feat(#1195): spin-then-park thread pool for the packed matmul kernels — 153 → 61 ms/step

skainet-backends/skainet-backend-jni-cpu/native/CMakeLists.txt

@@ -14,6 +14,7 @@ set(SKAINET_KERNEL_SOURCES
     ${SKAINET_KERNELS_ROOT}/src/q5_0_matmul.c
     ${SKAINET_KERNELS_ROOT}/src/q5_1_matmul.c
     ${SKAINET_KERNELS_ROOT}/src/q8_0_matmul.c
+    ${SKAINET_KERNELS_ROOT}/src/skainet_row_threads.c
     ${SKAINET_KERNELS_ROOT}/src/q4k_matmul.c
     ${SKAINET_KERNELS_ROOT}/src/q5k_matmul.c
     ${SKAINET_KERNELS_ROOT}/src/q6k_matmul.c

skainet-backends/skainet-backend-jni-cpu/src/androidTest/kotlin/sk/ainet/exec/harness/android/M2A5DeviceMeasurement.kt

@@ -105,6 +105,10 @@ class M2A5DeviceMeasurement {
         val prepack = args.getString("prepack")?.toBoolean() ?: false
         val steps = arg("steps", 16)
         val warmup = arg("warmup", 4)
+        // `residency=heap` (default mapped) restores heap staging — the #1193 A/B lever for
+        // models that fit the cap. The plan below prices the same form the load uses (#1190).
+        val residency = if (args.getString("residency") == "heap") WeightResidency.HEAP else WeightResidency.MAPPED
+        val loadForm = WeightForm(shape = WeightShapeOrientation.OUT_IN, residency = residency)
 
         val report = StringBuilder()
         fun line(s: String = "") { report.append(s).append('\n') }
@@ -116,13 +120,10 @@ class M2A5DeviceMeasurement {
         line("- device: ${Build.MANUFACTURER} ${Build.MODEL}, Android ${Build.VERSION.RELEASE} (SDK ${Build.VERSION.SDK_INT}), ABI ${Build.SUPPORTED_ABIS.firstOrNull()}")
         line("- ART heap cap (Runtime.maxMemory): ${mb(heapCap)}")
         line("- model: ${modelFile.name}, ${mb(modelFile.length())} on disk")
-        line("- ctx=$ctxLen, decode steps=$steps (warm-up $warmup), prepack=$prepack")
+        line("- ctx=$ctxLen, decode steps=$steps (warm-up $warmup), prepack=$prepack, residency=${residency.name.lowercase()}")
         line()
 
         // ---- plan, from the header only --------------------------------------------------
-        // The plan gets the same WeightForm the load below uses, so mapped-servable weights are
-        // budgeted against the page cache instead of the heap cap (#1189).
-        val loadForm = WeightForm(shape = WeightShapeOrientation.OUT_IN, residency = WeightResidency.MAPPED)
         val (plan, geometry) = MappedRandomAccessSource.open(modelPath).let { src ->
             StreamingGGUFReader.open(src).use { reader ->
                 val input = reader.planInput(ctx = ctxLen, formFor = { loadForm })

skainet-backends/skainet-backend-native-cpu/native/CMakeLists.txt

@@ -12,6 +12,7 @@ endif()
 set(SKAINET_KERNEL_SOURCES
     src/skainet_smoke.c
     src/skainet_cpu_features.c
+    src/skainet_row_threads.c
     src/q4k_matmul.c
     src/q5k_matmul.c
     src/q6k_matmul.c

skainet-backends/skainet-backend-native-cpu/native/include/skainet_kernels.h

@@ -48,6 +48,11 @@ SKAINET_API void skainet_smoke_double(const float* input, float* output, int32_t
  *
  * Caller owns input/weight/output memory; the kernel does not retain
  * pointers past return. input_dim must be a multiple of 256.
+ *
+ * Threads over output rows (disjoint out[] slices, pthreads, up to 4)
+ * when output_dim >= 512 (#1195); single-threaded below that and on
+ * MSVC. Results are bit-identical either way — per output row the
+ * accumulation order over blocks never changes.
  */
 SKAINET_API void skainet_q4k_matmul(
     const float* input,
@@ -68,6 +73,9 @@ SKAINET_API void skainet_q4k_matmul(
  * i.e. the bytes exactly as they sit in a .gguf file — which is what lets
  * mmap'd weights be fed to this kernel with no relayout copy (each row's
  * blocks are read strictly sequentially).
+ *
+ * Threads over output rows when output_dim >= 512 (#1195) — see
+ * skainet_q4k_matmul; bit-identical to the single-threaded result.
  */
 SKAINET_API void skainet_q4k_matmul_rm(
     const float* input,
@@ -120,6 +128,9 @@ SKAINET_API void skainet_q5k_matmul(
  *   weight + weight_byte_offset + (block_idx * output_dim + o) * 210
  *
  * input_dim must be a multiple of 256.
+ *
+ * Threads over output rows when output_dim >= 512 (#1195) — see
+ * skainet_q4k_matmul; bit-identical to the single-threaded result.
  */
 SKAINET_API void skainet_q6k_matmul(
     const float* input,
@@ -139,6 +150,9 @@ SKAINET_API void skainet_q6k_matmul(
  *   weight + weight_byte_offset + (o * blocks_per_row + block_idx) * 210
  * — the bytes exactly as they sit in a .gguf file, so an mmap'd weight
  * needs no relayout copy.
+ *
+ * Threads over output rows when output_dim >= 512 (#1195) — see
+ * skainet_q4k_matmul; bit-identical to the single-threaded result.
  */
 SKAINET_API void skainet_q6k_matmul_rm(
     const float* input,