Skip to content

Commit ee4725a

Browse files
slarenslaren
authored
ggml : group mul_mat_id rows by matrix (cpu only) (#4480)
* ggml : group mul_mat_id rows by matrix (cpu only) * remove mmid parameters from mm forward * store row groups in wdata and calculate only once in GGML_TASK_INIT ggml-ci
1 parent 6744dbe commit ee4725a

File tree

1 file changed

+195
-42
lines changed

1 file changed

+195
-42
lines changed

‎ggml.c

Lines changed: 195 additions & 42 deletions
Original file line numberDiff line numberDiff line change
@@ -9580,16 +9580,11 @@ static bool ggml_compute_forward_mul_mat_use_blas(
95809580
}
95819581
#endif
95829582

9583-
// off1 = offset in i11 and i1
9584-
// cne1 = ne11 and ne1
9585-
// in a normal matrix multiplication, off1 = 0 and cne1 = ne1
9586-
// during GGML_TASK_INIT, the full src1 is converted regardless of off1 and cne1
95879583
static void ggml_compute_forward_mul_mat(
95889584
const struct ggml_compute_params * params,
95899585
const struct ggml_tensor * src0,
95909586
const struct ggml_tensor * src1,
9591-
struct ggml_tensor * dst,
9592-
int64_t off1, int64_t cne1) {
9587+
struct ggml_tensor * dst) {
95939588
int64_t t0 = ggml_perf_time_us();
95949589
UNUSED(t0);
95959590

@@ -9657,9 +9652,9 @@ static void ggml_compute_forward_mul_mat(
96579652
const int64_t i03 = i13/r3;
96589653
const int64_t i02 = i12/r2;
96599654

9660-
const void * x = (char *) src0->data + i02*nb02 + i03*nb03;
9661-
const float * y = (float *) ((char *) src1->data + off1*nb11 + i12*nb12 + i13*nb13);
9662-
float * d = (float *) ((char *) dst->data + off1*nb1 + i12*nb2 + i13*nb3);
9655+
const void * x = (char *) src0->data + i02*nb02 + i03*nb03;
9656+
const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
9657+
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
96639658

96649659
if (type != GGML_TYPE_F32) {
96659660
float * const wdata = params->wdata;
@@ -9676,7 +9671,7 @@ static void ggml_compute_forward_mul_mat(
96769671
}
96779672

96789673
cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
9679-
cne1, ne01, ne10,
9674+
ne1, ne01, ne10,
96809675
1.0f, y, ne10,
96819676
x, ne00,
96829677
0.0f, d, ne01);
@@ -9717,8 +9712,8 @@ static void ggml_compute_forward_mul_mat(
97179712
const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
97189713
const size_t row_size = ggml_row_size(vec_dot_type, ne10);
97199714

9720-
const int64_t nr0 = ne01; // src0 rows
9721-
const int64_t nr1 = cne1*ne12*ne13; // src1 rows
9715+
const int64_t nr0 = ne01; // src0 rows
9716+
const int64_t nr1 = ne1*ne12*ne13; // src1 rows
97229717

97239718
//printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
97249719

@@ -9760,9 +9755,9 @@ static void ggml_compute_forward_mul_mat(
97609755
for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
97619756
for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
97629757
for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
9763-
const int64_t i13 = (ir1/(ne12*cne1));
9764-
const int64_t i12 = (ir1 - i13*ne12*cne1)/cne1;
9765-
const int64_t i11 = (ir1 - i13*ne12*cne1 - i12*cne1) + off1;
9758+
const int64_t i13 = (ir1/(ne12*ne1));
9759+
const int64_t i12 = (ir1 - i13*ne12*ne1)/ne1;
9760+
const int64_t i11 = (ir1 - i13*ne12*ne1 - i12*ne1);
97669761

97679762
// broadcast src0 into src1
97689763
const int64_t i03 = i13/r3;
@@ -9802,28 +9797,191 @@ static void ggml_compute_forward_mul_mat(
98029797

98039798
static void ggml_compute_forward_mul_mat_id(
98049799
const struct ggml_compute_params * params,
9805-
const struct ggml_tensor * src0,
9800+
const struct ggml_tensor * ids,
98069801
const struct ggml_tensor * src1,
98079802
struct ggml_tensor * dst) {
98089803

