Skip to content

Commit 96fdca0

Browse files
relent95relent95
authored
Vulkan: add conv_transpose_2d operation (#16022)
* Vulkan: add conv_transpose_2d operation * Vulkan: fix typo in conv_transpose_2d shader(s0mp, s0L, s1mp, s1L) * Vulkan: fix incorrect indentation in conv_transpose_2d shader * Vulkan: add checking the push constants size limit and reuse conv2d_mm.comp for conv_transpose_2d operation * Vulkan: revert the order of the index calculation and bound check in conv_2d shader * Vulkan: explicity check push constants limit in supports_op() for conv_transpose_2d operation. * Vulkan: remove unnecessary lower bound checks for H/W_idx in the conv_2d shader.
1 parent b2d980f commit 96fdca0

File tree

4 files changed

+229
-38
lines changed

4 files changed

+229
-38
lines changed

‎ggml/src/ggml-vulkan/ggml-vulkan.cpp‎

Lines changed: 185 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -574,6 +574,8 @@ struct vk_device_struct {
574574
vk_pipeline pipeline_opt_step_sgd_f32;
575575
vk_pipeline pipeline_conv2d_f32[CONV_SHAPE_COUNT];
576576
vk_pipeline pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
577+
vk_pipeline pipeline_conv_transpose_2d_f32[CONV_SHAPE_COUNT];
578+
vk_pipeline pipeline_conv_transpose_2d_f16_f32[CONV_SHAPE_COUNT];
577579
vk_pipeline pipeline_conv2d_dw_whcn_f32, pipeline_conv2d_dw_whcn_f16_f32;
578580
vk_pipeline pipeline_conv2d_dw_cwhn_f32, pipeline_conv2d_dw_cwhn_f16_f32;
579581

@@ -1117,6 +1119,56 @@ template <> void init_pushconst_fastdiv(vk_op_conv2d_push_constants &p) {
11171119
init_fastdiv_values(p.OW*p.OH, p.OWOHmp, p.OWOHL);
11181120
}
11191121

1122+
struct vk_op_conv_transpose_2d_push_constants {
1123+
uint32_t Cout;
1124+
uint32_t Cin;
1125+
uint32_t N;
1126+
1127+
uint32_t KW;
1128+
uint32_t KH;
1129+
uint32_t W;
1130+
uint32_t H;
1131+
uint32_t OW;
1132+
uint32_t OH;
1133+
1134+
uint32_t s0;
1135+
uint32_t s1;
1136+
uint32_t p0;
1137+
uint32_t p1;
1138+
uint32_t d0;
1139+
uint32_t d1;
1140+
1141+
uint32_t nb01;
1142+
uint32_t nb02;
1143+
uint32_t nb03;
1144+
1145+
uint32_t nb11;
1146+
uint32_t nb12;
1147+
uint32_t nb13;
1148+
1149+
uint32_t nb1;
1150+
uint32_t nb2;
1151+
uint32_t nb3;
1152+
1153+
// init_fastdiv_values constants for dividing by KW, KW*KH, OW, OW*OH, s0, s1
1154+
uint32_t KWmp; uint32_t KWL;
1155+
uint32_t KWKHmp; uint32_t KWKHL;
1156+
uint32_t OWmp; uint32_t OWL;
1157+
uint32_t OWOHmp; uint32_t OWOHL;
1158+
uint32_t s0mp; uint32_t s0L;
1159+
uint32_t s1mp; uint32_t s1L;
1160+
};
1161+
1162+
template <> void init_pushconst_fastdiv(vk_op_conv_transpose_2d_push_constants &p) {
1163+
// Compute magic values to divide by KW, KW*KH, OW, OW*OH, s0, s1
1164+
init_fastdiv_values(p.KW, p.KWmp, p.KWL);
1165+
init_fastdiv_values(p.KW*p.KH, p.KWKHmp, p.KWKHL);
1166+
init_fastdiv_values(p.OW, p.OWmp, p.OWL);
1167+
init_fastdiv_values(p.OW*p.OH, p.OWOHmp, p.OWOHL);
1168+
init_fastdiv_values(p.s0, p.s0mp, p.s0L);
1169+
init_fastdiv_values(p.s1, p.s1mp, p.s1L);
1170+
}
1171+
11201172
struct vk_op_conv2d_dw_push_constants {
11211173
uint32_t ne;
11221174
uint32_t batches;
@@ -1322,7 +1374,7 @@ class vk_perf_logger {
13221374
flops[name].push_back(m * n * (k + (k - 1)) * batch);
13231375
return;
13241376
}
1325-
if (node->op == GGML_OP_CONV_2D) {
1377+
if (node->op == GGML_OP_CONV_2D || node->op == GGML_OP_CONV_TRANSPOSE_2D) {
13261378
std::string name = ggml_op_name(node->op);
13271379
ggml_tensor * knl = node->src[0];
13281380
uint64_t OW = node->ne[0];
@@ -1331,7 +1383,7 @@ class vk_perf_logger {
13311383
uint64_t Cout = node->ne[2];
13321384
uint64_t KW = knl->ne[0];
13331385
uint64_t KH = knl->ne[1];
1334-
uint64_t Cin = knl->ne[2];
1386+
uint64_t Cin = node->src[1]->ne[2];
13351387
// KxCRS @ CRSxNPQ = KxNPQ -> M=K, K=CRS, N=NPQ
13361388
uint64_t size_M = Cout;
13371389
uint64_t size_K = Cin * KW * KH;
@@ -3492,7 +3544,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
34923544

