_1d_stage_1_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - if (params->type == GGML_TASK_INIT) { - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - GGML_TENSOR_BINARY_OP_LOCALS; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb0 == sizeof(float)); - - const int N = ne12; - const int OL = ne11; - - const int OC = ne02; - const int IC = ne01; - const int K = ne00; - - const int ith = params->ith; - const int nth = params->nth; - - int64_t m = OC; - int64_t n = OL; - int64_t k = IC * K; - - // [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m, n] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_1d( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_f16_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_conv_1d_f32(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_1d_stage_0( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_stage_0_f32(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_1d_stage_1( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_stage_1_f16(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - // ggml_compute_forward_conv_transpose_1d static void ggml_compute_forward_conv_transpose_1d_f16_f32( @@ -12258,12 +11711,10 @@ static void ggml_compute_forward_conv_transpose_1d( } } -// ggml_compute_forward_conv_2d - // src0: kernel [OC, IC, KH, KW] // src1: image [N, IC, IH, IW] // dst: result [N, OH, OW, IC*KH*KW] -static void ggml_compute_forward_conv_2d_stage_0_f32( +static void ggml_compute_forward_im2col_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -12277,28 +11728,30 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( GGML_TENSOR_BINARY_OP_LOCALS; - const int64_t N = ne13; - const int64_t IC = ne12; - const int64_t IH = ne11; - const int64_t IW = ne10; - - // const int64_t OC = ne03; - // const int64_t IC = ne02; - const int64_t KH = ne01; - const int64_t KW = ne00; - - const int64_t OH = ne2; - const int64_t OW = ne1; + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; const int ith = params->ith; const int nth = params->nth; - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + const int64_t N = is_2D ? ne13 : ne12; + const int64_t IC = is_2D ? ne12 : ne11; + const int64_t IH = is_2D ? ne11 : 1; + const int64_t IW = ne10; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne2 : 1; + const int64_t OW = ne1; + + int ofs0 = is_2D ? nb13 : nb12; + int ofs1 = is_2D ? nb12 : nb11; GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb10 == sizeof(float)); @@ -12317,15 +11770,15 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; for (int64_t in = 0; in < N; in++) { - for (int64_t ioh = 0; ioh < OH; ioh++) { + for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 for (int64_t iow = 0; iow < OW; iow++) { - for (int64_t iic = ith; iic < IC; iic+=nth) { + for (int64_t iic = ith; iic < IC; iic += nth) { // micro kernel ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW] + const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] - for (int64_t ikh = 0; ikh < KH; ikh++) { + for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 for (int64_t ikw = 0; ikw < KW; ikw++) { const int64_t iiw = iow*s0 + ikw*d0 - p0; const int64_t iih = ioh*s1 + ikh*d1 - p1; @@ -12342,180 +11795,7 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( } } -// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] -// src0: [OC, IC, KH, KW] -// src1: [N, OH, OW, IC * KH * KW] -// result: [N, OC, OH, OW] -static void ggml_compute_forward_conv_2d_stage_1_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - if (params->type == GGML_TASK_INIT) { - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - GGML_TENSOR_BINARY_OP_LOCALS; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb0 == sizeof(float)); - - const int N = ne13; - const int OH = ne12; - const int OW = ne11; - - const int OC = ne03; - const int IC = ne02; - const int KH = ne01; - const int KW = ne00; - - const int ith = params->ith; - const int nth = params->nth; - - int64_t m = OC; - int64_t n = OH * OW; - int64_t k = IC * KH * KW; - - // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m, n] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_2d_f16_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - // src1: image [N, IC, IH, IW] - // src0: kernel [OC, IC, KH, KW] - // dst: result [N, OC, OH, OW] - // ne12: IC - // ne0: OW - // ne1: OH - // nk0: KW - // nk1: KH - // ne13: N - - const int N = ne13; - const int IC = ne12; - const int IH = ne11; - const int IW = ne10; - - const int OC = ne03; - // const int IC = ne02; - const int KH = ne01; - const int KW = ne00; - - const int OH = ne1; - const int OW = ne0; - - const int ith = params->ith; - const int nth = params->nth; - - // const int nk0 = ne00; - // const int nk1 = ne01; - - // size of the convolution row - the kernel size unrolled across all channels - // const int ew0 = nk0*nk1*ne02; - // ew0: IC*KH*KW - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - // prepare source data (src1) - // im2col: [N, IC, IH, IW] => [N*OH*OW, IC*KH*KW] - - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int in = 0; in < N; in++) { - for (int iic = 0; iic < IC; iic++) { - for (int ioh = 0; ioh < OH; ioh++) { - for (int iow = 0; iow < OW; iow++) { - - // micro kernel - ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW] - - for (int ikh = 0; ikh < KH; ikh++) { - for (int ikw = 0; ikw < KW; ikw++) { - const int iiw = iow*s0 + ikw*d0 - p0; - const int iih = ioh*s1 + ikh*d1 - p1; - - if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { - dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); - } - } - } - } - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - // wdata: [N*OH*OW, IC*KH*KW] - // dst: result [N, OC, OH, OW] - // src0: kernel [OC, IC, KH, KW] - - int64_t m = OC; - int64_t n = OH * OW; - int64_t k = IC * KH * KW; - - // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)wdata + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m * k] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_2d( +static void ggml_compute_forward_im2col( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -12523,50 +11803,7 @@ static void ggml_compute_forward_conv_2d( switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_conv_2d_f16_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - //ggml_compute_forward_conv_2d_f32(params, src0, src1, dst); - GGML_ASSERT(false); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_2d_stage_0( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_2d_stage_0_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - GGML_ASSERT(false); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_2d_stage_1( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_2d_stage_1_f16(params, src0, src1, dst); + ggml_compute_forward_im2col_f16(params, src0, src1, dst); } break; case GGML_TYPE_F32: { @@ -14783,33 +14020,13 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_clamp(params, tensor->src[0], tensor); } break; - case GGML_OP_CONV_1D: - { - ggml_compute_forward_conv_1d(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - ggml_compute_forward_conv_1d_stage_0(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - ggml_compute_forward_conv_1d_stage_1(params, tensor->src[0], tensor->src[1], tensor); - } break; case GGML_OP_CONV_TRANSPOSE_1D: { ggml_compute_forward_conv_transpose_1d(params, tensor->src[0], tensor->src[1], tensor); } break; - case GGML_OP_CONV_2D: + case GGML_OP_IM2COL: { - ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - ggml_compute_forward_conv_2d_stage_0(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_2D_STAGE_1: - { - ggml_compute_forward_conv_2d_stage_1(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_im2col(params, tensor->src[0], tensor->src[1], tensor); } break; case GGML_OP_CONV_TRANSPOSE_2D: { @@ -15780,31 +14997,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // TODO: not implemented } break; - case GGML_OP_CONV_1D: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - GGML_ASSERT(false); // TODO: not implemented - } break; case GGML_OP_CONV_TRANSPOSE_1D: { GGML_ASSERT(false); // TODO: not