9809-
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
9810-
// during GGML_TASK_INIT the entire src1 is converted to vec_dot_type
9811-
ggml_compute_forward_mul_mat(params, dst->src[2], src1, dst, 0, dst->ne[1]);
9812-
return;
9813-
}
9804+
const struct ggml_tensor * src0 = dst->src[2]; // only for GGML_TENSOR_BINARY_OP_LOCALS
9805+
9806+
GGML_TENSOR_BINARY_OP_LOCALS
9807+
9808+
const int ith = params->ith;
9809+
const int nth = params->nth;
9810+
9811+
const enum ggml_type type = src0->type;
9812+
9813+
const bool src1_cont = ggml_is_contiguous(src1);
9814+
9815+
ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot;
9816+
enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type;
9817+
ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
9818+
9819+
GGML_ASSERT(ne0 == ne01);
9820+
GGML_ASSERT(ne1 == ne11);
9821+
GGML_ASSERT(ne2 == ne12);
9822+
GGML_ASSERT(ne3 == ne13);
9823+
9824+
// we don't support permuted src0 or src1
9825+
GGML_ASSERT(nb00 == ggml_type_size(type));
9826+
GGML_ASSERT(nb10 == ggml_type_size(src1->type));
9827+
9828+
// dst cannot be transposed or permuted
9829+
GGML_ASSERT(nb0 == sizeof(float));
9830+
GGML_ASSERT(nb0 <= nb1);
9831+
GGML_ASSERT(nb1 <= nb2);
9832+
GGML_ASSERT(nb2 <= nb3);
98149833

9815-
const struct ggml_tensor * ids = src0;
9834+
// broadcast factors
9835+
const int64_t r2 = ne12/ne02;
9836+
const int64_t r3 = ne13/ne03;
9837+
9838+
// row groups
98169839
const int id = ggml_get_op_params_i32(dst, 0);
98179840
const int n_as = ggml_get_op_params_i32(dst, 1);
98189841

9819-
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
9820-
const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
9842+
char * wdata_src1_end = (src1->type == vec_dot_type) ?
9843+
(char *) params->wdata :
9844+
(char *) params->wdata + GGML_PAD(ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t));
9845+
9846+
int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
9847+
int64_t * matrix_rows = matrix_row_counts + n_as; // [n_as][ne11]
9848+
9849+
#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne11 + (i1)]
98219850

9822-
GGML_ASSERT(row_id >= 0 && row_id < n_as);
9851+
if (params->type == GGML_TASK_INIT) {
9852+
char * wdata = params->wdata;
9853+
if (src1->type != vec_dot_type) {
9854+
const size_t row_size = ggml_row_size(vec_dot_type, ne10);
98239855

9824-
const struct ggml_tensor * src0_row = dst->src[row_id + 2];
9825-
ggml_compute_forward_mul_mat(params, src0_row, src1, dst, i01, 1);
9856+
assert(params->wsize >= ne11*ne12*ne13*row_size);
9857+
assert(src1->type == GGML_TYPE_F32);
9858+
9859+
for (int64_t i13 = 0; i13 < ne13; ++i13) {
9860+
for (int64_t i12 = 0; i12 < ne12; ++i12) {
9861+
for (int64_t i11 = 0; i11 < ne11; ++i11) {
9862+
from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
9863+
wdata += row_size;
9864+
}
9865+
}
9866+
}
9867+
}
9868+
9869+
// initialize matrix_row_counts
9870+
GGML_ASSERT(wdata == wdata_src1_end);
9871+
memset(matrix_row_counts, 0, n_as*sizeof(int64_t));
9872+
9873+
// group rows by src0 matrix
9874+
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
9875+
const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
9876+
9877+
GGML_ASSERT(row_id >= 0 && row_id < n_as);
9878+
MMID_MATRIX_ROW(row_id, matrix_row_counts[row_id]) = i01;
9879+
matrix_row_counts[row_id] += 1;
9880+
}
9881+
9882+
return;
98269883
}
9884+
9885+
if (params->type == GGML_TASK_FINALIZE) {
9886+
return;
9887+
}
9888+
9889+
// compute each matrix multiplication in sequence
9890+
for (int cur_a = 0; cur_a < n_as; ++cur_a) {
9891+
const int64_t cne1 = matrix_row_counts[cur_a];
9892+
9893+
if (cne1 == 0) {
9894+
continue;
9895+
}
9896+
9897+
const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
9898+
9899+
const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
9900+
const size_t row_size = ggml_row_size(vec_dot_type, ne10);
9901+
9902+
const int64_t nr0 = ne01; // src0 rows
9903+
const int64_t nr1 = cne1*ne12*ne13; // src1 rows
9904+
9905+
//printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
9906+
9907+
// distribute the thread work across the inner or outer loop based on which one is larger
9908+
9909+
const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
9910+
const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
9911+
9912+
const int64_t ith0 = ith % nth0;
9913+
const int64_t ith1 = ith / nth0;
9914+
9915+
const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
9916+
const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
9917+
9918+
const int64_t ir010 = dr0*ith0;
9919+
const int64_t ir011 = MIN(ir010 + dr0, nr0);
9920+
9921+
const int64_t ir110 = dr1*ith1;
9922+
const int64_t ir111 = MIN(ir110 + dr1, nr1);
9923+
9924+
//printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
9925+
9926+
// threads with no work simply yield (not sure if it helps)
9927+
if (ir010 >= ir011 || ir110 >= ir111) {
9928+
sched_yield();
9929+
continue;
9930+
}
9931+
9932+
assert(ne12 % ne02 == 0);
9933+
assert(ne13 % ne03 == 0);
9934+
9935+
// block-tiling attempt
9936+
const int64_t blck_0 = 16;
9937+
const int64_t blck_1 = 16;
9938+
9939+
// attempt to reduce false-sharing (does not seem to make a difference)
9940+
float tmp[16];
9941+
9942+
for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
9943+
for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
9944+
for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
9945+
const int64_t i13 = (ir1/(ne12*cne1)); // Note: currently, src1 is always a matrix
9946+
const int64_t i12 = (ir1 - i13*ne12*cne1)/cne1;
9947+
const int64_t _i11 = (ir1 - i13*ne12*cne1 - i12*cne1);
9948+
const int64_t i11 = MMID_MATRIX_ROW(cur_a, _i11);
9949+
9950+
// broadcast src0 into src1
9951+
const int64_t i03 = i13/r3;
9952+
const int64_t i02 = i12/r2;
9953+
9954+
const int64_t i1 = i11;
9955+
const int64_t i2 = i12;
9956+
const int64_t i3 = i13;
9957+
9958+
const char * src0_row = (const char *) src0_cur->data + (0 + i02*nb02 + i03*nb03);
9959+
9960+
// desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
9961+
// if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
9962+
// the original src1 data pointer, so we should index using the indices directly
9963+
// TODO: this is a bit of a hack, we should probably have a better way to handle this
9964+
const char * src1_col = (const char *) wdata +
9965+
(src1_cont || src1->type != vec_dot_type
9966+
? (i11 + i12*ne11 + i13*ne12*ne11)*row_size
9967+
: (i11*nb11 + i12*nb12 + i13*nb13));
9968+
9969+
float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
9970+
9971+
//for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
9972+
// vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
9973+
//}
9974+
9975+
for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
9976+
vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
9977+
}
9978+
memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
9979+
}
9980+
}
9981+
}
9982+
}
9983+
9984+
#undef MMID_MATRIX_ROW
98279985
}
98289986

