Skip to content

Commit

Permalink
refactor: Improving the efficiency of the AVX-implementations (#866)
Browse files Browse the repository at this point in the history
  • Loading branch information
@RealTimeChris
RealTimeChris authored Sep 20, 2023
1 parent 2e75d3d commit e0e4603
Show file tree
Hide file tree
Showing 4 changed files with 138 additions and 196 deletions.
110 changes: 46 additions & 64 deletions include/dpp/isa/avx.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,80 +23,22 @@
#if defined _MSC_VER || defined __GNUC__ || defined __clang__

#include <immintrin.h>

#ifdef max
#undef max
#endif
#ifdef min
#undef min
#endif
#include <numeric>

namespace dpp {

using avx_float = __m128;
using avx_int = __m128i;

/*
* @brief Extracts a 32-bit integer from a 128-bit AVX register.
* @param value The AVX register containing packed 32-bit integers.
* @param index The index of the 32-bit integer to extract (0-3).
* @return The extracted 32-bit integer.
*/
inline int32_t extract_int32_from_avx(const avx_int& value, int64_t index) {
switch (index) {
case 0: {
return _mm_extract_epi32(value, 0);
}
case 1: {
return _mm_extract_epi32(value, 1);
}
case 2: {
return _mm_extract_epi32(value, 2);
}
case 3: {
return _mm_extract_epi32(value, 3);
}
default: {
return _mm_extract_epi32(value, 0);
}
}
}


/**
* @brief A class for audio mixing operations using AVX instructions.
*/
class audio_mixer {
public:
/*
* @brief The number of 32-bit values per CPU register.
*/
inline static constexpr int32_t byte_blocks_per_register{ 4 };

/*
* @brief Stores values from a 128-bit AVX vector to a storage location.
* @tparam value_type The target value type for storage.
* @param values_to_store The 128-bit AVX vector containing values to store.
* @param storage_location Pointer to the storage location.
*/
template<typename value_type> inline static void store_values(const avx_int& values_to_store, value_type* storage_location) {
for (int64_t x = 0; x < byte_blocks_per_register; ++x) {
storage_location[x] = static_cast<value_type>(extract_int32_from_avx(values_to_store, x));
}
}

/**
* @brief Specialization for gathering non-float values into an AVX register.
* @tparam value_type The type of values being gathered.
* @tparam Indices Parameter pack of indices for gathering values.
* @return An AVX register containing gathered values.
* @brief The number of 32-bit values per CPU register.
*/
template<typename value_type> inline static avx_float gather_values(value_type* values) {
alignas(16) float new_array[byte_blocks_per_register]{};
for (size_t x = 0; x < byte_blocks_per_register; ++x) {
new_array[x] = static_cast<float>(values[x]);
}
return _mm_load_ps(new_array);
}
inline static constexpr int32_t byte_blocks_per_register{ 4 };

/**
* @brief Collect a single register worth of data from data_in, apply gain and increment, and store the result in data_out.
Expand All @@ -115,7 +57,7 @@ namespace dpp {
_mm_min_ps(current_samples_new, _mm_set1_ps(static_cast<float>(std::numeric_limits<int16_t>::max()))),
_mm_cmp_ps(current_samples_new, _mm_set1_ps(0.0f), _CMP_GE_OQ));

store_values(_mm_cvtps_epi32(current_samples_new), data_out);
store_values(current_samples_new, data_out);
}

/**
Expand All @@ -126,9 +68,49 @@ namespace dpp {
* @param decoded_data Pointer to the array of int16_t values.
*/
inline static void combine_samples(int32_t* up_sampled_vector, const int16_t* decoded_data) {
auto newValues{ _mm_cvtps_epi32(_mm_add_ps(gather_values(up_sampled_vector), gather_values(decoded_data))) };
auto newValues{ _mm_add_ps(gather_values(up_sampled_vector), gather_values(decoded_data)) };
store_values(newValues, up_sampled_vector);
}

protected:

/**
* @brief Stores values from a 128-bit AVX vector to a storage location.
* @tparam value_type The target value type for storage.
* @param values_to_store The 128-bit AVX vector containing values to store.
* @param storage_location Pointer to the storage location.
*/
template<typename value_type> inline static void store_values(const avx_float& values_to_store, value_type* storage_location) {
for (int64_t x = 0; x < byte_blocks_per_register; ++x) {
storage_location[x] = static_cast<value_type>(extract_float_from_avx(values_to_store, x));
}
}

/**
* @brief Specialization for gathering non-float values into an AVX register.
* @tparam value_type The type of values being gathered.
* @tparam Indices Parameter pack of indices for gathering values.
* @return An AVX register containing gathered values.
*/
template<typename value_type> inline static avx_float gather_values(value_type* values) {
alignas(16) float new_array[byte_blocks_per_register]{};
for (uint64_t x = 0; x < byte_blocks_per_register; ++x) {
new_array[x] = static_cast<float>(values[x]);
}
return _mm256_load_ps(new_array);
}

/**
* @brief Extracts a 32-bit integer from a 128-bit AVX register.
* @param value The AVX register containing packed 32-bit integers.
* @param index The index of the 32-bit integer to extract (0-3).
* @return The extracted 32-bit integer.
*/
inline static float extract_float_from_avx(const avx_float& value, int64_t index) {
alignas(16) float new_array[4]{};
_mm_store_ps(new_array, value);
return new_array[index];
}
};

