Skip to content

Gemv fp16 fix #9331

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Sep 9, 2022
Merged

Gemv fp16 fix #9331

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
29f6446
fix fp16 gemv compute error
lxwlaq Jul 5, 2022
5c0c4b4
fix
lxwlaq Jul 6, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions lite/backends/arm/math/fp16/conv_impl_fp16.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -338,8 +338,8 @@ void conv1x1s1_gemm_fp16(CONV_PARAM(float16_t)) {
}
}

gemv_fp16(weights_group,
din_group,
gemv_fp16(din_group,
weights_group,
dout_group,
true,
n,
Expand Down Expand Up @@ -453,8 +453,8 @@ void conv_im2col_gemm_fp16(CONV_PARAM(float16_t)) {
}
}

gemv_fp16(weights_group,
dB,
gemv_fp16(dB,
weights_group,
dout_group,
true,
n,
Expand Down
10 changes: 3 additions & 7 deletions lite/backends/arm/math/fp16/gemv_fp16.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -277,7 +277,7 @@ void gemv_fp16_trans(const float16_t *A,
ARMContext *ctx) {
int Nup = (N + 7) / 8 * 8;
int Mup = (M + 7) / 8 * 8;
auto size = (Mup * 3 + Nup);
auto size = (Mup * 2 + Nup);
ctx->ExtendWorkspace(size * sizeof(float16_t));
auto ptr_zero = ctx->workspace_data<float16_t>();
memset(ptr_zero, 0, Mup * sizeof(float16_t));
Expand All @@ -289,12 +289,8 @@ void gemv_fp16_trans(const float16_t *A,
memset(bias_ptr, 0, Mup * sizeof(float16_t));
}
float16_t *ptr_w = bias_ptr + Mup;
lite::TargetWrapperHost::MemcpySync(ptr_w, A, N * sizeof(float16_t));
lite::TargetWrapperHost::MemcpySync(ptr_w, x, N * sizeof(float16_t));
memset(ptr_w + N, 0, (Nup - N) * sizeof(float16_t));
float16_t *data_in = ptr_w + Nup;
lite::TargetWrapperHost::MemcpySync(
data_in, x + (N - 1) * M, M * sizeof(float16_t));
memset(data_in + M, 0, (Mup - M) * sizeof(float16_t));
memset(y, 0, M * sizeof(float16_t));
float16_t local_alpha = 0.f;
float16_t offset = 0.f;
Expand All @@ -317,7 +313,7 @@ void gemv_fp16_trans(const float16_t *A,
int y_index = j * 8;
const float16_t *ptr_in = ptr_w + y_index;
const float16_t *inptr_row[8];
inptr_row[0] = x + y_index * M;
inptr_row[0] = A + y_index * M;
for (int i = 1; i < 8; i++) {
inptr_row[i] = inptr_row[i - 1] + M;
}
Expand Down
4 changes: 2 additions & 2 deletions lite/tests/math/gemv_fp16_compute_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,7 +136,7 @@ bool test_sgemv_fp16(bool tra,
dc_basic,
static_cast<float16_t>(1.f),
static_cast<float16_t>(0.f),
false,
tra,
has_bias,
flag_act,
alpha);
Expand Down Expand Up @@ -211,7 +211,7 @@ TEST(TestLiteGemvFP16, gemv_fp16) {
LOG(INFO) << "run basic sgemm test";
for (auto& m : {3, 8, 32, 397}) {
for (auto& n : {3, 13, 141, 512, 789}) {
for (auto& tra : {false}) {
for (auto& tra : {false, true}) {
for (auto& has_bias : {false, true}) {
for (auto& flag_act : {0, 1}) {
for (auto& th : {1, 2, 4}) {
Expand Down