vulkan: optimize mat_mul_id row_ids search to batch loads, and port to coopmat1 path

Author: jeffbolznv

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1821,7 +1821,7 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
     const uint32_t warps = warptile[0] / warptile[10];
 
     const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size;
-    const uint32_t mmid_row_ids = mul_mat_id ? 4096 * sizeof(uint32_t) : 0;
+    const uint32_t mmid_row_ids = mul_mat_id ? (4096 * sizeof(uint32_t) + 4/*_ne1*/) : 0;
     const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0;
 
     const uint32_t total_size = load_bufs + mmid_row_ids + coopmat_stage + lut_size;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -18,6 +18,7 @@
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_KHR_memory_scope_semantics : enable
 #extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
 #endif
 
 #ifdef MUL_MAT_ID
@@ -104,6 +105,10 @@ shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE];
 
 #ifdef MUL_MAT_ID
 shared u16vec2 row_ids[4096];
+uint _ne1;
+#ifdef COOPMAT
+shared uint _ne1_sh;
+#endif
 #endif // MUL_MAT_ID
 
 #define NUM_WARPS (BLOCK_SIZE / WARP)
@@ -172,7 +177,47 @@ void main() {
     const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
 
 #ifdef MUL_MAT_ID
-    uint _ne1 = 0;
+#ifdef COOPMAT
+    // Spread the search across all elements in the first subgroup
+    if (gl_SubgroupID == 0) {
+        _ne1 = 0;
+        uint num_elements = p.nei1 * p.nei0;
+
+        uint ids[16];
+        uint iter = 0;
+
+        for (uint j = 0; j < num_elements; j += gl_SubgroupSize) {
+            // prefetch up to 16 elements
+            if (iter == 0) {
+                [[unroll]] for (uint k = 0; k < 16; ++k) {
+                    uint i = j + gl_SubgroupInvocationID + k*gl_SubgroupSize;
+                    bool in_range = i < num_elements;
+                    uint ii1 = i / p.nei0;
+                    uint ii0 = i % p.nei0;
+                    ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+                }
+            }
+            uint i = j + gl_SubgroupInvocationID;
+            bool in_range = i < num_elements;
+            uint ii1 = i / p.nei0;
+            uint ii0 = i % p.nei0;
+            uint id = ids[iter++];
+            uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+            uint idx = subgroupBallotExclusiveBitCount(ballot);
+            if (in_range && id == expert_idx) {
+                row_ids[_ne1 + idx] = u16vec2(ii0, ii1);
+            }
+            _ne1 += subgroupBallotBitCount(ballot);
+            iter &= 15;
+        }
+        _ne1_sh = _ne1;
+    }
+
+    barrier();
+
+    _ne1 = _ne1_sh;
+#else
+    _ne1 = 0;
     for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
         for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
             if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
@@ -183,6 +228,7 @@ void main() {
     }
 
     barrier();
+#endif
 
     // Workgroup has no work
     if (ic * BN >= _ne1) return;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -162,17 +162,32 @@ void main() {
         _ne1 = 0;
         uint num_elements = p.nei1 * p.nei0;
 
-        for (uint i = gl_SubgroupInvocationID; subgroupAny(i < num_elements); i += gl_SubgroupSize) {
+        uint ids[16];
+        uint iter = 0;
+
+        for (uint j = 0; j < num_elements; j += gl_SubgroupSize) {
+            // prefetch up to 16 elements
+            if (iter == 0) {
+                [[unroll]] for (uint k = 0; k < 16; ++k) {
+                    uint i = j + gl_SubgroupInvocationID + k*gl_SubgroupSize;
+                    bool in_range = i < num_elements;
+                    uint ii1 = i / p.nei0;
+                    uint ii0 = i % p.nei0;
+                    ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+                }
+            }
+            uint i = j + gl_SubgroupInvocationID;
             bool in_range = i < num_elements;
-            uint ii0 = i % p.nei0;
             uint ii1 = i / p.nei0;
-            uint id = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            uint ii0 = i % p.nei0;
+            uint id = ids[iter++];
             uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
             uint idx = subgroupBallotExclusiveBitCount(ballot);
             if (in_range && id == expert_idx) {
                 row_ids[_ne1 + idx] = u16vec4(ii0 % p.ne11, ii1, ii0, 0);
             }
             _ne1 += subgroupBallotBitCount(ballot);
+            iter &= 15;
         }
         _ne1_sh = _ne1;
     }