34933545
ggml_vk_create_pipeline(device, device->pipeline_opt_step_sgd_f32, "opt_step_sgd_f32", opt_step_sgd_f32_len, opt_step_sgd_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
34943546

3495-
// conv2d
3547+
// conv2d, conv_transpose_2d
34963548
for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
34973549
uint32_t conv2d_WG_SIZE = 256;
34983550
uint32_t conv2d_BS_K = 128;
@@ -3567,31 +3619,30 @@ static void ggml_vk_load_shaders(vk_device& device) {
35673619
std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
35683620
std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
35693621

3622+
#define CREATE_CONV(name, type_suffix, spv_suffix) \
3623+
ggml_vk_create_pipeline( \
3624+
device, device->pipeline_##name##type_suffix[s], #name #type_suffix, \
3625+
name##type_suffix##spv_suffix##_len, name##type_suffix##spv_suffix##_data, "main", 3, \
3626+
sizeof(vk_op_##name##_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
3627+
#define CREATE_CONVS(spv_suffix) \
3628+
CREATE_CONV(conv2d, _f32, spv_suffix) \
3629+
CREATE_CONV(conv2d, _f16_f32, spv_suffix) \
3630+
if (device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_conv_transpose_2d_push_constants)) { \
3631+
CREATE_CONV(conv_transpose_2d, _f32, spv_suffix) \
3632+
CREATE_CONV(conv_transpose_2d, _f16_f32, spv_suffix) \
3633+
}
35703634
#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
35713635
if (device->coopmat2) {
3572-
ggml_vk_create_pipeline(
3573-
device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_cm2_len, conv2d_f32_cm2_data, "main", 3,
3574-
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
3575-
ggml_vk_create_pipeline(
3576-
device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_cm2_len, conv2d_f16_f32_cm2_data, "main", 3,
3577-
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
3636+
CREATE_CONVS(_cm2)
35783637
} else
35793638
#endif
35803639
if (conv2d_UNROLL) {
3581-
ggml_vk_create_pipeline(
3582-
device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
3583-
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
3584-
ggml_vk_create_pipeline(
3585-
device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_unroll_len, conv2d_f16_f32_unroll_data, "main", 3,
3586-
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
3640+
CREATE_CONVS(_unroll)
35873641
} else {
3588-
ggml_vk_create_pipeline(
3589-
device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
3590-
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
3591-
ggml_vk_create_pipeline(
3592-
device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
3593-
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
3642+
CREATE_CONVS( )
35943643
}
3644+
#undef CREATE_CONV
3645+
#undef CREATE_CONVS
35953646
}
35963647

