[Snippets] SplitM pass removed

Author: v-Golubev

src/common/snippets/include/snippets/lowered/pass/mha_parallel_wa_optimizer.hpp

@@ -40,7 +40,41 @@ class MHAParallelWAOptimizer : public lowered::pass::RuntimeOptimizer {
         return !m_loops_to_split.empty();
     }
 
+    /**
+     * @brief Tries to split M dimension in "shape" in accordance to optimal parallel work amount
+     * @param shape Original shape
+     * @param optimal_parallelism_work_amount Optimal work amount
+     * @param batch_m_dim reference on batch's part of the split M
+     * @param new_m_dim reference on new M dim after the split
+     * @return true if split was successfull, otherwise false
+     */
+    static bool split(const ov::Shape& shape,
+                     size_t optimal_parallelism_work_amount,
+                     size_t& batch_m_dim,
+                     size_t& new_m_dim);
+
 private:
+    /**
+     * @brief Contains splitM approaches allowing to get the batch ideally divisible by
+     * optimal_parallelism_work_amount
+     */
+    static std::pair<size_t, size_t> split_ideally(size_t batch_dim,
+                                                   size_t m_dim,
+                                                   size_t optimal_parallelism_work_amount);
+    /**
+     * @brief Splits m_dim to minimize kernel_m in order to reduce waiting time for idle threads at the last
+     * parallel loop iteration.
+     */
+    static std::pair<size_t, size_t> split_minimize_kernel_wa(size_t batch_dim,
+                                                              size_t m_dim,
+                                                              size_t optimal_parallelism_work_amount);
+    /**
+     * @brief Splits m_dim to get the batch in (optimal_parallelism_work_amount, 2 *
+     * optimal_parallelism_work_amount) interval
+     */
+    static std::pair<size_t, size_t> split_fallback_increase_parallel_wa(size_t batch_dim,
+                                                                         size_t m_dim,
+                                                                         size_t optimal_parallelism_work_amount);
     static std::unordered_set<lowered::ExpressionPtr> find_applicable_brgemms(const lowered::LinearIRCPtr& linear_ir,
                                                                               bool check_dynamic_wa = true);
 
@@ -58,6 +92,7 @@ class MHAParallelWAOptimizer : public lowered::pass::RuntimeOptimizer {
     size_t m_concurrency = 0;
 
     static const size_t m_dim_M_idx;
+    static const size_t m_min_kernel_m;
 };
 
 }  // namespace ov::snippets::lowered::pass

src/common/snippets/src/lowered/pass/mha_parallel_wa_optimizer.cpp

@@ -5,8 +5,10 @@
 #include "snippets/lowered/pass/mha_parallel_wa_optimizer.hpp"
 
 #include <algorithm>
+#include <cmath>
 #include <cstddef>
 #include <iterator>
+#include <numeric>
 #include <set>
 #include <unordered_set>
 #include <vector>
@@ -22,14 +24,60 @@
 #include "snippets/lowered/loop_port.hpp"
 #include "snippets/lowered/pass/runtime_optimizer.hpp"
 #include "snippets/op/brgemm.hpp"
-#include "snippets/pass/split_dimension_m.hpp"
 #include "snippets/runtime_configurator.hpp"
 #include "snippets/utils/loop_utils.hpp"
 #include "snippets/utils/utils.hpp"
 
 namespace ov::snippets::lowered::pass {
+namespace {
+std::vector<size_t> get_updated_order(const std::vector<size_t>& order, size_t m_index) {
+    std::vector<size_t> new_order(order.size() + 1, 0);
+    size_t shift_idx = 0;
+    for (size_t i = 0; i < order.size(); ++i) {
+        if (order[i] < m_index) {
+            new_order[i + shift_idx] = order[i];
+        } else if (order[i] == m_index) {
+            new_order[i + shift_idx++] = order[i];
+            new_order[i + shift_idx] = order[i] + 1;
+        } else {
+            new_order[i + shift_idx] = order[i] + 1;
+        }
+    }
+    return new_order;
+}
+
+ov::snippets::VectorDims unsqueeze_m_dim(ov::snippets::VectorDims shape, size_t m_index) {
+    shape.insert(shape.begin() + m_index, 1);
+    return shape;
+}
+
+ov::snippets::VectorDims reshape_m_dim(ov::snippets::VectorDims shape,
+                                       size_t m_index,
+                                       size_t batch_m_dim,
+                                       size_t new_m_dim) {
+    if (shape[m_index] == 1)
+        return unsqueeze_m_dim(std::move(shape), m_index);
+    shape[m_index] = new_m_dim;
+    shape.insert(shape.begin() + m_index, batch_m_dim);
+    return shape;
+}
+
+bool is_prime_number(size_t value) {
+    if (value == 2lu || value == 3lu)
+        return true;
+    if (value == 1 || value % 2 == 0 || value % 3 == 0)
+        return false;
+    const auto root = std::sqrt(value) + 1;
+    for (size_t divisor = 5; divisor < root; divisor += 6) {
+        if ((value % divisor == 0) || (value % (divisor + 2) == 0))
+            return false;
+    }
+    return true;
+}
+}  // namespace
 
 const size_t MHAParallelWAOptimizer::m_dim_M_idx = 1;
+const size_t MHAParallelWAOptimizer::m_min_kernel_m = 32;
 
