Skip to content

Commit

Permalink
[Snippets][CPU][Tests] Added tests for dynamic BF16/INT8 MHA (openvin…
Browse files Browse the repository at this point in the history 
…otoolkit#27169)

### Details:
 - *Added more tests for the validation of INT8/BF16 MHA in CPU Plugin*
- *Split the large "mha.cpp" file into the several small files with the
same test semantic (comment
openvinotoolkit#26547 (comment)

### Tickets:
 - *N/A*
  • Loading branch information
@a-sidorova
a-sidorova authored Oct 31, 2024
1 parent 86083e0 commit c685d44
Show file tree
Hide file tree
Showing 17 changed files with 807 additions and 620 deletions.
8 changes: 6 additions & 2 deletions src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -237,7 +237,6 @@ std::vector<std::string> disabledTestPatterns() {
R"(.*smoke_FakeQuantize.*/FakeQuantizeLayerTest.Inference.*TS=.*3.4.2.5.*LEVELS=255.*)",
R"(.*smoke_FakeQuantizePerChannel.*/FakeQuantizeLayerTest.Inference.*TS=.*11.10.22.19.*LEVELS=(255|256).*netPRC=f32.*)",
R"(.*smoke_MVN_5D/Mvn6LayerTest.Inference.*TS=.*3.4.2.5.*LEVELS=255.*netPRC=f16.*)",
R"(.*smoke_Snippets_MHAINT8MatMul/MHAINT8MatMul.*)",
R"(.*smoke_static/ConvertFqRnnToQuantizedRnn.*2.1.5.*2.1.1.*2.1.1.*)",
R"(.*smoke_InterpolateBicubicPillow_Layout_Test/InterpolateLayerCPUTest.CompareWithRefs/ShapeCalcMode=sizes_IS=\[?.2..20.?.?\]_TS.*1.17.4.4.*2.3.10.12.*1.17.4.4.*Sizes.*4.4.*10.20.*10.4.*PARAMETER.*0.0.0.0.*0.0.1.1.*2.3.*)",
R"(.*smoke_LoopForCommon/LoopLayerCPUTest.CompareWithRefs/.*_netType=bf16.*)",
Expand Down Expand Up @@ -563,7 +562,7 @@ std::vector<std::string> disabledTestPatterns() {
// ignored for not supported bf16 platforms
retVector.emplace_back(R"(.*smoke_Snippets_EnforcePrecision_bf16.*)");
retVector.emplace_back(R"(.*smoke_Snippets_MHAWOTransposeEnforceBF16.*)");
retVector.emplace_back(R"(.*smoke_Snippets_MHAEnforceBF16.*)");
retVector.emplace_back(R"(.*smoke_Snippets_MHA.*EnforceBF16.*)");
retVector.emplace_back(R"(.*ConcatSDPTest.*bf16.*)");
}
// [150842] Need to support dynamic K dimension of BF16|INT8 MatMul on AMX systems
Expand All @@ -572,6 +571,11 @@ std::vector<std::string> disabledTestPatterns() {
retVector.emplace_back(R"(.*smoke_Snippets_MatMul/MatMul.CompareWithRefImpl/.*IS\[0\]=\[\?.\?.\?.\?\].*T\[0\]=(u8|i8|bf16)_T\[1\]=(i8|bf16).*)");
retVector.emplace_back(R"(.*smoke_Snippets_MatMulTransposeB.*IS\[0\]=\[\?.\?.\?.\?\].*T\[0\]=(u8|i8|bf16)_T\[1\]=(i8|bf16).*)");
retVector.emplace_back(R"(.*smoke_Snippets_MatMulBias.*IS\[0\]=\[\?.\?.\?.\?\].*T\[0\]=(u8|i8|bf16)_T\[1\]=(i8|bf16).*)");

retVector.emplace_back(R"(.*smoke_Snippets_MHAWOTransposeEnforceBF16_3D.*IS\[1\]=\[2.64.\?\].*)");
retVector.emplace_back(R"(.*smoke_Snippets_MHA.*BF16.*/MHA.*IS\[0\]=\[(\?|1).(\?|4).(\?|12).(\?|64)\].*)");
retVector.emplace_back(R"(.*smoke_Snippets_MHA.*BF16.*/MHA.*IS\[0\]=\[\?.\?.\?\].*)");
retVector.emplace_back(R"(.*smoke_Snippets_(MHAINT8MatMul|MHAQuantMatMul0|MHAFQAfterMatMul_4D|smoke_Snippets_MHAFQ).*IS\[0\]=\[\?.\?.\?\.\?].*)");
}
#ifdef SNIPPETS_LIBXSMM_TPP
// GN in TPP requires exposing tmp Buffer results outside the loop (ticket: 151234)
Expand Down
38 changes: 10 additions & 28 deletions src/plugins/intel_cpu/tests/functional/shared_tests_instances/snippets/matmul.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,44 +4,26 @@

