From be8c9c245bd129ebabb80e0a7a8dd7daeb4d30af Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Thu, 7 Sep 2023 15:45:01 +0200 Subject: [PATCH] metal : parallel RoPE on Metal (#3024) * Parallel RoPE on metal * PR suggestion --------- Co-authored-by: Iwan Kawrakow --- ggml-metal.m | 2 +- ggml-metal.metal | 26 ++++++++++++++------------ 2 files changed, 15 insertions(+), 13 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 521ca180f..7e2355ce6 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -1141,7 +1141,7 @@ void ggml_metal_graph_compute( [encoder setBytes:&freq_base length:sizeof(float) atIndex:21]; [encoder setBytes:&freq_scale length:sizeof(float) atIndex:22]; - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)]; } break; case GGML_OP_DUP: case GGML_OP_CPY: diff --git a/ggml-metal.metal b/ggml-metal.metal index 5edf6d521..5070561fb 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -682,25 +682,27 @@ kernel void kernel_rope( constant int & mode, constant float & freq_base, constant float & freq_scale, - uint3 tpig[[thread_position_in_grid]]) { - const int64_t i3 = tpig[2]; - const int64_t i2 = tpig[1]; - const int64_t i1 = tpig[0]; + uint tiitg[[thread_index_in_threadgroup]], + uint3 tptg[[threads_per_threadgroup]], + uint3 tgpig[[threadgroup_position_in_grid]]) { + const int64_t i3 = tgpig[2]; + const int64_t i2 = tgpig[1]; + const int64_t i1 = tgpig[0]; const bool is_neox = mode & 2; - const float theta_scale = pow(freq_base, -2.0f/n_dims); const int64_t p = ((mode & 1) == 0 ? n_past + i2 : i2); - float theta = freq_scale * (float)p; + const float theta_0 = freq_scale * (float)p; + const float inv_ndims = -1.f/n_dims; if (!is_neox) { - for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) { + + const float theta = theta_0 * pow(freq_base, inv_ndims*i0); const float cos_theta = cos(theta); const float sin_theta = sin(theta); - theta *= theta_scale; - device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); device float * dst_data = (device float *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -712,12 +714,12 @@ kernel void kernel_rope( } } else { for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { - for (int64_t ic = 0; ic < n_dims; ic += 2) { + for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) { + + const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib); const float cos_theta = cos(theta); const float sin_theta = sin(theta); - theta *= theta_scale; - const int64_t i0 = ib*n_dims + ic/2; device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);