vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 17:44:38 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

ggml: add conv3d op (#15182)

Browse source 
* add conv3d

* bump GGML_OP_COUNT
This commit is contained in:
rmatif 2025-08-22 15:33:15 +02:00 • committed by GitHub
parent b1ab91821f
commit 92f7f0a53c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 345 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

124
tests/test-backend-ops.cpp
View file

@ -4091,6 +4091,75 @@ struct test_conv_2d_dw : public test_case {
}
};
// GGML_OP_CONV_3D
struct test_conv_3d : public test_case {
// Logical 5D dimensions
const int64_t N, IC, ID, IH, IW;
const int64_t OC, KD, KH, KW;
// Conv params
const int s0, s1, s2;
const int p0, p1, p2;
const int d0, d1, d2;
// Types
const ggml_type type_kernel;
std::string op_desc(ggml_tensor * t) override {
GGML_UNUSED(t);
return "CONV_3D";
}
std::string vars() override {
return VARS_TO_STR11(N, IC, ID, IH, IW, OC, KD, KH, KW, s0, s1) + "," +
VARS_TO_STR8(s2, p0, p1, p2, d0, d1, d2, type_kernel);
}
double max_nmse_err() override {
return 5e-4;
}
uint64_t op_flops(ggml_tensor * t) override {
GGML_UNUSED(t);
auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
};
const int64_t OD = calc_conv_output_size(ID, KD, s2, p2, d2);
const int64_t OH = calc_conv_output_size(IH, KH, s1, p1, d1);
const int64_t OW = calc_conv_output_size(IW, KW, s0, p0, d0);
return (uint64_t)N * OC * OD * OH * OW * (2 * IC * KD * KH * KW - 1);
}
test_conv_3d(
int64_t N, int64_t IC, int64_t ID, int64_t IH, int64_t IW,
int64_t OC, int64_t KD, int64_t KH, int64_t KW,
int s0, int s1, int s2,
int p0, int p1, int p2,
int d0, int d1, int d2,
ggml_type type_kernel
) : N(N), IC(IC), ID(ID), IH(IH), IW(IW),
OC(OC), KD(KD), KH(KH), KW(KW),
s0(s0), s1(s1), s2(s2),
p0(p0), p1(p1), p2(p2),
d0(d0), d1(d1), d2(d2),
type_kernel(type_kernel) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
// GGML input tensor is packed as [W, H, D, C*N]
const int64_t ne_input[] = {IW, IH, ID, IC * N};
ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input);
ggml_set_name(input, "input");
// GGML kernel tensor is packed as [KW, KH, KD, IC*OC]
const int64_t ne_kernel[] = {KW, KH, KD, IC * OC};
ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel);
ggml_set_name(kernel, "kernel");
ggml_tensor * out = ggml_conv_3d(ctx, kernel, input, s0, s1, s2, p0, p1, p2, d0, d1, d2, (int)IC, (int)N, (int)OC);
ggml_set_name(out, "out");
return out;
}
};
// GGML_OP_CONCAT
struct test_concat : public test_case {
const ggml_type type;
@ -5528,6 +5597,61 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, false));
test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, true));
// CONV_3D
auto calc_conv_output_size_3d = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
};
for (ggml_type kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
for (int N : {1, 2}) {
for (int IC : {1, 3}) {
for (int OC : {1, 4}) {
for (int s0 : {1, 2}) {
for (int p1 : {0, 1}) {
for (int d2 : {1, 2}) {
int64_t IW = 20, IH = 22, ID = 18;
int64_t KW = 3, KH = 3, KD = 3;
int s1 = s0, s2 = s0;
int p0 = p1, p2 = p1;
int d0 = d2, d1 = d2;
if (calc_conv_output_size_3d(IW, KW, s0, p0, d0) <= 0 ||
calc_conv_output_size_3d(IH, KH, s1, p1, d1) <= 0 ||
calc_conv_output_size_3d(ID, KD, s2, p2, d2) <= 0) {
continue;
}
test_cases.emplace_back(new test_conv_3d(
N, IC, ID, IH, IW,
OC, KD, KH, KW,
s0, s1, s2, p0, p1, p2, d0, d1, d2,
kernel_type));
// Asymmetric kernel and params
int64_t asym_KW = 5, asym_KH = 1, asym_KD = 3;
int asym_s0 = 2, asym_s1 = 1, asym_s2 = 1;
int asym_p0 = 2, asym_p1 = 0, asym_p2 = 1;
int asym_d0 = 1, asym_d1 = 1, asym_d2 = 2;
if (calc_conv_output_size_3d(IW, asym_KW, asym_s0, asym_p0, asym_d0) <= 0 ||
calc_conv_output_size_3d(IH, asym_KH, asym_s1, asym_p1, asym_d1) <= 0 ||
calc_conv_output_size_3d(ID, asym_KD, asym_s2, asym_p2, asym_d2) <= 0) {
continue;
}
test_cases.emplace_back(new test_conv_3d(
N, IC, ID, IH, IW,
OC, asym_KD, asym_KH, asym_KW,
asym_s0, asym_s1, asym_s2, asym_p0, asym_p1, asym_p2, asym_d0, asym_d1, asym_d2,
kernel_type));
}
}
}
}
}
}
// Case with kernel size 1
test_cases.emplace_back(new test_conv_3d(1, 4, 8, 8, 8, 8, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, kernel_type));
}
for(uint32_t Cout : {1, 9}){
for(uint32_t Cin : {1, 7}){
for(uint32_t K : {1, 3, 1337}){