35973648
ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
@@ -7548,6 +7599,33 @@ static std::array<uint32_t, 3> ggml_vk_get_conv_elements(const ggml_tensor *dst)
75487599
return elements;
75497600
}
75507601

7602+
static std::array<uint32_t, 3> ggml_vk_get_conv_transpose_2d_elements(const ggml_tensor *dst) {
7603+
const ggml_tensor *src0 = dst->src[0];
7604+
const ggml_tensor *src1 = dst->src[1];
7605+
7606+
// src0 - kernel: [KW, KH, Cout, Cin]
7607+
// src1 - input: [W, H, Cin, N]
7608+
// dst - result: [OW, OH, Cout, N]
7609+
7610+
auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
7611+
return (ins - 1) * s - 2 * p + (ks - 1) * d + 1;
7612+
};
7613+
// parallelize in {OW/BS_K, OH/BS_NPQ, 1}
7614+
int64_t W = src1->ne[0];
7615+
int64_t H = src1->ne[1];
7616+
int64_t KW = src0->ne[0];
7617+
int64_t KH = src0->ne[1];
7618+
int64_t Cout = src0->ne[2];
7619+
int64_t N = src1->ne[3];
7620+
int64_t OH = calc_conv_output_size(H, KH, dst->op_params[0], 0, 1);
7621+
int64_t OW = calc_conv_output_size(W, KW, dst->op_params[0], 0, 1);
7622+
int64_t NPQ = N * OW * OH;
7623+
7624+
// Tile output matrix to (K/NB_K, NPQ/NB_NPQ, 1) workgroups
7625+
std::array<uint32_t, 3> elements = { static_cast<uint32_t>(Cout), static_cast<uint32_t>(NPQ), 1 };
7626+
return elements;
7627+
}
7628+
75517629
static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * dst, ggml_op op) {
75527630
switch (op) {
75537631
case GGML_OP_GET_ROWS:
@@ -7925,9 +8003,12 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
79258003
}
79268004
return nullptr;
79278005
case GGML_OP_CONV_2D:
8006+
case GGML_OP_CONV_TRANSPOSE_2D:
79288007
if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
79298008
ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
7930-
auto elements = ggml_vk_get_conv_elements(dst);
8009+
std::array<uint32_t, 3> elements;
8010+
if (op == GGML_OP_CONV_2D) elements = ggml_vk_get_conv_elements(dst);
8011+
else if (op == GGML_OP_CONV_TRANSPOSE_2D) elements = ggml_vk_get_conv_transpose_2d_elements(dst);
79318012
vk_conv_shapes shape;
79328013

79338014
uint32_t tiles[CONV_SHAPE_COUNT];
@@ -7947,10 +8028,18 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
79478028
shape = CONV_SHAPE_64x32;
79488029
}
79498030

7950-
if (src0->type == GGML_TYPE_F32) {
7951-
return ctx->device->pipeline_conv2d_f32[shape];
7952-
} else if (src0->type == GGML_TYPE_F16) {
7953-
return ctx->device->pipeline_conv2d_f16_f32[shape];
8031+
if (op == GGML_OP_CONV_2D) {
8032+
if (src0->type == GGML_TYPE_F32) {
8033+
return ctx->device->pipeline_conv2d_f32[shape];
8034+
} else if (src0->type == GGML_TYPE_F16) {
8035+
return ctx->device->pipeline_conv2d_f16_f32[shape];
8036+
}
8037+
} else if (op == GGML_OP_CONV_TRANSPOSE_2D) {
8038+
if (src0->type == GGML_TYPE_F32) {
8039+
return ctx->device->pipeline_conv_transpose_2d_f32[shape];
8040+
} else if (src0->type == GGML_TYPE_F16) {
8041+
return ctx->device->pipeline_conv_transpose_2d_f16_f32[shape];
8042+
}
79548043
}
79558044
}
79568045
return nullptr;
@@ -8350,6 +8439,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
83508439
{
83518440
elements = ggml_vk_get_conv_elements(dst);
83528441
} break;
8442+
case GGML_OP_CONV_TRANSPOSE_2D:
8443+
{
8444+
elements = ggml_vk_get_conv_transpose_2d_elements(dst);
8445+
} break;
83538446
case GGML_OP_ADD:
83548447
case GGML_OP_SUB:
83558448
case GGML_OP_DIV:
@@ -9523,6 +9616,55 @@ static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx,
95239616
ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D, std::move(p), dryrun);
95249617
}
95259618