 MHAParallelWAOptimizer::MHAParallelWAOptimizer(const lowered::LinearIRCPtr& linear_ir,
                                                const RuntimeConfigurator* configurator)
@@ -56,18 +104,15 @@ MHAParallelWAOptimizer::MHAParallelWAOptimizer(const lowered::LinearIRCPtr& line
                                                             : utils::get_output_dim_idx(layout, m_dim_M_idx);
         m_dim_M_idces[i] = dim_idx;
         const auto m_idx = i < configurator->get_in_num() ? dim_idx : layout.size() - 2;
-        m_optimized_layouts[i] = ov::snippets::pass::SplitDimensionM::get_updated_order(layout, m_idx);
+        m_optimized_layouts[i] = get_updated_order(layout, m_idx);
     }
 }
 
 bool MHAParallelWAOptimizer::run(const lowered::LinearIR& linear_ir) {
     OV_ITT_SCOPED_TASK(ov::pass::itt::domains::SnippetsTransform, "Snippets::MHAParallelWAOptimizer")
     const auto& config = m_configurator->get_config();
     size_t new_batch_dim = 0, new_kernel_dim = 0;
-    if (!ov::snippets::pass::SplitDimensionM::split(config->master_shape,
-                                                    m_concurrency,
-                                                    new_batch_dim,
-                                                    new_kernel_dim)) {
+    if (!split(config->master_shape, m_concurrency, new_batch_dim, new_kernel_dim)) {
         return false;
     }
     auto& master_shape = config->master_shape;
@@ -98,13 +143,9 @@ bool MHAParallelWAOptimizer::run(const lowered::LinearIR& linear_ir) {
     }
 
     for (size_t i = 0; i < m_configurator->get_io_num(); ++i) {
-        config->io_shapes[i] =
-            m_unsqueezed_params.count(i)
-                ? ov::snippets::pass::SplitDimensionM::unsqueeze_m_dim(config->io_shapes[i], m_dim_M_idces[i])
-                : ov::snippets::pass::SplitDimensionM::reshape_m_dim(config->io_shapes[i],
-                                                                     m_dim_M_idces[i],
-                                                                     new_batch_dim,
-                                                                     new_kernel_dim);
+        config->io_shapes[i] = m_unsqueezed_params.count(i)
+                                   ? unsqueeze_m_dim(config->io_shapes[i], m_dim_M_idces[i])
+                                   : reshape_m_dim(config->io_shapes[i], m_dim_M_idces[i], new_batch_dim, new_kernel_dim);
     }
     config->io_layouts = m_optimized_layouts;
     return true;
@@ -211,4 +252,99 @@ std::vector<lowered::ExpandedLoopInfoPtr> MHAParallelWAOptimizer::find_loops_to_
     return loops_to_split;
 }
 
