whispercpp 1.3.5 → 1.3.7

data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h

@@ -1,6 +1,19 @@
 #ifndef GGML_METAL_IMPL
 #define GGML_METAL_IMPL
 
+// kernel parameters for mat-mat threadgroups
+//
+// TODO: become function constants
+
+#define SZ_SIMDGROUP 16
+#define N_MM_NK 2
+#define N_MM_NK_TOTAL (SZ_SIMDGROUP * N_MM_NK)
+
+#define N_MM_BLOCK_X 4
+#define N_MM_BLOCK_Y 2
+#define N_MM_SIMD_GROUP_X 2
+#define N_MM_SIMD_GROUP_Y 2
+
 // kernel parameters for mat-vec threadgroups
 //
 // N_R0: number of src0 rows to process per simdgroup
@@ -8,6 +21,9 @@
 //
 // TODO: for optimal performance, become function of the device and work size
 
+#define N_R0_Q1_0 8
+#define N_SG_Q1_0 2
+
 #define N_R0_Q4_0 4
 #define N_SG_Q4_0 2
 
@@ -35,7 +51,7 @@
 #define N_R0_Q4_K 2
 #define N_SG_Q4_K 2
 
-#define N_R0_Q5_K 2
+#define N_R0_Q5_K 1
 #define N_SG_Q5_K 2
 
 #define N_R0_Q6_K 2
@@ -78,15 +94,57 @@
 #define FC_MUL_MM                      700
 #define FC_ROPE                        800
 #define FC_SSM_CONV                    900
-#define FC_COUNT_EQUAL                 1000
+#define FC_SOLVE_TRI                   1000
+#define FC_COUNT_EQUAL                 1100
+#define FC_UNARY                       1200
+#define FC_BIN                         1300
+#define FC_SUM_ROWS                    1400
+#define FC_UPSCALE                     1500
+#define FC_GATED_DELTA_NET             1600
 
 // op-specific constants
-#define OP_FLASH_ATTN_EXT_NQPTG 8
+#define OP_FLASH_ATTN_EXT_NQPSG 8
 #define OP_FLASH_ATTN_EXT_NCPSG 64
 
-#define OP_FLASH_ATTN_EXT_VEC_NQPTG 1
+#define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
 #define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
 
+#define OP_UNARY_NUM_SCALE      10
+#define OP_UNARY_NUM_FILL       11
+#define OP_UNARY_NUM_CLAMP      12
+#define OP_UNARY_NUM_SQR        13
+#define OP_UNARY_NUM_SQRT       14
+#define OP_UNARY_NUM_SIN        15
+#define OP_UNARY_NUM_COS        16
+#define OP_UNARY_NUM_LOG        17
+#define OP_UNARY_NUM_LEAKY_RELU 18
+
+#define OP_UNARY_NUM_TANH        100
+#define OP_UNARY_NUM_RELU        101
+#define OP_UNARY_NUM_SIGMOID     102
+#define OP_UNARY_NUM_GELU        103
+#define OP_UNARY_NUM_GELU_ERF    104
+#define OP_UNARY_NUM_GELU_QUICK  105
+#define OP_UNARY_NUM_SILU        106
+#define OP_UNARY_NUM_ELU         107
+#define OP_UNARY_NUM_NEG         108
+#define OP_UNARY_NUM_ABS         109
+#define OP_UNARY_NUM_SGN         110
+#define OP_UNARY_NUM_STEP        111
+#define OP_UNARY_NUM_HARDSWISH   112
+#define OP_UNARY_NUM_HARDSIGMOID 113
+#define OP_UNARY_NUM_EXP         114
+#define OP_UNARY_NUM_SOFTPLUS    115
+#define OP_UNARY_NUM_EXPM1       116
+#define OP_UNARY_NUM_FLOOR       117
+#define OP_UNARY_NUM_CEIL        118
+#define OP_UNARY_NUM_ROUND       119
+#define OP_UNARY_NUM_TRUNC       120
+#define OP_UNARY_NUM_XIELU       121
+
+#define OP_SUM_ROWS_NUM_SUM_ROWS 10
+#define OP_SUM_ROWS_NUM_MEAN     11
+
 // kernel argument structs
 //
 // - element counters (e.g. ne00) typically use int32_t to reduce register usage
@@ -122,6 +180,31 @@ typedef struct {
     int32_t  dim;
 } ggml_metal_kargs_concat;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    float    slope;
+    float    scale;
+    float    bias;
+    float    val;
+    float    min;
+    float    max;
+} ggml_metal_kargs_unary;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -179,20 +262,6 @@ typedef struct {
     uint64_t nb3;
 } ggml_metal_kargs_repeat;
 
