@fugood/llama.node 0.3.2 → 0.3.4

package/src/llama.cpp/ggml/src/ggml-cann/kernels/CMakeLists.txt

@@ -1,7 +1,3 @@
-if (NOT SOC_TYPE)
-    set (SOC_TYPE "Ascend910B3")
-endif()
-
 file(GLOB SRC_FILES
     get_row_f32.cpp
     get_row_f16.cpp
@@ -13,7 +9,6 @@ file(GLOB SRC_FILES
     dup.cpp
 )
 
-string(TOLOWER ${SOC_TYPE} SOC_VERSION)
 set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
 set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
 
@@ -30,4 +25,6 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
+message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.")
+ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
 # ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

package/src/llama.cpp/ggml/src/ggml-cann/kernels/dup.cpp

@@ -5,6 +5,7 @@
 using namespace AscendC;
 
 #define BUFFER_NUM 2
+const int64_t SUPPORTED_MAX_DIM = 65535;  // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
 
 template <typename SRC_T, typename DST_T>
 class DupByRows {
@@ -51,24 +52,36 @@ class DupByRows {
 
     __aicore__ inline void copy_in() {
         LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
-
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = num_elem * sizeof(SRC_T);
-        DataCopyPadExtParams<SRC_T> padParams;
-        DataCopyPad(src_local, src_gm, dataCopyParams, padParams);
-
+        const size_t elem_per_block = 32 / sizeof(SRC_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
+        DataCopy(src_local, src_gm, cpy_elements_len);
         src_queue.EnQue(src_local);
     }
 
     __aicore__ inline void copy_out() {
         LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
-
+#ifdef ASCEND_310P
+        const size_t elem_per_block = 32 / sizeof(DST_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t len = num_elem & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(dst_gm, dst_local, len);
+        }
+        if(tail != 0) {
+            for (size_t i = tail; i < elem_per_block; i++) {
+                dst_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(dst_gm[len], dst_local[len], elem_per_block);
+            SetAtomicNone();
+        }
+#else
         DataCopyExtParams dataCopyParams;
         dataCopyParams.blockCount = 1;
         dataCopyParams.blockLen = num_elem * sizeof(DST_T);
         DataCopyPad(dst_gm, dst_local, dataCopyParams);
-
+#endif
         dst_queue.FreeTensor(dst_local);
     }
 

package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_f16.cpp

@@ -14,7 +14,7 @@ class GET_ROW_F16 {
                                 int64_t *output_ne_ub, size_t *output_nb_ub) {
         // TODO, use template for F16/f32
         int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
+        op_block_idx = GetBlockIdx();
 
         for (int i = 0; i < 4; i++) {
             input_ne[i] = input_ne_ub[i];
@@ -59,32 +59,42 @@ class GET_ROW_F16 {
     }
 
     __aicore__ inline void copy_in(uint32_t offset, size_t len) {
+        size_t origin_len = len;
         LocalTensor<half> input_local = input_queue.AllocTensor<half>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(input_local, input_gm[offset], len);
+        const size_t elem_per_block = 32 / sizeof(half);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
         if(tail != 0) {
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(half);
-            DataCopyPadExtParams<half> padParams;
-            DataCopyPad(input_local[len], input_gm[offset + len],
-                        dataCopyParams, padParams);
+            len += elem_per_block;
         }
+        DataCopy(input_local, input_gm[offset], len);
         input_queue.EnQue(input_local);
     }
 
     __aicore__ inline void copy_out(uint32_t offset, size_t len) {
         LocalTensor<float> output_local = output_queue.DeQue<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(output_gm[offset], output_local, len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(output_gm[offset], output_local, len);
+        }
+
         if(tail != 0) {
+#ifdef ASCEND_310P
+            for (size_t i = tail; i < elem_per_block; i++) {
+                output_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
+            SetAtomicNone();
+#else
             DataCopyExtParams dataCopyParams;
             dataCopyParams.blockCount = 1;
             dataCopyParams.blockLen = tail * sizeof(float);
             DataCopyPad(output_gm[offset + len], output_local[len],
                         dataCopyParams);
+#endif
         }
         output_queue.FreeTensor(output_local);
     }
@@ -150,6 +160,7 @@ class GET_ROW_F16 {
     GlobalTensor<float> output_gm;
     TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
     TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    int64_t op_block_idx;
 };
 
 template <typename T>

package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_f32.cpp

@@ -13,7 +13,7 @@ class GET_ROW_F32 {
                                 int64_t *indices_ne_ub, size_t *indices_nb_ub,
                                 int64_t *output_ne_ub, size_t *output_nb_ub) {
         int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
+        op_block_idx = GetBlockIdx();
 
         for (int i = 0; i < 4; i++) {
             input_ne[i] = input_ne_ub[i];
@@ -55,31 +55,40 @@ class GET_ROW_F32 {
 
     __aicore__ inline void copy_in(uint32_t offset, size_t len) {
         LocalTensor<float> input_local = input_queue.AllocTensor<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(input_local, input_gm[offset], len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
         if(tail != 0) {
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(float);
-            DataCopyPadExtParams<float> padParams;
-            DataCopyPad(input_local[len], input_gm[offset + len],
-                        dataCopyParams, padParams);
+            len += elem_per_block;
         }
+        DataCopy(input_local, input_gm[offset], len);
         input_queue.EnQue(input_local);
     }
 
