@fugood/llama.node 0.3.16 → 0.4.0

package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.hpp

@@ -2,29 +2,28 @@
 #define GGML_SYCL_ELEMENTWISE_HPP
 
 #include "common.hpp"
+#include "ggml.h"
+#include <limits.h>
 
-static __dpct_inline__ float op_repeat(const float a, const float b) {
-    return b;
-    GGML_UNUSED(a);
+template <typename T>
+T neg_infinity() {
+    return -std::numeric_limits<T>::infinity();
 }
 
-static __dpct_inline__ float op_add(const float a, const float b) {
-    return a + b;
+template<typename T>
+struct typed_data {
+    const T * src;
+    T * dst;
+};
+
+template<typename T>
+typed_data<T> cast_data(ggml_tensor * dst) {
+    return {
+        /* .src = */ static_cast<const T *>(dst->src[0]->data),
+        /* .dst = */ static_cast<T *>(dst->data)
+    };
 }
 
-static __dpct_inline__ float op_sub(const float a, const float b) {
-    return a - b;
-}
-
-static __dpct_inline__ float op_mul(const float a, const float b) {
-    return a * b;
-}
-
-static __dpct_inline__ float op_div(const float a, const float b) {
-    return a / b;
-}
-
-
 void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
@@ -65,12 +64,12 @@ void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
-void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
-void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
-void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
-void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
+void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 #endif // GGML_SYCL_ELEMENTWISE_HPP
+

package/src/llama.cpp/ggml/src/ggml-sycl/gemm.hpp

@@ -32,16 +32,36 @@ public:
         else static_assert(0);
     }
 
-    static inline void row_gemm(ggml_backend_sycl_context & ctx, bool a_trans, bool b_trans, int m, int n, int k,
-                                const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
+    // matrix A has m rows, k columns
+    // matrix B has k rows, n columns
+    // nra - number of elements to skip when moving into next row in A
+    // nrb - number of elements to skip when moving into next row in B
+    // nca - number of elements to skip when moving into next column in A
+    // ncb - number of elements to skip when moving into next column in B
+    // stride_a - number of elements to skip when moving to next A matrix
+    // stride_b - number of elements to skip when moving to next B matrix
+    // batches_a - number of A matrices
+    // batches_b - number of B matrices
+    static void gemm(ggml_backend_sycl_context & ctx, int m, int n, int k,
+        const void * a, dt at, dnnl_dim_t nra, dnnl_dim_t nca, dnnl_dim_t stride_a,
+        const void * b, dt bt, dnnl_dim_t nrb, dnnl_dim_t ncb, dnnl_dim_t stride_b,
+        void * c, dt ct, const queue_ptr & q, dnnl_dim_t batches_a, dnnl_dim_t batches_b) {
+
         auto stream = ctx.stream_dnnl(q);
         auto eng = ctx.engine_dnnl(q);
-        dnnl::memory::dims a_dims = { m, k };
-        dnnl::memory::dims b_dims = { k, n };
-        dnnl::memory::dims c_dims = { m, n };
-        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
-        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
-        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::ab);
+
+        // { # strides, # rows, # columns }
+        dnnl::memory::dims a_dims = { batches_a, m, k };
+        dnnl::memory::dims b_dims = { batches_b, k, n };
+        dnnl::memory::dims c_dims = { std::max(batches_a, batches_b), m, n };
+
+        // { # elements to skip to next stride, # elements to skip to next row, # elements to skip to next column }
+        dnnl::memory::dims a_strides = { stride_a, nra, nca };
+        dnnl::memory::dims b_strides = { stride_b, nrb, ncb };
+
+        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_strides);
+        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_strides);
+        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::abc);
 
         dnnl::primitive_attr primitive_attr;
         primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
@@ -63,6 +83,15 @@ public:
 
         matmul_prim.execute(stream, matmul_args);
     }
+
+    // matrices A and B are column major, both having k rows
+    // matrix A has m column, matrix B has n columns
+    // output: column major matrix C = A transposed * B
+    static void row_gemm(ggml_backend_sycl_context & ctx, int m, int n, int k,
+        const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
+
+        gemm(ctx, m, n, k, a, at, k, 1, k * m, b, bt, 1, k, n * k, c, ct, q, 1, 1);
+    }
 };
 
 #endif

package/src/llama.cpp/ggml/src/ggml-sycl/getrows.cpp

@@ -257,50 +257,54 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens
     GGML_UNUSED(ctx);
 }
 
-void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                  const ggml_tensor *src1, ggml_tensor *dst,
-                                  const float *src0_d, const float *src1_d,
-                                  float *dst_d, const queue_ptr &stream) {
+void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
 
-    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->src[1]->type == GGML_TYPE_I32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
-    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
+    GGML_ASSERT(dst->src[0]->nb[0] == ggml_type_size(dst->src[0]->type));
+    GGML_ASSERT(dst->src[1]->nb[0] == ggml_type_size(dst->src[1]->type));
     GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
 
-    const int32_t * src1_i32 = (const int32_t *) src1_d;
-
-    switch (src0->type) {
+    const int32_t * src1_i32 = (const int32_t *) dst->src[1]->data;
+    /* TODO: Refactor and remove duplicates */
+    switch (dst->src[0]->type) {
         case GGML_TYPE_F16:
-            get_rows_sycl_float(ctx, src0, src1, dst, (const sycl::half *)src0_d,
-                                src1_i32, dst_d, stream);
+            get_rows_sycl_float(ctx, dst->src[0], dst->src[1], dst, (const sycl::half *)dst->src[0]->data,
+                                src1_i32, (float *)dst->data, ctx.stream());
             break;
         case GGML_TYPE_F32:
-            get_rows_sycl_float(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_sycl_float(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+            src1_i32, (float *)dst->data, ctx.stream());
             break;
         case GGML_TYPE_Q4_0:
             if (ctx.opt_feature.reorder && dst->op == GGML_OP_MUL_MAT) {
-                get_rows_sycl_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+                get_rows_sycl_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+                src1_i32, (float *)dst->data, ctx.stream());
             } else {
-                get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+                get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+                src1_i32, (float *)dst->data, ctx.stream());
             }
             break;
         case GGML_TYPE_Q4_1:
-            get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+            src1_i32, (float *)dst->data, ctx.stream());
             break;
         case GGML_TYPE_Q5_0:
-            get_rows_sycl<QK5_0, QR5_0, dequantize_q5_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_sycl<QK5_0, QR5_0, dequantize_q5_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+            src1_i32, (float *)dst->data, ctx.stream());
             break;
         case GGML_TYPE_Q5_1:
-            get_rows_sycl<QK5_1, QR5_1, dequantize_q5_1>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_sycl<QK5_1, QR5_1, dequantize_q5_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+            src1_i32, (float *)dst->data, ctx.stream());
             break;
         case GGML_TYPE_Q8_0:
-            get_rows_sycl<QK8_0, QR8_0, dequantize_q8_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            get_rows_sycl<QK8_0, QR8_0, dequantize_q8_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+            src1_i32, (float *)dst->data, ctx.stream());
             break;
         default:
             // TODO: k-quants
-            GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
+            GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(dst->src[0]->type));
             GGML_ABORT("fatal error");
     }
 }

package/src/llama.cpp/ggml/src/ggml-sycl/getrows.hpp

@@ -15,9 +15,6 @@
 
 #include "common.hpp"
 
-void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-    const ggml_tensor *src1, ggml_tensor *dst,
-    const float *src0_d, const float *src1_d,
-    float *dst_d, const queue_ptr &stream);
+void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst);
 
 #endif // GGML_SYCL_GETROWS_HPP