-typedef struct {
-    float scale;
-    float bias;
-} ggml_metal_kargs_scale;
-
-typedef struct {
-    float val;
-} ggml_metal_kargs_fill;
-
-typedef struct {
-    float min;
-    float max;
-} ggml_metal_kargs_clamp;
-
 typedef struct {
     int64_t  nk0;
     int64_t  ne00;
@@ -496,8 +565,21 @@ typedef struct {
 
 typedef struct {
     int32_t  ne00;
-    int32_t  ne00_4;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
     uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
     float    eps;
 } ggml_metal_kargs_l2_norm;
 
@@ -582,6 +664,42 @@ typedef struct {
     int32_t  KHW; // KH * KW, pre-computed on CPU to save GPU resources
 } ggml_metal_kargs_im2col;
 
+typedef struct {
+    int32_t  IW;
+    int32_t  IH;
+    int32_t  ID;
+    int32_t  OW;
+    int32_t  OH;
+    int32_t  OD;
+    int32_t  KW;
+    int32_t  KH;
+    int32_t  KD;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  s2;
+    int32_t  p0;
+    int32_t  p1;
+    int32_t  p2;
+    int32_t  d0;
+    int32_t  d1;
+    int32_t  d2;
+    int32_t  IC;
+    int32_t  N;
+    int32_t  OC;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_conv_3d;
+
 typedef struct{
     int32_t  ne00;
     uint64_t nb01;
@@ -733,6 +851,71 @@ typedef struct {
     uint64_t nb0;
 } ggml_metal_kargs_ssm_scan;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne20;
+    int32_t  ne21;
+    int32_t  ne22;
+    int32_t  ne23;
+    uint64_t nb20;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
+    int32_t  ns02;
+    int32_t  ns12;
+    int32_t  ns22;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_gated_delta_net;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_solve_tri;
+
 typedef struct {
     int32_t  ne00t;
     int32_t  ne00;
@@ -764,6 +947,25 @@ typedef struct {
     uint64_t nb3;
 } ggml_metal_kargs_set_rows;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_diag;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;
@@ -785,6 +987,7 @@ typedef struct {
     float    sf1;
     float    sf2;
     float    sf3;
+    float    poffs;
 } ggml_metal_kargs_upscale;
 
 typedef struct {
@@ -827,16 +1030,35 @@ typedef struct {
     int32_t  p1;
 } ggml_metal_kargs_pad_reflect_1d;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  ne0;
+    int64_t  ne1;
+    int64_t  ne2;
+    int64_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  s2;
+    int32_t  s3;
+} ggml_metal_kargs_roll;
+
 typedef struct {
     uint64_t nb1;
     int      dim;
     int      max_period;
 } ggml_metal_kargs_timestep_embedding;
 
-typedef struct {
-    float    slope;
-} ggml_metal_kargs_leaky_relu;
-
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -928,6 +1150,15 @@ typedef struct {
     int64_t  np;
 } ggml_metal_kargs_pool_2d;
 
+typedef struct {
+    int32_t  k0;
+    int32_t  s0;
+    int32_t  p0;
+    int64_t  IW;
+    int64_t  OW;
+    int64_t  np;
+} ggml_metal_kargs_pool_1d;
+
 typedef struct {
      int64_t ne00;
     uint64_t nb01;