} // namespace dpp
Expand Down
120 changes: 45 additions & 75 deletions include/dpp/isa/avx2.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,92 +23,22 @@
#if defined _MSC_VER || defined __GNUC__ || defined __clang__

#include <immintrin.h>

#ifdef max
#undef max
#endif
#ifdef min
#undef min
#endif
#include <numeric>

namespace dpp {

using avx_2_float = __m256;
using avx_2_int = __m256i;

/*
* @brief Extracts a 32-bit integer from a 256-bit AVX2 register.
* @param value The AVX2 register containing packed 32-bit integers.
* @param index The index of the 32bit integer to extract (0-7).
* @return The extracted 32-bit integer.
*/
inline int32_t extract_int32_from_avx2(const avx_2_int& value, int64_t index) {
switch (index) {
case 0: {
return _mm256_extract_epi32(value, 0);
}
case 1: {
return _mm256_extract_epi32(value, 1);
}
case 2: {
return _mm256_extract_epi32(value, 2);
}
case 3: {
return _mm256_extract_epi32(value, 3);
}
case 4: {
return _mm256_extract_epi32(value, 4);
}
case 5: {
return _mm256_extract_epi32(value, 5);
}
case 6: {
return _mm256_extract_epi32(value, 6);
}
case 7: {
return _mm256_extract_epi32(value, 7);
}
default: {
return _mm256_extract_epi32(value, 0);
}
}
}

/**
* @brief A class for audio mixing operations using AVX2 instructions.
*/
class audio_mixer {
public:
/*
* @brief The number of 32-bit values per CPU register.
*/
inline static constexpr int32_t byte_blocks_per_register{ 8 };

/*
* @brief Stores values from a 256-bit AVX2 vector to a storage location.
* @tparam value_type The target value type for storage.
* @param values_to_store The 256-bit AVX2 vector containing values to store.
* @param storage_location Pointer to the storage location.
*/
template<typename value_type> inline static void store_values(const avx_2_int& values_to_store, value_type* storage_location) {
for (int64_t x = 0; x < byte_blocks_per_register; ++x) {
storage_location[x] = static_cast<value_type>(extract_int32_from_avx2(values_to_store, x));
}
}

/**
* @brief Specialization for gathering non-float values into an AVX2 register.
* @tparam value_type The type of values being gathered.
* @tparam Indices Parameter pack of indices for gathering values.
* @return An AVX2 register containing gathered values.
* @brief The number of 32-bit values per CPU register.
*/
template<typename value_type> inline static avx_2_float gather_values(value_type* values) {
alignas(32) float new_array[byte_blocks_per_register]{};
for (size_t x = 0; x < byte_blocks_per_register; ++x) {
new_array[x] = static_cast<float>(values[x]);
}
return _mm256_load_ps(new_array);
}
inline static constexpr int32_t byte_blocks_per_register{ 8 };

/**
* @brief Collect a single register worth of data from data_in, apply gain and increment, and store the result in data_out.
Expand All @@ -129,7 +59,7 @@ namespace dpp {
_mm256_min_ps(current_samples_new, _mm256_set1_ps(static_cast<float>(std::numeric_limits<int16_t>::max()))),
_mm256_cmp_ps(current_samples_new, _mm256_set1_ps(0.0f), _CMP_GE_OQ));

store_values(_mm256_cvtps_epi32(current_samples_new), data_out);
store_values(current_samples_new, data_out);
}

/**
Expand All @@ -141,9 +71,49 @@ namespace dpp {
* @param x Index to select a specific set of elements to combine.
*/
inline static void combine_samples(int32_t* up_sampled_vector, const int16_t* decoded_data) {
auto newValues{ _mm256_cvtps_epi32(_mm256_add_ps(gather_values(up_sampled_vector), gather_values(decoded_data))) };
auto newValues{ _mm256_add_ps(gather_values(up_sampled_vector), gather_values(decoded_data)) };
store_values(newValues, up_sampled_vector);
}

protected:

/**
* @brief Stores values from a 256-bit AVX2 vector to a storage location.
* @tparam value_type The target value type for storage.
* @param values_to_store The 256-bit AVX2 vector containing values to store.
* @param storage_location Pointer to the storage location.
*/
template<typename value_type> inline static void store_values(const avx_2_float& values_to_store, value_type* storage_location) {
for (int64_t x = 0; x < byte_blocks_per_register; ++x) {
storage_location[x] = static_cast<value_type>(extract_float_from_avx_2(values_to_store, x));
}
}

/**
* @brief Specialization for gathering non-float values into an AVX2 register.
* @tparam value_type The type of values being gathered.
* @tparam Indices Parameter pack of indices for gathering values.
* @return An AVX2 register containing gathered values.
*/
template<typename value_type> inline static avx_2_float gather_values(value_type* values) {
alignas(32) float new_array[byte_blocks_per_register]{};
for (uint64_t x = 0; x < byte_blocks_per_register; ++x) {
new_array[x] = static_cast<float>(values[x]);
}
return _mm256_load_ps(new_array);
}

/**
* @brief Extracts a 32-bit integer from a 256-bit AVX2 register.
* @param value The AVX2 register containing packed 32-bit integers.
* @param index The index of the 32-bit integer to extract (0-7).
* @return The extracted 32-bit integer.
*/
inline static float extract_float_from_avx_2(const avx_2_float& value, int64_t index) {
alignas(32) float new_array[byte_blocks_per_register]{};
_mm256_store_ps(new_array, value);
return new_array[index];
}
};

} // namespace dpp
Expand Down
Loading

0 comments on commit e0e4603

Please sign in to comment.