9619+
static void ggml_vk_conv_transpose_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
9620+
const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
9621+
GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
9622+
GGML_ASSERT(src1->type == GGML_TYPE_F32);
9623+
GGML_ASSERT(dst->type == GGML_TYPE_F32);
9624+
9625+
GGML_TENSOR_BINARY_OP_LOCALS
9626+
9627+
GGML_ASSERT(nb00 == sizeof(float) || nb00 == sizeof(ggml_fp16_t));
9628+
GGML_ASSERT(nb10 == sizeof(float));
9629+
GGML_ASSERT(nb0 == sizeof(float));
9630+
9631+
vk_op_conv_transpose_2d_push_constants p{};
9632+
p.Cout = static_cast<uint32_t>(ne02);
9633+
p.Cin = static_cast<uint32_t>(ne03);
9634+
p.N = static_cast<uint32_t>(ne13);
9635+
9636+
p.KW = static_cast<uint32_t>(ne00);
9637+
p.KH = static_cast<uint32_t>(ne01);
9638+
p.W = static_cast<uint32_t>(ne10);
9639+
p.H = static_cast<uint32_t>(ne11);
9640+
p.OW = static_cast<uint32_t>(ne0);
9641+
p.OH = static_cast<uint32_t>(ne1);
9642+
9643+
p.s0 = static_cast<uint32_t>(dst->op_params[0]);
9644+
p.s1 = static_cast<uint32_t>(dst->op_params[0]);
9645+
p.p0 = 0;
9646+
p.p1 = 0;
9647+
p.d0 = 1;
9648+
p.d1 = 1;
9649+
9650+
p.nb01 = static_cast<uint32_t>(nb01 / nb00);
9651+
p.nb02 = static_cast<uint32_t>(nb02 / nb00);
9652+
p.nb03 = static_cast<uint32_t>(nb03 / nb00);
9653+
9654+
p.nb11 = static_cast<uint32_t>(nb11 / nb10);
9655+
p.nb12 = static_cast<uint32_t>(nb12 / nb10);
9656+
p.nb13 = static_cast<uint32_t>(nb13 / nb10);
9657+
9658+
p.nb1 = static_cast<uint32_t>(nb1 / nb0);
9659+
p.nb2 = static_cast<uint32_t>(nb2 / nb0);
9660+
p.nb3 = static_cast<uint32_t>(nb3 / nb0);
9661+
9662+
GGML_ASSERT(ne02 == ne2);
9663+
GGML_ASSERT(ne03 == ne12);
9664+
9665+
ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_2D, std::move(p), dryrun);
9666+
}
9667+
95269668
static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
95279669
vk_op_conv2d_dw_push_constants p{};
95289670
p.ne = ggml_nelements(dst);
@@ -10615,6 +10757,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
1061510757
case GGML_OP_CONV_TRANSPOSE_1D:
1061610758
case GGML_OP_POOL_2D:
1061710759
case GGML_OP_CONV_2D:
10760+
case GGML_OP_CONV_TRANSPOSE_2D:
1061810761
case GGML_OP_CONV_2D_DW:
1061910762
case GGML_OP_RWKV_WKV6:
1062010763
case GGML_OP_RWKV_WKV7:
@@ -10686,6 +10829,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
1068610829
case GGML_OP_CONV_TRANSPOSE_1D:
1068710830
case GGML_OP_POOL_2D:
1068810831
case GGML_OP_CONV_2D:
10832+
case GGML_OP_CONV_TRANSPOSE_2D:
1068910833
case GGML_OP_CONV_2D_DW:
1069010834
case GGML_OP_LEAKY_RELU:
1069110835
case GGML_OP_OPT_STEP_SGD:
@@ -10997,6 +11141,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
1099711141
case GGML_OP_CONV_2D:
1099811142
ggml_vk_conv_2d(ctx, compute_ctx, src0, src1, node, dryrun);
1099911143