implemented } break; - case GGML_OP_CONV_2D: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_2D_STAGE_1: + case GGML_OP_IM2COL: { GGML_ASSERT(false); // TODO: not implemented } break; @@ -16533,31 +15730,11 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { { n_tasks = 1; //TODO } break; - case GGML_OP_CONV_1D: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - n_tasks = n_threads; - } break; case GGML_OP_CONV_TRANSPOSE_1D: { n_tasks = n_threads; } break; - case GGML_OP_CONV_2D: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONV_2D_STAGE_1: + case GGML_OP_IM2COL: { n_tasks = n_threads; } break; @@ -16642,6 +15819,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { } break; default: { + printf("%s: op %s not implemented\n", __func__, ggml_op_name(node->op)); GGML_ASSERT(false); } break; } @@ -16844,38 +16022,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; } } break; - case GGML_OP_CONV_1D: - { - GGML_ASSERT(node->src[0]->ne[3] == 1); - GGML_ASSERT(node->src[1]->ne[2] == 1); - GGML_ASSERT(node->src[1]->ne[3] == 1); - - const int64_t ne00 = node->src[0]->ne[0]; - const int64_t ne01 = node->src[0]->ne[1]; - const int64_t ne02 = node->src[0]->ne[2]; - - const int64_t ne10 = node->src[1]->ne[0]; - const int64_t ne11 = node->src[1]->ne[1]; - - const int64_t ne0 = node->ne[0]; - const int64_t ne1 = node->ne[1]; - const int64_t nk = ne00; - const int64_t ew0 = nk * ne01; - - UNUSED(ne02); - UNUSED(ne10); - UNUSED(ne11); - - if (node->src[0]->type == GGML_TYPE_F16 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(ggml_fp16_t)*(ne0*ne1*ew0); - } else if (node->src[0]->type == GGML_TYPE_F32 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(float)*(ne0*ne1*ew0); - } else { - GGML_ASSERT(false); - } - } break; case GGML_OP_CONV_TRANSPOSE_1D: { GGML_ASSERT(node->src[0]->ne[3] == 1); @@ -16901,37 +16047,9 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { GGML_ASSERT(false); } } break; - case GGML_OP_CONV_2D: + case GGML_OP_IM2COL: { - const int64_t ne00 = node->src[0]->ne[0]; // W - const int64_t ne01 = node->src[0]->ne[1]; // H - const int64_t ne02 = node->src[0]->ne[2]; // C - const int64_t ne03 = node->src[0]->ne[3]; // N - - const int64_t ne10 = node->src[1]->ne[0]; // W - const int64_t ne11 = node->src[1]->ne[1]; // H - const int64_t ne12 = node->src[1]->ne[2]; // C - - const int64_t ne0 = node->ne[0]; - const int64_t ne1 = node->ne[1]; - const int64_t ne2 = node->ne[2]; - const int64_t ne3 = node->ne[3]; - const int64_t nk = ne00*ne01; - const int64_t ew0 = nk * ne02; - - UNUSED(ne03); - UNUSED(ne2); - - if (node->src[0]->type == GGML_TYPE_F16 && - node->src[1]->type == GGML_TYPE_F32) { - // im2col: [N*OH*OW, IC*KH*KW] - cur = sizeof(ggml_fp16_t)*(ne3*ne0*ne1*ew0); - } else if (node->src[0]->type == GGML_TYPE_F32 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(float)* (ne10*ne11*ne12); - } else { - GGML_ASSERT(false); - } + n_tasks = n_threads; } break; case GGML_OP_CONV_TRANSPOSE_2D: { diff --git a/ggml.h b/ggml.h index e56a8337..52ae6755 100644 --- a/ggml.h +++ b/ggml.h @@ -403,13 +403,8 @@ extern "C" { GGML_OP_ROPE_BACK, GGML_OP_ALIBI, GGML_OP_CLAMP, - GGML_OP_CONV_1D, - GGML_OP_CONV_1D_STAGE_0, // internal - GGML_OP_CONV_1D_STAGE_1, // internal GGML_OP_CONV_TRANSPOSE_1D, - GGML_OP_CONV_2D, - GGML_OP_CONV_2D_STAGE_0, // internal - GGML_OP_CONV_2D_STAGE_1, // internal + GGML_OP_IM2COL, GGML_OP_CONV_TRANSPOSE_2D, GGML_OP_POOL_1D, GGML_OP_POOL_2D, @@ -1398,6 +1393,18 @@ extern "C" { float min, float max); + GGML_API struct ggml_tensor * ggml_im2col( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1, + bool is_2D); + GGML_API struct ggml_tensor * ggml_conv_1d( struct ggml_context * ctx, struct ggml_tensor * a, diff --git a/whisper.cpp b/whisper.cpp index 26ac965d..80ca5c9b 100644 --- a/whisper.cpp +++ b/whisper.cpp @@ -588,7 +588,6 @@ struct whisper_kv_cache { }; struct whisper_model_data { - ggml_backend_buffer_t buffer_conv; // TODO: tmp until GPU support for conv ggml_backend_buffer_t buffer_main; }; @@ -818,23 +817,9 @@ struct