+bool MHAParallelWAOptimizer::split(const ov::Shape& shape,
+                                   size_t optimal_parallelism_work_amount,
+                                   size_t& batch_m_dim,
+                                   size_t& new_m_dim) {
+    const auto batch_dim = std::accumulate(shape.rbegin() + 2, shape.rend(), size_t(1), std::multiplies<size_t>());
+    const auto m_dim = *(shape.rbegin() + 1);
+    if (is_prime_number(m_dim))
+        return false;
+
+    // We skip optimization if the current batch is optimal for concurrency
+    if (batch_dim % optimal_parallelism_work_amount == 0)
+        return false;
+
+    auto split_is_done = [&batch_m_dim]() {
+        return batch_m_dim != 1;
+    };
+
+    std::tie(batch_m_dim, new_m_dim) = split_ideally(batch_dim, m_dim, optimal_parallelism_work_amount);
+    if (split_is_done())
+        return true;
+
+    std::tie(batch_m_dim, new_m_dim) = split_minimize_kernel_wa(batch_dim, m_dim, optimal_parallelism_work_amount);
+    if (split_is_done())
+        return true;
+    // If all the previous heuristics failed, fallback heuristic is used, which reflects the old splitting behavior
+    if (batch_dim < optimal_parallelism_work_amount)
+        std::tie(batch_m_dim, new_m_dim) =
+            split_fallback_increase_parallel_wa(batch_dim, m_dim, optimal_parallelism_work_amount);
+    return split_is_done();
+}
+
+std::pair<size_t, size_t> MHAParallelWAOptimizer::split_ideally(size_t batch_dim,
+                                                                size_t m_dim,
+                                                                size_t optimal_parallelism_work_amount) {
+    // Ideal case #1: M can be split on the parts one of which complements the batch dimension to the optimal parallel
+    // work amount In this case, each thread will execute the Snippets kernel once
+    const size_t lower_bound = optimal_parallelism_work_amount / batch_dim;
+    if (lower_bound * batch_dim == optimal_parallelism_work_amount && m_dim % lower_bound == 0)
+        return std::make_pair(lower_bound, m_dim / lower_bound);
+
+    // Ideal case #2: M is divisible by optimal parallel work amount, and the new_m_dim is big enough
+    // In this case, each thread will execute the Snippets kernel 'batch_dim' times
+    if (m_dim % optimal_parallelism_work_amount == 0) {
+        const auto new_m_dim = m_dim / optimal_parallelism_work_amount;
+        if (new_m_dim >= m_min_kernel_m)
+            return std::make_pair(optimal_parallelism_work_amount, new_m_dim);
+    }
+
+    return std::make_pair(1, m_dim);
+}
+
+std::pair<size_t, size_t> MHAParallelWAOptimizer::split_fallback_increase_parallel_wa(
+    size_t batch_dim,
+    size_t m_dim,
+    size_t optimal_parallelism_work_amount) {
+    std::pair<size_t, size_t> splited = {1, m_dim};
+    const size_t upper_bound = utils::div_up(2 * optimal_parallelism_work_amount, batch_dim);
+    for (size_t divisor_0 = upper_bound - 1; divisor_0 > 1; divisor_0--) {
+        size_t divisor_1 = m_dim / divisor_0;
+        if (divisor_1 * divisor_0 == m_dim)
+            return divisor_0 * batch_dim >= optimal_parallelism_work_amount ? std::make_pair(divisor_0, divisor_1)
+                                                                            : splited;
+    }
+    return splited;
+}
+
+std::pair<size_t, size_t> MHAParallelWAOptimizer::split_minimize_kernel_wa(size_t batch_dim,
+                                                                           size_t m_dim,
+                                                                           size_t optimal_parallelism_work_amount) {
+    // This heuristic minimizes 'm_kernel' (=> maximizes 'm_batch') with a limitation that 'm_kernel >= min_kernel_m'.
+    // In other words, it tries to find 'm_kernel' bigger than 'm_min_kernel_m' and at the same time as close as possible
+    // to this value.
+    std::pair<size_t, size_t> best_result = {1, m_dim};
+    for (size_t divisor = 2; divisor < std::sqrt(m_dim); ++divisor) {
+        if (m_dim % divisor != 0)
+            continue;
+        // If divisor is more than 'm_min_kernel_m', divisor becomes 'm_kernel',
+        // guaranteeing the most optimal implementation from 'm_kernel' minimization perspective.
+        if (divisor >= m_min_kernel_m)
+            return std::make_pair(m_dim / divisor, divisor);
+
+        // If divisor is less than 'm_min_kernel_m', divisor becomes m_batch.
+        // However, it is not guaranteed that the current 'm_kernel = m_dim / divisor' is minimized, as one of the next
+        // divisors can be more optimal. So in this case the best result is remembered
+        const size_t m_kernel = m_dim / divisor;
+        if (m_kernel >= m_min_kernel_m) {
+            best_result.first = divisor;
+            best_result.second = m_kernel;
+        }
+    }
+    if (best_result.first * batch_dim >= optimal_parallelism_work_amount)
+        return best_result;
+    return std::make_pair(1, m_dim);
+}
+
 }  // namespace ov::snippets::lowered::pass

src/common/snippets/src/pass/common_optimizations.cpp

@@ -18,7 +18,6 @@
 #include "snippets/pass/extract_unsupported_transposes.hpp"
 #include "snippets/pass/fq_decomposition.hpp"
 #include "snippets/pass/softmax_reshape_elimination.hpp"
-#include "snippets/pass/split_dimension_m.hpp"
 #include "snippets/pass/subgraph_manager.hpp"
 #include "snippets/pass/transform_convert.hpp"
 #include "snippets/pass/validate.hpp"
@@ -60,10 +59,6 @@ CommonOptimizations::CommonOptimizations(const CommonOptimizations::Config& conf
                                ov::snippets::pass::ExtractUnsupportedTransposes,
                                is_domain_sensitive,
                                config.get_transpose_support_callback());
-        REGISTER_SNIPPETS_PASS(subgraph_manager,
-                               ov::snippets::pass::SplitDimensionM,
-                               is_domain_sensitive && config.get_split_m_dimension(),
-                               config.get_concurrency());
         subgraph_manager.run_passes(subgraph);
 
         // Validate the body after all common optimizations