#include "snippets/matmul.hpp"

#include "common_test_utils/test_constants.hpp"
#include "openvino/runtime/system_conf.hpp"
#include "utils.hpp"

namespace ov {
namespace test {
namespace snippets {

#define STATIC_SHAPES(...) static_shapes_to_test_representation(std::vector<std::vector<ov::Shape>>{__VA_ARGS__})

namespace {

static inline std::vector<std::vector<element::Type>> quantized_precisions() {
std::vector<std::vector<element::Type>> prc = {};
// In Snippets MatMul INT8 is supported only on VNNI/AMX platforms
if (ov::with_cpu_x86_avx512_core_vnni() || ov::with_cpu_x86_avx512_core_amx_int8()) {
prc.emplace_back(std::vector<element::Type>{element::i8, element::i8});
prc.emplace_back(std::vector<element::Type>{element::u8, element::i8});
}
return prc;
}

static inline std::vector<std::vector<element::Type>> precisions() {
std::vector<std::vector<element::Type>> prc = {
{element::f32, element::f32},
};
std::vector<std::vector<element::Type>> prc = precision_f32(2);
// Note: TPP doesn't support low precisions yet
#ifndef SNIPPETS_LIBXSMM_TPP
auto quant = quantized_precisions();
auto quant = quantized_precisions_if_supported();
std::copy(quant.begin(), quant.end(), std::back_inserter(prc));
// In Snippets MatMul BF16 is supported only on bf16/AMX platforms
if (ov::with_cpu_x86_bfloat16() || ov::with_cpu_x86_avx512_core_amx_bf16()) {
prc.emplace_back(std::vector<element::Type>{element::bf16, element::bf16});
}
auto bfloat = precision_bf16_if_supported(2);
std::copy(bfloat.begin(), bfloat.end(), std::back_inserter(prc));
#endif
return prc;
}


std::vector<std::vector<ov::test::InputShape>> input_shapes{
{ {{}, {{2, 1, 3, 5}}}, {{}, {{1, 3, 5, 3}}} },
{ {{}, {{3, 1, 32, 14}}}, {{}, {{1, 3, 14, 37}}} },
Expand Down Expand Up @@ -158,7 +140,7 @@ INSTANTIATE_TEST_SUITE_P(smoke_Snippets_MatMulBias, MatMulBias,
INSTANTIATE_TEST_SUITE_P(smoke_Snippets_MatMulBiasQuantized, MatMulBiasQuantized,
::testing::Combine(
::testing::ValuesIn(input_shapes_bias),
::testing::ValuesIn(quantized_precisions()),
::testing::ValuesIn(quantized_precisions_if_supported()),
::testing::Values(MatMulType::MatMul),
::testing::Values(1), // Subgraph
::testing::Values(1), // Tokenized MatMul+Bias
Expand All @@ -167,8 +149,8 @@ INSTANTIATE_TEST_SUITE_P(smoke_Snippets_MatMulBiasQuantized, MatMulBiasQuantized

INSTANTIATE_TEST_SUITE_P(smoke_Snippets_MatMulsQuantized, MatMulsQuantized,
::testing::Combine(
::testing::ValuesIn(STATIC_SHAPES({{1, 16, 128, 64}, {1, 16, 64, 128}, {128, 64}})),
::testing::ValuesIn(quantized_precisions()),
::testing::ValuesIn(SNIPPETS_TESTS_STATIC_SHAPES({{1, 16, 128, 64}, {1, 16, 64, 128}, {128, 64}})),
::testing::ValuesIn(quantized_precisions_if_supported()),
::testing::Values(MatMulType::MatMul),
::testing::Values(3), // Subgraph + Reshape + Subgraph
::testing::Values(2), // Tokenized [MatMul+FQ+Matmul] and [FQ]
Expand All @@ -177,8 +159,8 @@ INSTANTIATE_TEST_SUITE_P(smoke_Snippets_MatMulsQuantized, MatMulsQuantized,

INSTANTIATE_TEST_SUITE_P(smoke_Snippets_MatMulsQuantizedSoftmax, MatMulsQuantizedSoftmax,
::testing::Combine(
::testing::ValuesIn(STATIC_SHAPES({{1, 16, 128, 64}, {1, 16, 64, 128}, {128, 64}})),
::testing::ValuesIn(quantized_precisions()),
::testing::ValuesIn(SNIPPETS_TESTS_STATIC_SHAPES({{1, 16, 128, 64}, {1, 16, 64, 128}, {128, 64}})),
::testing::ValuesIn(quantized_precisions_if_supported()),
::testing::Values(MatMulType::MatMul),
::testing::Values(3), // Subgraph + Reshape + Subgraph
::testing::Values(2), // Tokenized [MatMul+FQ+Matmul] and [FQ]
Expand Down
Loading

0 comments on commit c685d44

Please sign in to comment.