attempted speedups 2

Author: apage43

ggml-vulkan.cpp

@@ -989,26 +989,27 @@ void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq,
         nb1, nb2
     };
 
+    const uint32_t local_x = ggml_vk_current_device().subgroupSize;
     std::shared_ptr<kp::Algorithm> s_algo = nullptr;
     if (!komputeManager()->hasAlgorithm(__func__)) {
-        //std::cerr << "init f32 matmat shader" << std::endl;
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(),
+        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(),
         {inA, inB, out}, spirv,
         {unsigned(ne01),
          unsigned(ne11),
-         unsigned(ne12)},
-        {},
+         unsigned(std::max(ne12, ne02))
+         },
+        {local_x},
         {pushConsts});
     } else {
         s_algo = komputeManager()->getAlgorithm(__func__);
         s_algo->setTensors({inA, inB, out});
         s_algo->setWorkgroup({unsigned(ne01),
                               unsigned(ne11),
-                              unsigned(std::max(ne12, ne02))});
+                              unsigned(std::max(ne12, ne02)),
+                              });
         s_algo->setPushConstants<PushConstants>({pushConsts});
         s_algo->updateDescriptors(s_kompute_context->pool.get());
     }
-    //seq.record<kp::OpTensorFill>({out});
     seq.record<kp::OpAlgoDispatch>(s_algo);
 }
 
@@ -1038,15 +1039,16 @@ void ggml_vk_mul_mat_mat_f16(kp::Sequence& seq,
         nb1, nb2
     };
 
+    const uint32_t local_x = ggml_vk_current_device().subgroupSize;
     std::shared_ptr<kp::Algorithm> s_algo = nullptr;
     if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(),
+        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(),
         {inA, inB, out}, spirv,
         {unsigned(ne01),
          unsigned(ne11),
          unsigned(std::max(ne12, ne02))
          },
-        {},
+        {local_x},
         {pushConsts});
     } else {
         s_algo = komputeManager()->getAlgorithm(__func__);
@@ -1141,7 +1143,7 @@ void ggml_vk_mul_mat_mat_q6_k(
     if (!komputeManager()->hasAlgorithm(__func__)) {
         s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(),
         {inA, inB, out}, spirv,
-        {unsigned(ne01)/32,
+        {unsigned(ne01)/256,
          unsigned(ne11),
          unsigned(std::max(ne12, ne02))
          },
@@ -1150,7 +1152,7 @@ void ggml_vk_mul_mat_mat_q6_k(
     } else {
         s_algo = komputeManager()->getAlgorithm(__func__);
         s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01)/32,
+        s_algo->setWorkgroup({unsigned(ne01)/256,
                               unsigned(ne11),
                               unsigned(std::max(ne12, ne02)),
                               });
@@ -1192,7 +1194,7 @@ void ggml_vk_mul_mat_mat_q4_x(const std::vector<uint32_t>& spirv,
         {unsigned(ne01),
          unsigned(ne11),
          unsigned(std::max(ne12, ne02))},
-        {local_x, 4},
+        {local_x, 1},
         {pushConsts});
     } else {
         s_algo = komputeManager()->getAlgorithm(__func__);

kompute/op_mul_mat_mat_f16.comp

@@ -14,7 +14,8 @@
 #extension GL_KHR_shader_subgroup_arithmetic : require
 #extension GL_EXT_debug_printf : enable
 
-// layout(local_size_x = 8) in;
+// device subgroup size
+layout (local_size_x_id = 0) in;
 
 layout(binding = 0) readonly buffer tensorInA { float16_t inA[]; };
 layout(binding = 1) readonly buffer tensorInB { float inB[]; };
@@ -40,17 +41,20 @@ pcs;
 
 
 void main() {
-  uvec3 gid = gl_GlobalInvocationID;
+  uvec3 gid = gl_WorkGroupID;
 
   uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
   uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
 
   const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 2 + pcs.inAOff; // Based from inA
   const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
   float sum = 0.0f;
-  for (uint i = 0; i < pcs.ne00; i ++) {
+  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
       sum += float(inA[x+i]) * float(inB[y+i]);
   }
 
-  out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = sum;
+  const float all_sum = subgroupAdd(sum);
+  if (subgroupElect()) {
+    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
+  }
 }

kompute/op_mul_mat_mat_f32.comp

@@ -14,7 +14,8 @@
 #extension GL_KHR_shader_subgroup_arithmetic : require
 #extension GL_EXT_debug_printf : enable
 
-// layout(local_size_x = 8) in;
+// device subgroup size
+layout (local_size_x_id = 0) in;
 
 layout(binding = 0) readonly buffer tensorInA { float inA[]; };
 layout(binding = 1) readonly buffer tensorInB { float inB[]; };
@@ -40,14 +41,20 @@ pcs;
 
 
 void main() {
-  uvec3 gid = gl_GlobalInvocationID;
+  uvec3 gid = gl_WorkGroupID;
 
-  const uint x = (gid.x*pcs.nb01 + gid.z/(pcs.ne12/pcs.ne02)*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
-  const uint y = (gid.y*pcs.nb11 + gid.z/(pcs.ne02/pcs.ne12)*pcs.nb12) / 4 + pcs.inBOff; // based from inB
+  uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
+  uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
+
+  const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
+  const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
   float sum = 0.0f;
-  for (uint i = 0; i < pcs.ne00; i ++) {
+  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
       sum += float(inA[x+i]) * float(inB[y+i]);
   }
 
-  out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = sum;
-}
+  const float all_sum = subgroupAdd(sum);
+  if (subgroupElect()) {
+    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
+  }
+}

kompute/op_mul_mat_mat_q6_k.comp

@@ -14,7 +14,7 @@
 #extension GL_KHR_shader_subgroup_arithmetic : require
 #extension GL_EXT_debug_printf : enable
 
-layout(local_size_x = 32) in;
+layout(local_size_x = 256) in;
 
 layout(binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
 layout(binding = 1) readonly buffer tensorInB { float inB[]; };