ggml : minor naming changes (ggml-org#8433)

* ggml : minor naming changes

ggml-ci

* ggml : use PRId64 [no ci]

* ggml : revert FA K/Q names

Author: ggerganov

examples/quantize-stats/quantize-stats.cpp

@@ -154,7 +154,7 @@ static void test_roundtrip_on_chunk(
     }
 
     if (use_reference) {
-        qfns.from_float_reference(input_scratch, quantized_scratch, chunk_size);
+        qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);
     } else {
         qfns.from_float(input_scratch, quantized_scratch, chunk_size);
     }

ggml/include/ggml.h

@@ -714,9 +714,9 @@ extern "C" {
     GGML_API GGML_CALL size_t  ggml_nbytes      (const struct ggml_tensor * tensor);
     GGML_API           size_t  ggml_nbytes_pad  (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
 
-    GGML_API GGML_CALL int    ggml_blck_size(enum ggml_type type);
-    GGML_API GGML_CALL size_t ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
-    GGML_API GGML_CALL size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
+    GGML_API GGML_CALL int64_t ggml_blck_size(enum ggml_type type);
+    GGML_API GGML_CALL size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
+    GGML_API GGML_CALL size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
 
     GGML_DEPRECATED(
     GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
@@ -2410,31 +2410,31 @@ extern "C" {
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
-    typedef void (*ggml_vec_dot_t)   (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
-                                      const void * GGML_RESTRICT y, size_t by, int nrc);
-    typedef void (*ggml_from_float_to_mat_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t nr,
-                                             int64_t k, int64_t bx);
-    typedef void (*ggml_gemv_t)      (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                      const void * GGML_RESTRICT y, int nr, int nc);
-    typedef void (*ggml_gemm_t)      (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                      const void * GGML_RESTRICT y, int nr, int nc);
+    typedef void (*ggml_from_float_to_mat_t)
+                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
+    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+                                       const void * GGML_RESTRICT y, size_t by, int nrc);
+    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
+    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
 
     typedef struct {
-        const char      * type_name;
-        int               blck_size;
-        size_t            type_size;
-        bool              is_quantized;
-        ggml_to_float_t   to_float;
-        ggml_from_float_t from_float;
-        ggml_from_float_t from_float_reference;
-        ggml_vec_dot_t    vec_dot;
-        enum ggml_type    vec_dot_type;
-        int64_t           nrows; // number of rows to process simultaneously;
-        int64_t           ncols; // number of columns to process simultaneously;
-        int64_t           interleave_blcksize; // interleave elements in blocks of interleave_blcksize;
+        const char             * type_name;
+        int64_t                  blck_size;
+        int64_t                  blck_size_interleave; // interleave elements in blocks
+        size_t                   type_size;
+        bool                     is_quantized;
+        ggml_to_float_t          to_float;
+        ggml_from_float_t        from_float;
+        ggml_from_float_t        from_float_ref;
         ggml_from_float_to_mat_t from_float_to_mat;
-        ggml_gemv_t       gemv;
-        ggml_gemm_t       gemm;
+        ggml_vec_dot_t           vec_dot;
+        enum ggml_type           vec_dot_type;
+        int64_t                  nrows; // number of rows to process simultaneously
+        int64_t                  ncols; // number of columns to process simultaneously
+        ggml_gemv_t              gemv;
+        ggml_gemm_t              gemm;
     } ggml_type_traits_t;
 
     GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);

ggml/src/ggml-aarch64.c

@@ -20,51 +20,51 @@
 
 // Functions to create the interleaved data layout formats
 
-// interleave 4 block_q4_0s in blocks of interleave_blcksize
+// interleave 4 block_q4_0s in blocks of blck_size_interleave
 // returns an interleaved block_q4_0x4
 // in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
-// first, then interleave quants from 4 block_q4_0s in blocks of interleave_blcksize
+// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
 //
 // - in                  : an array of block_q4_0 pointers
-// - interleave_blcksize : the block_q4_0 quants bytes are interleaved in blocks of
-//                         interleave_blcksize bytes
+// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
+//                         blck_size_interleave bytes
 // - xor_mask            : the mask to convert the nibbles in block_q4_0 quants bytes
 //                         from bias offset form to pure sign form (this saves subtract
 //                         operations durin unpacking)
 //
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int interleave_blcksize, unsigned int xor_mask) {
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
     block_q4_0x4 out;
 