whisper_context { whisper_state * state = nullptr; - ggml_backend_t backend_cpu = nullptr; - ggml_backend_t backend_gpu = nullptr; + ggml_backend_t backend = nullptr; std::string path_model; // populated by whisper_init_from_file_with_params() - - ggml_backend_t backend_kv() const { - return backend_gpu ? backend_gpu : backend_cpu; - } - - // TODO: always on CPU until we have a GPU support for conv - ggml_backend_t backend_conv() const { - return backend_cpu; - } - - ggml_backend_t backend_main() const { - return backend_gpu ? backend_gpu : backend_cpu; - } }; struct whisper_global { @@ -1192,16 +1177,16 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con // encoder { - model.e_pe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_audio_state, n_audio_ctx); + model.e_pe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_audio_state, n_audio_ctx); - model.e_conv_1_w = ggml_new_tensor_3d(ctx, vtype, 3, n_mels, n_audio_state); - model.e_conv_1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n_audio_state); + model.e_conv_1_w = ggml_new_tensor_3d(ctx, vtype, 3, n_mels, n_audio_state); + model.e_conv_1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2*n_audio_ctx, n_audio_state); - model.e_conv_2_w = ggml_new_tensor_3d(ctx, vtype, 3, n_audio_state, n_audio_state); - model.e_conv_2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n_audio_state); + model.e_conv_2_w = ggml_new_tensor_3d(ctx, vtype, 3, n_audio_state, n_audio_state); + model.e_conv_2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_audio_ctx, n_audio_state); - model.e_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state); - model.e_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state); + model.e_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state); + model.e_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state); // map by name model.tensors["encoder.positional_embedding"] = model.e_pe; @@ -1394,45 +1379,30 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con #endif if (backend_gpu) { - wctx.backend_gpu = backend_gpu; + wctx.backend = backend_gpu; } else { - wctx.backend_gpu = nullptr; + wctx.backend = ggml_backend_cpu_init(); } - - // always add the CPU backend as a fallback - wctx.backend_cpu = ggml_backend_cpu_init(); } { - size_t size_conv = 0; size_t size_main = 0; for (const auto & t : model.tensors) { - if (t.first.find("conv") != std::string::npos) { - size_conv += ggml_nbytes(t.second) + ggml_tensor_overhead(); - } else { - size_main += ggml_nbytes(t.second) + ggml_tensor_overhead(); - } + size_main += ggml_nbytes(t.second) + ggml_tensor_overhead(); } - model.data->buffer_conv = ggml_backend_alloc_buffer(wctx.backend_conv(), size_conv); - model.data->buffer_main = ggml_backend_alloc_buffer(wctx.backend_main(), size_main); + model.data->buffer_main = ggml_backend_alloc_buffer(wctx.backend, size_main); - WHISPER_LOG_INFO("%s: %8s buffer size = %8.2f MB\n", __func__, ggml_backend_name(wctx.backend_conv()), size_conv / 1024.0 / 1024.0); - WHISPER_LOG_INFO("%s: %8s buffer size = %8.2f MB\n", __func__, ggml_backend_name(wctx.backend_main()), size_main / 1024.0 / 1024.0); + WHISPER_LOG_INFO("%s: %8s buffer size = %8.2f MB\n", __func__, ggml_backend_name(wctx.backend), size_main / 1024.0 / 1024.0); } - ggml_allocr * alloc_conv = ggml_allocr_new_from_buffer(model.data->buffer_conv); ggml_allocr * alloc_main = ggml_allocr_new_from_buffer(model.data->buffer_main); // allocate tensors in the backend buffers { for (const auto & t : model.tensors) { - if (t.first.find("conv") != std::string::npos) { - ggml_allocr_alloc(alloc_conv, t.second); - } else { - ggml_allocr_alloc(alloc_main, t.second); - } + ggml_allocr_alloc(alloc_main, t.second); } } @@ -1475,45 +1445,67 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con } auto tensor = model.tensors[name.data()]; - if (ggml_nelements(tensor) != nelements) { - WHISPER_LOG_ERROR("%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); - WHISPER_LOG_ERROR("%s: shape: [%d, %d, %d], expected: [%d, %d, %d]\n", - __func__, ne[0], ne[1], ne[2], (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2]); - return false; + + const bool is_conv_bias = (name == "encoder.conv1.bias" || name == "encoder.conv2.bias"); + + if (!is_conv_bias) { + if (ggml_nelements(tensor) != nelements) { + WHISPER_LOG_ERROR("%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); + WHISPER_LOG_ERROR("%s: shape: [%d, %d, %d], expected: [%d, %d, %d]\n", + __func__, ne[0], ne[1], ne[2], (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2]); + return false; + } + + if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1] || tensor->ne[2] != ne[2]) { + WHISPER_LOG_ERROR("%s: tensor '%s' has wrong shape in model file: got [%d, %d, %d], expected [%d, %d, %d]\n", + __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], ne[0], ne[1], ne[2]); + return false; + } + + const size_t bpe = ggml_type_size(ggml_type(ttype)); + + if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) { + WHISPER_LOG_ERROR("%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", + __func__, name.data(), ggml_nbytes(tensor), nelements*bpe); + return false; + } } - if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1] || tensor->ne[2] != ne[2]) { - WHISPER_LOG_ERROR("%s: tensor '%s' has wrong shape in model file: got [%d, %d, %d], expected [%d, %d, %d]\n", - __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], ne[0], ne[1], ne[2]); - return false; - } - - const size_t bpe = ggml_type_size(ggml_type(ttype)); - - if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) { - WHISPER_LOG_ERROR("%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", - __func__, name.data(), ggml_nbytes(tensor), nelements*bpe); - return false; - } - - const bool is_conv = name.find("conv") != std::string::npos; - - ggml_backend * backend = is_conv ? wctx.backend_conv() : wctx.backend_main(); + ggml_backend_t backend = wctx.backend; //printf("%s: [%5.5s] %s\n", __func__, ggml_backend_name(backend), name.c_str()); - if (ggml_backend_is_cpu(backend) + if ((ggml_backend_is_cpu(backend) #ifdef GGML_USE_METAL || ggml_backend_is_metal(backend) #endif - ) { + ) && !is_conv_bias) { // for the CPU and Metal backend, we can read directly into the tensor loader->read(loader->context, tensor->data, ggml_nbytes(tensor)); BYTESWAP_TENSOR(tensor); } else { // read into a temporary buffer first, then copy to device memory read_buf.resize(ggml_nbytes(tensor)); - loader->read(loader->context, read_buf.data(), read_buf.size()); + + // we repeat the 2 bias tensors along dim 0: + // [1, 512] -> [3000, 512] (conv1.bias) + // [1, 512] -> [1500, 512] (conv2.bias) + if (is_conv_bias) { + loader->read(loader->context, read_buf.data(), read_buf.size() / tensor->ne[0]); + + float * data_f32 = (float *) read_buf.data(); + for (int64_t y = 0; y < tensor->ne[1]; ++y) { + const int64_t yy = tensor->ne[1] - y - 1; + const float val = data_f32[yy]; + + for (int64_t x = 0; x < tensor->ne[0]; ++x) { + data_f32[yy*tensor->ne[0] + x] = val; + } + } + } else { + loader->read(loader->context, read_buf.data(), read_buf.size()); + } + ggml_backend_tensor_set(tensor, read_buf.data(), 0, ggml_nbytes(tensor)); } @@ -1532,7 +1524,6 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con } } - ggml_allocr_free(alloc_conv); ggml_allocr_free(alloc_main); wctx.t_load_us = ggml_time_us() - t_start_us; @@ -1595,7 +1586,7 @@ static struct ggml_cgraph * whisper_build_graph_conv( float * dst = wstate.inp_mel.data(); memset(dst, 0, ggml_nbytes(mel)); - const int i0 = std::min(mel_offset, mel_inp.n_len); + const int i0 = std::min(mel_offset, mel_inp.n_len); const int i1 = std::min(mel_offset + 2*n_ctx, mel_inp.n_len); for (int j = 0; j < mel_inp.n_mel; ++j) { @@ -1613,20 +1604,22 @@ static struct ggml_cgraph * whisper_build_graph_conv( // convolution + gelu { cur = ggml_conv_1d_ph(ctx0, model.e_conv_1_w, mel, 1, 1); - cur = ggml_add(ctx0, - ggml_repeat(ctx0, - model.e_conv_1_b, - cur), - cur); + cur = ggml_add(ctx0, cur, model.e_conv_1_b); + //cur = ggml_add(ctx0, + // ggml_repeat(ctx0, + // model.e_conv_1_b, + // cur), + // cur); cur = ggml_gelu(ctx0, cur); cur = ggml_conv_1d_ph(ctx0, model.e_conv_2_w, cur, 2, 1); - cur = ggml_add(ctx0, - ggml_repeat(ctx0, - model.e_conv_2_b, - cur), - cur); + cur = ggml_add(ctx0, cur, model.e_conv_2_b); + //cur = ggml_add(ctx0, + // ggml_repeat(ctx0, + // model.e_conv_2_b, + // cur), + // cur); cur = ggml_gelu(ctx0, cur); } @@ -1685,6 +1678,14 @@ static struct ggml_cgraph * whisper_build_graph_encoder( ggml_allocr * alloc = wstate.alloc_encode.alloc; + //struct ggml_tensor * cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_ctx, n_state); + //ggml_allocr_alloc(alloc, cur); + + //if (!ggml_allocr_is_measure(alloc)) { + // ggml_backend_tensor_copy(wstate.embd_conv, cur); + //} + struct ggml_tensor * cur = ggml_view_tensor(ctx0, wstate.embd_conv); + struct ggml_tensor * KQscale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); ggml_allocr_alloc(alloc, KQscale); @@ -1693,13 +1694,6 @@ static struct ggml_cgraph * whisper_build_graph_encoder( ggml_backend_tensor_set(KQscale, &val, 0, sizeof(float)); } - struct ggml_tensor * cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_ctx, n_state); - ggml_allocr_alloc(alloc, cur); - - if (!ggml_allocr_is_measure(alloc)) { - ggml_backend_tensor_copy(wstate.embd_conv, cur); - } - // =================================================================== // NOTE: experimenting with partial evaluation of the encoder (ignore) //static int iter = -1; @@ -1939,12 +1933,13 @@ static struct ggml_cgraph * whisper_build_graph_cross( ggml_allocr * alloc = wstate.alloc_cross.alloc; - struct ggml_tensor * cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx); - ggml_allocr_alloc(alloc, cur); + //struct ggml_tensor * cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx); + //ggml_allocr_alloc(alloc, cur); - if (!ggml_allocr_is_measure(alloc)) { - ggml_backend_tensor_copy(wstate.embd_enc, cur); - } + //if (!ggml_allocr_is_measure(alloc)) { + // ggml_backend_tensor_copy(wstate.embd_enc, cur); + //} + struct ggml_tensor * cur = ggml_view_tensor(ctx0, wstate.embd_enc); struct ggml_tensor * Kscale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); ggml_allocr_alloc(alloc, Kscale); @@ -2022,15 +2017,15 @@ static bool whisper_encode_internal( ggml_allocr_alloc_graph(alloc, gf); if (!whisper_encode_external(wstate)) { - if (ggml_backend_is_cpu(wctx.backend_conv())) { - ggml_backend_cpu_set_n_threads(wctx.backend_conv(), n_threads); + if (ggml_backend_is_cpu(wctx.backend)) { + ggml_backend_cpu_set_n_threads(wctx.backend, n_threads); } #ifdef GGML_USE_METAL - if (ggml_backend_is_metal(wctx.backend_conv())) { - ggml_backend_metal_set_n_cb(wctx.backend_conv(), n_threads); + if (ggml_backend_is_metal(wctx.backend)) { + ggml_backend_metal_set_n_cb(wctx.backend, n_threads); } #endif - ggml_backend_graph_compute(wctx.backend_conv(), gf); + ggml_backend_graph_compute(wctx.backend, gf); } } @@ -2044,15 +2039,15 @@ static bool whisper_encode_internal( ggml_allocr_alloc_graph(alloc, gf); - if (ggml_backend_is_cpu(wctx.backend_main())) { - ggml_backend_cpu_set_n_threads(wctx.backend_main(), n_threads); + if (ggml_backend_is_cpu(wctx.backend)) { + ggml_backend_cpu_set_n_threads(wctx.backend, n_threads); } #ifdef GGML_USE_METAL - if (ggml_backend_is_metal(wctx.backend_main())) { - ggml_backend_metal_set_n_cb(wctx.backend_main(), n_threads); + if (ggml_backend_is_metal(wctx.backend)) { + ggml_backend_metal_set_n_cb(wctx.backend, n_threads); } #endif - ggml_backend_graph_compute(wctx.backend_main(), gf); + ggml_backend_graph_compute(wctx.backend, gf); } // cross @@ -2065,15 +2060,15 @@ static bool whisper_encode_internal( ggml_allocr_alloc_graph(alloc, gf); - if (ggml_backend_is_cpu(wctx.backend_main())) { - ggml_backend_cpu_set_n_threads(wctx.backend_main(), n_threads); + if (ggml_backend_is_cpu(wctx.backend)) { + ggml_backend_cpu_set_n_threads(wctx.backend, n_threads); } #ifdef GGML_USE_METAL - if (ggml_backend_is_metal(wctx.backend_main())) { - ggml_backend_metal_set_n_cb(wctx.backend_main(), n_threads); + if (ggml_backend_is_metal(wctx.backend)) { + ggml_backend_metal_set_n_cb(wctx.backend, n_threads); } #endif - ggml_backend_graph_compute(wctx.backend_main(), gf); + ggml_backend_graph_compute(wctx.backend, gf); } wstate.t_encode_us += ggml_time_us() - t_start_us; @@ -2464,15 +2459,15 @@ static bool whisper_decode_internal( logits = gf->nodes[gf->n_nodes - 1]; - if (ggml_backend_is_cpu(wctx.backend_main())) { - ggml_backend_cpu_set_n_threads(wctx.backend_main(), n_threads); + if (ggml_backend_is_cpu(wctx.backend)) { + ggml_backend_cpu_set_n_threads(wctx.backend, n_threads); } #ifdef GGML_USE_METAL - if (ggml_backend_is_metal(wctx.backend_main())) { - ggml_backend_metal_set_n_cb(wctx.backend_main(), n_threads); + if (ggml_backend_is_metal(wctx.backend)) { + ggml_backend_metal_set_n_cb(wctx.backend, n_threads); } #endif - ggml_backend_graph_compute(wctx.backend_main(), gf); + ggml_backend_graph_compute(wctx.backend, gf); } // extract logits for all N tokens @@ -2915,7 +2910,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) { whisper_state * state = new whisper_state; - if (!kv_cache_init(ctx->model.hparams, state->decoders[0].kv_self, ctx->backend_kv(), ctx->itype, ctx->model.hparams.n_text_ctx)) { + if (!kv_cache_init(ctx->model.hparams, state->decoders[0].kv_self, ctx->backend, ctx->itype, ctx->model.hparams.n_text_ctx)) { WHISPER_LOG_ERROR("%s: kv_cache_init() failed for self-attention cache\n", __func__); delete state; return nullptr; @@ -2926,7 +2921,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) { WHISPER_LOG_INFO("%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0); } - if (!kv_cache_init(ctx->model.hparams, state->kv_cross, ctx->backend_kv(), ctx->itype, ctx->model.hparams.n_audio_ctx)) { + if (!kv_cache_init(ctx->model.hparams, state->kv_cross, ctx->backend, ctx->itype, ctx->model.hparams.n_audio_ctx)) { WHISPER_LOG_ERROR("%s: kv_cache_init() failed for cross-attention cache\n", __func__); delete state; return nullptr; @@ -2968,7 +2963,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) { // conv allocator { - whisper_allocr_graph_init(state->alloc_conv, ctx->backend_conv(), + whisper_allocr_graph_init(state->alloc_conv, ctx->backend, [&]() { return whisper_build_graph_conv(*ctx, *state, 0); }); @@ -2978,7 +2973,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) { // encoder allocator if (!whisper_encode_external(*state)) { - whisper_allocr_graph_init(state->alloc_encode, ctx->backend_main(), + whisper_allocr_graph_init(state->alloc_encode, ctx->backend, [&]() { return whisper_build_graph_encoder(*ctx, *state); }); @@ -2988,7 +2983,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) { // cross allocator { - whisper_allocr_graph_init(state->alloc_cross, ctx->backend_main(), + whisper_allocr_graph_init(state->alloc_cross, ctx->backend, [&]() { return whisper_build_graph_cross(*ctx, *state); }); @@ -2998,7 +2993,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) { // decoder allocator { - whisper_allocr_graph_init(state->alloc_decode, ctx->backend_main(), + whisper_allocr_graph_init(state->alloc_decode, ctx->backend, [&]() { const auto & hparams = ctx->model.hparams; @@ -3273,7 +3268,6 @@ void whisper_free(struct whisper_context * ctx) { ggml_free(ctx->model.ctx); } if (ctx->model.data) { - ggml_backend_buffer_free(ctx->model.data->buffer_conv); ggml_backend_buffer_free(ctx->model.data->buffer_main); delete ctx->model.data; @@ -3281,11 +3275,7 @@ void whisper_free(struct whisper_context * ctx) { whisper_free_state(ctx->state); - ggml_backend_free(ctx->backend_cpu); - - if (ctx->backend_gpu) { - ggml_backend_free(ctx->backend_gpu); - } + ggml_backend_free(ctx->backend); delete ctx; } @@ -4583,7 +4573,7 @@ int whisper_full_with_state( if (decoder.kv_self.ctx == nullptr) { decoder.kv_self = state->decoders[0].kv_self; - if (!kv_cache_reinit(decoder.kv_self, ctx->backend_kv())) { + if (!kv_cache_reinit(decoder.kv_self, ctx->backend)) { WHISPER_LOG_ERROR("%s: kv_cache_reinit() failed for self-attention, decoder %d\n", __func__, j); return -4; }