backend cpu: add online flow for aarch64 Q4_0 GEMV/GEMM kernels

Makefile

@@ -874,6 +874,10 @@ ggml/src/ggml-cuda/%.o: \
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
 endif # GGML_HIPBLAS
 
+ifndef GGML_NO_CPU_AARCH64
+	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
+endif
+
 ifdef GGML_METAL
 	MK_CPPFLAGS += -DGGML_USE_METAL
 	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
@@ -883,7 +887,7 @@ ifdef GGML_METAL_NDEBUG
 endif
 ifdef GGML_METAL_EMBED_LIBRARY
 	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJ_GGML   += ggml/src/ggml-metal-embed.o
+	OBJ_GGML    += ggml/src/ggml-metal-embed.o
 endif
 endif # GGML_METAL
 

ggml/CMakeLists.txt

@@ -92,6 +92,7 @@ else()
 endif()
 
 option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
+option(GGML_CPU_AARCH64 "ggml: use runtime weight conversionn of Q4_0 to Q4_X_X" ON)
 
 option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
 option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})

ggml/include/ggml-cpu.h

@@ -145,6 +145,11 @@ extern "C" {
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
 #endif
 
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
+    GGML_API bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft);
+
+
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/CMakeLists.txt

@@ -880,6 +880,12 @@ if (GGML_CPU_HBM)
     target_link_libraries(ggml PUBLIC memkind)
 endif()
 
+if (GGML_CPU_AARCH64)
+    message(STATUS "Using runtime weight conversion of Q4_0 to Q4_0_x_x to enable optimized GEMM/GEMV kernels")
+
+    add_compile_definitions(GGML_USE_CPU_AARCH64)
+endif()
+
 if (GGML_CANN)
     if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
         set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})

ggml/src/ggml-aarch64.c

@@ -3476,3 +3476,111 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         }
     }
 }
+
+static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * restrict data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+
+    block_q4_0x4 * dst = (block_q4_0x4 *)t->data;
+    const block_q4_0 * src = (const block_q4_0 *)data;
+    block_q4_0 dst_tmp[4];
+    int nrow = t->ne[1]; // Number of rows
+    int nrows_interleaved = 4;
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_0x4(dst_tmp, interleave_block, 0x88);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor *t, int interleave_block, const void * restrict data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(interleave_block == 8);
+
+    block_q4_0x8 * dst = (block_q4_0x8*)t->data;
+    const block_q4_0 * src = (const block_q4_0*) data;
+    block_q4_0 dst_tmp[8];
+    int nrow = t->ne[1]; // Number of rows
+    int nrows_interleaved = 8;
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_0x8(dst_tmp, interleave_block, 0x88);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+// Prepare for optimized kernels if applicable
+void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * restrict data, size_t data_size) {
+    int ret = -1;
+
+    if (cur->type == repack_type) {
+        memcpy(cur->data, data, data_size);
+        return;
+    }
+
+    GGML_ASSERT(cur->type == GGML_TYPE_Q4_0);
+
+    switch (repack_type) {
+        case GGML_TYPE_Q4_0_8_8:
+            ret = repack_q4_0_to_q4_0_8_bl(cur, 8, data, data_size);
+            break;
+        case GGML_TYPE_Q4_0_4_8:
+            ret = repack_q4_0_to_q4_0_4_bl(cur, 8, data, data_size);
+            break;
+        case GGML_TYPE_Q4_0_4_4:
+            ret = repack_q4_0_to_q4_0_4_bl(cur, 4, data, data_size);
+            break;
+        default:
+            GGML_ABORT("Unsupported type");
+    }
+    if (ret == -1) {
+        memcpy(cur->data, data, data_size);
+    }
+}
+
+enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur) {
+    if (cur->type == GGML_TYPE_Q4_0) {
+        // TODO: enable for AVX2 - currently disabled due to bad gemv performance
+        if (/* ggml_cpu_has_avx2() || */ (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
+            return GGML_TYPE_Q4_0_8_8;
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            return GGML_TYPE_Q4_0_4_8;
+        }
+        if (ggml_cpu_has_neon()) {
+            return GGML_TYPE_Q4_0_4_4;
+        }
+    }
+
+    return cur->type;
+}

ggml/src/ggml-aarch64.h

@@ -33,6 +33,9 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
+void           ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size);
+enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur);
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-backend.cpp