     __aicore__ inline void copy_out(uint32_t offset, size_t len) {
         LocalTensor<float> output_local = output_queue.DeQue<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(output_gm[offset], output_local, len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(output_gm[offset], output_local, len);
+        }
+
         if(tail != 0) {
+#ifdef ASCEND_310P
+            for (size_t i = tail; i < elem_per_block; i++) {
+                output_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
+            SetAtomicNone();
+#else
             DataCopyExtParams dataCopyParams;
             dataCopyParams.blockCount = 1;
             dataCopyParams.blockLen = tail * sizeof(float);
             DataCopyPad(output_gm[offset + len], output_local[len],
                         dataCopyParams);
+#endif
         }
         output_queue.FreeTensor(output_local);
     }
@@ -144,6 +153,7 @@ class GET_ROW_F32 {
     GlobalTensor<float> output_gm;
     TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
     TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    int64_t op_block_idx;
 };
 
 template <typename T>

package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp

@@ -2,6 +2,15 @@
 
 // optimize me. Use template to avoid copy code.
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support 4bit get row
+    extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
+        GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
+        GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
+        GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support 4bit get row.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 
@@ -191,3 +200,5 @@ extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
             indices_nb_ub, output_ne_ub, output_nb_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P

package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp

@@ -1,6 +1,14 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P
+    extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f16->8bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define QK8_0 32
@@ -206,3 +214,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P

package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp

@@ -1,6 +1,14 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support f32->8bit quantization
+    extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f32->8bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define QK8_0 32
@@ -204,3 +212,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P

package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp

@@ -1,6 +1,21 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support float->4bit quantization
+    extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f32->4bit quantization.\n");
+    }
+
+    extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f16->4bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define Group_Size 32
@@ -276,3 +291,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P

package/src/llama.cpp/ggml/src/ggml-common.h

@@ -6,7 +6,20 @@
 typedef uint16_t ggml_half;
 typedef uint32_t ggml_half2;
 
-#define GGML_COMMON_AGGR
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_CPP)
+#include <cstdint>
+
+typedef uint16_t ggml_half;
+typedef uint32_t ggml_half2;
+
+// std-c++ allow anonymous unions but some compiler warn on it
+#define GGML_COMMON_AGGR_U data
+// std-c++ do not allow it.
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_METAL)
@@ -15,7 +28,8 @@ typedef uint32_t ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_CUDA)
@@ -29,7 +43,8 @@ typedef half2 ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_HIP)
@@ -39,7 +54,8 @@ typedef half2 ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_SYCL)
@@ -49,7 +65,8 @@ typedef half2 ggml_half2;
 typedef sycl::half  ggml_half;
 typedef sycl::half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #endif
@@ -154,9 +171,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t qs[QK4_1 / 2]; // nibbles / quants
 } block_q4_1;
 static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
@@ -175,9 +192,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t qh[4];         // 5-th bit of quants
     uint8_t qs[QK5_1 / 2]; // nibbles / quants
 } block_q5_1;
@@ -196,37 +213,13 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half s; // d * sum(qs[i])
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 ds;
-    };
+    } GGML_COMMON_AGGR_U;
     int8_t qs[QK8_1]; // quants
 } block_q8_1;
 static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
 
-typedef struct {
-    ggml_half d[4];        // deltas for 4 q4_0 blocks
-    uint8_t qs[QK4_0 * 2]; // nibbles / quants for 4 q4_0 blocks
-} block_q4_0x4;
-static_assert(sizeof(block_q4_0x4) == 4 * sizeof(ggml_half) + QK4_0 * 2, "wrong q4_0x4 block size/padding");
-
-typedef struct {
-    ggml_half d[8];        // deltas for 8 q4_0 blocks
-    uint8_t qs[QK4_0 * 4]; // nibbles / quants for 8 q4_0 blocks
-} block_q4_0x8;
-static_assert(sizeof(block_q4_0x8) == 8 * sizeof(ggml_half) + QK4_0 * 4, "wrong q4_0x8 block size/padding");
-
-typedef struct {
-    ggml_half d[4];        // deltas for 4 q8_0 blocks
-    int8_t qs[QK8_0 * 4];  // quants for 4 q8_0 blocks
-} block_q8_0x4;
-static_assert(sizeof(block_q8_0x4) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong q8_0x4 block size/padding");
-
-typedef struct {
-    ggml_half d[8];        // deltas for 8 q8_0 blocks
-    int8_t qs[QK8_0 * 8];  // quants for 8 q8_0 blocks
-} block_q8_0x8;
-static_assert(sizeof(block_q8_0x8) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong q8_0x8 block size/padding");
-
 //
 // Ternary quantization
 //
@@ -261,9 +254,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
 } block_q2_K;
 static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
 
@@ -288,9 +281,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
@@ -305,9 +298,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qh[QK_K/8];           // quants, high bit
     uint8_t qs[QK_K/2];           // quants, low 4 bits
@@ -431,6 +424,13 @@ static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_
 #define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
 #define GGML_TABLE_END() };
 
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_CPP)
+#include <cstdint>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
+
 #define GGML_COMMON_IMPL
 #elif defined(GGML_COMMON_IMPL_METAL)
 #include <metal_stdlib>
@@ -473,7 +473,7 @@ GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
     240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
 GGML_TABLE_END()
 
-//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+//#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A // lowest compute capability for integer intrinsics
 GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
     0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
     0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,