98299987
// ggml_compute_forward_out_prod
@@ -14191,7 +14349,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
1419114349
} break;
1419214350
case GGML_OP_MUL_MAT:
1419314351
{
14194-
ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor, 0, tensor->ne[1]);
14352+
ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor);
1419514353
} break;
1419614354
case GGML_OP_MUL_MAT_ID:
1419714355
{
@@ -15991,7 +16149,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
1599116149
} break;
1599216150
case GGML_OP_MUL_MAT_ID:
1599316151
{
15994-
// FIXME: blas
1599516152
n_tasks = n_threads;
1599616153
} break;
1599716154
case GGML_OP_OUT_PROD:
@@ -16325,20 +16482,16 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
1632516482
} break;
1632616483
case GGML_OP_MUL_MAT_ID:
1632716484
{
16328-
const struct ggml_tensor * a = node->src[2];
16329-
const struct ggml_tensor * b = node->src[1];
16330-
const enum ggml_type vec_dot_type = type_traits[a->type].vec_dot_type;
16331-
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
16332-
if (ggml_compute_forward_mul_mat_use_blas(a, b, node)) {
16333-
if (a->type != GGML_TYPE_F32) {
16334-
// here we need memory just for single 2D matrix from src0
16335-
cur = ggml_type_size(GGML_TYPE_F32)*(a->ne[0]*a->ne[1]);
16336-
}
16337-
} else
16338-
#endif
16339-
if (b->type != vec_dot_type) {
16340-
cur = ggml_row_size(vec_dot_type, ggml_nelements(b));
16485+
const struct ggml_tensor * src0 = node->src[2];
16486+
const struct ggml_tensor * src1 = node->src[1];
16487+
const enum ggml_type vec_dot_type = type_traits[src0->type].vec_dot_type;
16488+
if (src1->type != vec_dot_type) {
16489+
cur = ggml_row_size(vec_dot_type, ggml_nelements(src1));
1634116490
}
16491+
const int n_as = ggml_get_op_params_i32(node, 1);
16492+
cur = GGML_PAD(cur, sizeof(int64_t)); // align
16493+
cur += n_as * sizeof(int64_t); // matrix_row_counts
16494+
cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows
1634216495
} break;
1634316496
case GGML_OP_OUT_PROD:
1634416497
{

0 commit comments

Comments
 (0)