@@ -2239,15 +2239,104 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
 }
 #endif
 
+// buffer type AARCH64
+
+#ifdef __GNUC__
+    #pragma GCC diagnostic push
+    #pragma GCC diagnostic ignored "-Wpedantic"
+#endif
+
+#include "ggml-aarch64.h"
+
+#ifdef __GNUC__
+    #pragma GCC diagnostic pop
+#endif
+
+static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    tensor->extra = (void *)ggml_aarch64_get_optimal_repack_type(tensor); // NOLINT
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset == 0);
+    GGML_ASSERT(size == ggml_nbytes(tensor));
+
+    enum ggml_type repack_type = (enum ggml_type)(intptr_t)tensor->extra;
+
+    ggml_aarch64_repack_tensor(tensor, repack_type, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static const struct ggml_backend_buffer_i ggml_backend_cpu_aarch64_buffer_i = {
+    /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_cpu_aarch64_buffer_init_tensor,
+    /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cpu_aarch64_buffer_set_tensor,
+    /* .get_tensor      = */ NULL,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_AARCH64";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * data = ggml_aligned_malloc(size);
+
+    if (data == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cpu_aarch64_buffer_i, data, size);
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cpu_aarch64_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ NULL,
+        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_aarch64;
+}
+
+bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft) {
+    return buft == ggml_backend_cpu_aarch64_buffer_type();
+}
+
 static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
-    static ggml_backend_buffer_type_t bufts[] = {
+    static std::vector<ggml_backend_buffer_type_t> bufts = []() {
+        std::vector<ggml_backend_buffer_type_t> bufts;
+
 #ifdef GGML_USE_CPU_HBM
-        ggml_backend_cpu_hbm_buffer_type(),
+        bufts.push_back(ggml_backend_cpu_hbm_buffer_type());
 #endif
-        NULL
-    };
 
-    return bufts;
+#ifdef GGML_USE_CPU_AARCH64
+        bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+#endif
+
+        bufts.push_back(NULL);
+
+        return bufts;
+    }();
+
+    return bufts.data();
 
     GGML_UNUSED(device);
 }
@@ -2558,6 +2647,21 @@ static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_b
 }
 
 static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    if (src0 && src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
+        if (op->op != GGML_OP_MUL_MAT || src0->type != GGML_TYPE_Q4_0 || ggml_aarch64_get_optimal_repack_type(src0) == GGML_TYPE_Q4_0) {
+            return false;
+        }
+    }
+
+    for (int i = 1; i < GGML_MAX_SRC; i++) {
+        if (op->src[i] && op->src[i]->buffer && ggml_backend_cpu_buft_is_aarch64(op->src[i]->buffer->buft)) {
+            return false;
+        }
+    }
+
     switch (op->op) {
         case GGML_OP_CPY:
             return
@@ -2566,13 +2670,13 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
                 op->type != GGML_TYPE_IQ1_S   &&
                 op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
-            return op->src[1]->type == GGML_TYPE_F32;// FIXME || op->src[1]->type == ggml_get_type_traits(op->src[0]->type)->vec_dot_type;
+            return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
         case GGML_OP_ROPE_BACK:
             return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
         case GGML_OP_IM2COL_BACK:
-            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+            return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
         case GGML_OP_OUT_PROD:
-            return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
+            return (src0->type == GGML_TYPE_F32 || ggml_is_quantized(src0->type)) && src1->type == GGML_TYPE_F32;
         default:
             return true;
     }
@@ -2581,7 +2685,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
 }
 
 static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_buft_is_host(buft);
+    return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_buft_is_aarch64(buft);
 
     GGML_UNUSED(dev);
 }
@@ -2652,6 +2756,9 @@ static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
 };
 
 ggml_backend_reg_t ggml_backend_cpu_reg(void) {
+    // init CPU feature detection
+    ggml_cpu_init();
+
     static struct ggml_backend_reg ggml_backend_cpu_reg = {
         /* .iface   = */ ggml_backend_cpu_reg_i,
         /* .context = */ NULL,