vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

vulkan: Optimize contiguous copies (#10254)

Browse source 
* tests: Fix memory bandwidth calculation for perf tests

Add a flops calculation for flash attention.

Add one GGML_OP_CPY perf test.

* vulkan: Optimize contiguous copies

Add a variant of the copy shader for when the tensors are contiguous. Avoid
the complex addressing calculations, and do four elements per invocation
to hide some other overhead.

Apply similar changes to the scale shader, since scale is always contiguous.

Add a "progress bar" for shader compiles.
This commit is contained in:
Jeff Bolz 2024-11-13 00:58:57 -06:00 • committed by GitHub
parent 54ef9cfc72
commit 80dd7ff22f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
13 changed files with 144 additions and 27 deletions
Download patch file Download diff file Expand all files Collapse all files

13
tests/test-backend-ops.cpp
View file

@ -681,6 +681,7 @@ struct test_case {
// run
int64_t total_time_us = 0;
int64_t total_mem = 0;
int total_runs = 0;
do {
int64_t start_time = ggml_time_us();
@ -688,6 +689,7 @@ struct test_case {
int64_t end_time = ggml_time_us();
total_time_us += end_time - start_time;
total_mem += mem;
total_runs += n_runs;
} while (total_time_us < 1000*1000); // run for at least 1 second
@ -717,7 +719,7 @@ struct test_case {
} else {
printf("%8zu kB/run - \033[1;34m%7.2f GB/s\033[0m",
op_size(out) / 1024,
mem / (total_time_us / 1e6) / 1024.0 / 1024.0 / 1024.0);
total_mem / (total_time_us / 1e6) / 1024.0 / 1024.0 / 1024.0);
}
printf("\n");
@ -2740,6 +2742,13 @@ struct test_flash_attn_ext : public test_case {
return 5e-4;
}
uint64_t op_flops(ggml_tensor * t) override {
GGML_UNUSED(t);
// Just counting matmul costs:
// Q*K^T is nb x hs x kv, P*V is nb x kv x hs, per head
return 2 * 2 * nh * nb * hs * kv;
}
test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8,
bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16)
: hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV) {}
@ -3779,6 +3788,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 1, 1, 1}));
test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 512, 1, 1}));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F16, {512, 3072, 1, 1}));
for (int bs : {1, 512}) {
for (ggml_type type_a : all_types) {
for (ggml_type type_b : {GGML_TYPE_F32}) {