11144+
break;
11145+
case GGML_OP_CONV_TRANSPOSE_2D:
11146+
ggml_vk_conv_transpose_2d(ctx, compute_ctx, src0, src1, node, dryrun);
11147+
1100011148
break;
1100111149
case GGML_OP_CONV_2D_DW:
1100211150
ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -11137,6 +11285,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
1113711285
case GGML_OP_CONV_TRANSPOSE_1D:
1113811286
case GGML_OP_POOL_2D:
1113911287
case GGML_OP_CONV_2D:
11288+
case GGML_OP_CONV_TRANSPOSE_2D:
1114011289
case GGML_OP_CONV_2D_DW:
1114111290
case GGML_OP_RWKV_WKV6:
1114211291
case GGML_OP_RWKV_WKV7:
@@ -11794,10 +11943,10 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
1179411943
ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
1179511944
if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
1179611945
total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
11797-
} else if (cgraph->nodes[i]->op == GGML_OP_CONV_2D) {
11946+
} else if (cgraph->nodes[i]->op == GGML_OP_CONV_2D || cgraph->nodes[i]->op == GGML_OP_CONV_TRANSPOSE_2D) {
1179811947
// Return CRSxNPQxsizeof(*) to account as many bytes as mul_mat has in im2col->mul_mat mode.
1179911948
auto CRS_size =
11800-
cgraph->nodes[i]->src[0]->ne[0] * cgraph->nodes[i]->src[0]->ne[1] * cgraph->nodes[i]->src[0]->ne[2];
11949+
cgraph->nodes[i]->src[0]->ne[0] * cgraph->nodes[i]->src[0]->ne[1] * cgraph->nodes[i]->src[1]->ne[2];
1180111950
auto NPQ_size = cgraph->nodes[i]->ne[0] * cgraph->nodes[i]->ne[1] * cgraph->nodes[i]->ne[3];
1180211951
total_mat_mul_bytes += NPQ_size * CRS_size * ggml_type_size(cgraph->nodes[i]->type);
1180311952
}
@@ -12618,10 +12767,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
1261812767
case GGML_OP_CONV_TRANSPOSE_1D:
1261912768
return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
1262012769
case GGML_OP_CONV_2D:
12770+
case GGML_OP_CONV_TRANSPOSE_2D:
1262112771
{
1262212772
// Op is disabled for Apple because it segfaults at pipeline create time on MoltenVK
1262312773
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
1262412774
const vk_device& device = ggml_vk_get_device(ctx->device);
12775+
if (op->op == GGML_OP_CONV_TRANSPOSE_2D &&
12776+
device->properties.limits.maxPushConstantsSize < sizeof(vk_op_conv_transpose_2d_push_constants)) {
12777+
return false;
12778+
}
1262512779
// Channel-contiguous format is not supported yet.
1262612780
return ((op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
1262712781
op->src[1]->type == GGML_TYPE_F32 &&
@@ -13240,6 +13394,9 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
1324013394
const int32_t d0 = tensor->op_params[4];
1324113395
const int32_t d1 = tensor->op_params[5];
1324213396
tensor_clone = ggml_conv_2d(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1);
13397+
} else if (tensor->op == GGML_OP_CONV_TRANSPOSE_2D) {
13398+
const int32_t s = tensor->op_params[0];
13399+
tensor_clone = ggml_conv_transpose_2d_p0(ggml_ctx, src_clone[0], src_clone[1], s);
1324313400
} else if (tensor->op == GGML_OP_LEAKY_RELU) {
1324413401
const float * op_params = (const float *)tensor->op_params;
1324513402
tensor_clone = ggml_leaky_relu(ggml_ctx, src_clone[0], op_params[0], false);

0 commit comments

Comments
 (0)