     for (int i = 0; i < 4; i++) {
         out.d[i] = in[i].d;
     }
 
     for (int i = 0; i < QK4_0 * 2; i++) {
-        int src_offset = (i / (4 * interleave_blcksize)) * interleave_blcksize;
-        int src_id = (i % (4 * interleave_blcksize)) / interleave_blcksize;
-        src_offset += (i % interleave_blcksize);
+        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
+        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
+        src_offset += (i % blck_size_interleave);
 
         out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
     }
 
     return out;
 }
 
-// interleave 8 block_q4_0s in blocks of interleave_blcksize
+// interleave 8 block_q4_0s in blocks of blck_size_interleave
 // returns an interleaved block_q4_0x8
 // in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
-// first, then interleave quants from 8 block_q4_0s in blocks of interleave_blcksize
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int interleave_blcksize, unsigned int xor_mask) {
+// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
     block_q4_0x8 out;
 
     for (int i = 0; i < 8; i++) {
         out.d[i] = in[i].d;
     }
 
     for (int i = 0; i < QK4_0 * 4; i++) {
-        int src_offset = (i / (8 * interleave_blcksize)) * interleave_blcksize;
-        int src_id = (i % (8 * interleave_blcksize)) / interleave_blcksize;
-        src_offset += (i % interleave_blcksize);
+        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
+        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
+        src_offset += (i % blck_size_interleave);
 
         out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
     }
@@ -135,7 +135,7 @@ void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k)
     }
 #else
     // scalar
-    const int interleave_blcksize = 4;
+    const int blck_size_interleave = 4;
     float srcv[4][QK8_0];
     float id[4];
 
@@ -155,12 +155,12 @@ void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k)
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * interleave_blcksize)) * interleave_blcksize;
-            int src_id = (j % (4 * interleave_blcksize)) / interleave_blcksize;
-            src_offset += (j % interleave_blcksize);
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
 
             float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);;
+            y[i].qs[j] = roundf(x0);
         }
     }
 #endif
@@ -253,7 +253,7 @@ void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k)
     }
 #else
     // scalar
-    const int interleave_blcksize = 8;
+    const int blck_size_interleave = 8;
     float srcv[4][QK8_0];
     float id[4];
 
@@ -273,26 +273,30 @@ void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k)
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
-            int src_offset = (j / (4 * interleave_blcksize)) * interleave_blcksize;
-            int src_id = (j % (4 * interleave_blcksize)) / interleave_blcksize;
-            src_offset += (j % interleave_blcksize);
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
 
             float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);;
+            y[i].qs[j] = roundf(x0);
         }
     }
 #endif
 }
 
-void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t interleave_blcksize) {
+void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
     assert(nrow == 4);
     UNUSED(nrow);
-    if (interleave_blcksize == 4) quantize_q8_0_4x4(x, vy, n_per_row);
-    else if (interleave_blcksize == 8) quantize_q8_0_4x8(x, vy, n_per_row);
-    else assert(false);
+    if (blck_size_interleave == 4) {
+        quantize_q8_0_4x4(x, vy, n_per_row);
+    } else if (blck_size_interleave == 8) {
+        quantize_q8_0_4x8(x, vy, n_per_row);
+    } else {
+        assert(false);
+    }
 }
 
-static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int interleave_blcksize) {
+static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
     assert(n_per_row % QK4_0 == 0);
     const int nb = n_per_row / QK4_0;
 
@@ -311,15 +315,15 @@ static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict ds
         for (int64_t x = 0; x < nb; x++) {
 
             for (int i  = 0; i < nrows_interleaved; i++ ) {
-                quantize_row_q4_0_reference(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
+                quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
             }
 
             if (nrows_interleaved == 8) {
-                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, interleave_blcksize, 0x88);
+                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
                 out_ptr = (block_q4_0x8 *) out_ptr + 1;
             }
             else if (nrows_interleaved == 4) {
-                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, interleave_blcksize, 0x88);
+                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
                 out_ptr = (block_q4_0x4 *) out_ptr + 1;
             }
         }

ggml/src/ggml-aarch64.h

@@ -16,22 +16,22 @@ extern "C" {
 void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
-void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t interleave_blcksize);
+void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
 
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
 size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
 // GEMV
-void ggml_gemv_q4_0_4x4_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_gemv_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_gemv_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_gemv_q4_0_4x4_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_gemv_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_gemv_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);
 
 // GEMM
-void ggml_gemm_q4_0_4x4_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_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_gemm_q4_0_4x4_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_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);
 
 #ifdef __cplusplus
 }