CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

sycl: Cleanup codepaths in Get Rows in sycl backend (#14215)

Browse Source 
Addresses unused reorder path
This commit is contained in:
Anton Mitkov
2025-06-19 11:40:21 +01:00
committed by GitHub
parent fffcce535e
commit 600e3e9b50
1 changed files with 2 additions and 96 deletions
Download Patch File Download Diff File Expand all files Collapse all files

98
ggml/src/ggml-sycl/getrows.cpp
View File

@@ -60,54 +60,6 @@ static void k_get_rows(
dst_row[iybs + iqs + y_offset] = v.y(); dst_row[iybs + iqs + y_offset] = v.y();
} }
template<int qk, int qr, dequantize_kernel_t_reorder dequantize_kernel_recorder, typename dst_t>
static void k_get_rows_reorder(
const void * src0, const void *src0_dq, const int32_t * src1, dst_t * dst,
int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
/*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
/*size_t s0,*/ size_t s1, size_t s2, size_t s3,
/*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
size_t s10, size_t s11, size_t s12,
const sycl::nd_item<3> &item_ct1/*, size_t s13*/) {
const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) +
item_ct1.get_local_id(2)) *
2;
const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
item_ct1.get_local_id(1);
const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
item_ct1.get_local_id(0)) /
ne12;
const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
item_ct1.get_local_id(0)) %
ne12;
if (i00 >= ne00) {
return;
}
auto ncols = ne00;
const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
const int src0_off = i01 * ncols + i00;
const int ib = src0_off / QK4_0; // block index
const int iqs = (i00%qk)/qr; // x quant index
const int iybs = i00 - i00%qk; // dst block start index
const int y_offset = qr == 1 ? 1 : qk/2;
// dequantize
dfloat2 v;
dequantize_kernel_recorder((const void *)src0_dq, ib, (const void *)src0, src0_off/2, v);
dst_row[iybs + iqs + 0] = v.x();
dst_row[iybs + iqs + y_offset] = v.y();
GGML_UNUSED(nb01);
GGML_UNUSED(nb02);
GGML_UNUSED(nb03);
}
template<typename src0_t, typename dst_t> template<typename src0_t, typename dst_t>
static void k_get_rows_float( static void k_get_rows_float(
const src0_t * src0, const int32_t * src1, dst_t * dst, const src0_t * src0, const int32_t * src1, dst_t * dst,
@@ -177,47 +129,6 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
GGML_UNUSED(ctx); GGML_UNUSED(ctx);
} }
template <int qk, int qr, dequantize_kernel_t_reorder dq_reorder>
static void get_rows_sycl_reorder(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const void *src0_dd,
const int32_t *src1_dd, float *dst_dd,
queue_ptr stream) {
GGML_TENSOR_BINARY_OP_LOCALS
const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE);
const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE);
const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x);
// strides in elements
//const size_t s0 = nb0 / ggml_element_size(dst);
const size_t s1 = nb1 / ggml_element_size(dst);
const size_t s2 = nb2 / ggml_element_size(dst);
const size_t s3 = nb3 / ggml_element_size(dst);
const size_t s10 = nb10 / ggml_element_size(src1);
const size_t s11 = nb11 / ggml_element_size(src1);
const size_t s12 = nb12 / ggml_element_size(src1);
//const size_t s13 = nb13 / ggml_element_size(src1);
GGML_ASSERT(ne00 % 2 == 0);
const uint8_t* src0_q = (const uint8_t*)src0_dd;
const size_t ncols = ne00;
const size_t nrows = ne01;
const sycl::half* src0_dq = (const sycl::half*)(src0_q + nrows * ncols / 2);
stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
k_get_rows_reorder<qk, qr, dq_reorder>(
src0_dd, src0_dq, src1_dd, dst_dd, ne00, ne12, s1, s2,
s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
});
GGML_UNUSED(dst);
GGML_UNUSED(ctx);
}
template <typename src0_t> template <typename src0_t>
static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst, const ggml_tensor *src1, ggml_tensor *dst,
@@ -277,13 +188,8 @@ void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
src1_i32, (float *)dst->data, ctx.stream()); src1_i32, (float *)dst->data, ctx.stream());
break; break;
case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_0:
if (ctx.opt_feature.reorder && dst->op == GGML_OP_MUL_MAT) { get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
get_rows_sycl_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, src1_i32, (float *)dst->data, ctx.stream());
src1_i32, (float *)dst->data, ctx.stream());
} else {
get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
src1_i32, (float *)dst->data, ctx.stream());
}
break; break;
case GGML_TYPE_Q4_1: case GGML_TYPE_Q4_1:
get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,