ggml-hexagon: add HTP unary ops for ABS and LOG (#27786)
Add HVX-accelerated implementations for GGML_OP_LOG and GGML_UNARY_OP_ABS on the HTP backend. - Register HTP_OP_UNARY_ABS and HTP_OP_UNARY_LOG in op_remap_to_htp() - Add ABS and LOG to ggml_backend_hexagon_device_supports_op() - Implement hvx_abs_f32_aa() in hvx-arith.h using hvx_vec_abs_f32() - Implement hvx_log_f32_aa() in hvx-log.h using hvx_vec_log_f32() - Add abs_f32() and log_f32() row-wise dispatch in unary-ops.c - Define tiled and non-tiled task functions via DEFINE_UNARY_TASK and DEFINE_UNARY_TILED_TASK macros - Route HTP_OP_UNARY_ABS and HTP_OP_UNARY_LOG through execute_op() in main.c
This commit is contained in:
@@ -4643,6 +4643,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
|
||||
case GGML_OP_CLAMP: return HTP_OP_CLAMP;
|
||||
case GGML_OP_SQR: return HTP_OP_SQR;
|
||||
case GGML_OP_SQRT: return HTP_OP_SQRT;
|
||||
case GGML_OP_LOG: return HTP_OP_UNARY_LOG;
|
||||
case GGML_OP_SOFT_MAX: return HTP_OP_SOFTMAX;
|
||||
case GGML_OP_SSM_CONV: return HTP_OP_SSM_CONV;
|
||||
case GGML_OP_GATED_DELTA_NET: return HTP_OP_GATED_DELTA_NET;
|
||||
@@ -4666,6 +4667,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
|
||||
case GGML_UNARY_OP_EXP: return HTP_OP_UNARY_EXP;
|
||||
case GGML_UNARY_OP_SOFTPLUS: return HTP_OP_UNARY_SOFTPLUS;
|
||||
case GGML_UNARY_OP_TANH: return HTP_OP_UNARY_TANH;
|
||||
case GGML_UNARY_OP_ABS: return HTP_OP_UNARY_ABS;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
@@ -5463,6 +5465,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
|
||||
|
||||
case GGML_OP_SQR:
|
||||
case GGML_OP_SQRT:
|
||||
case GGML_OP_LOG:
|
||||
supp = ggml_hexagon_supported_unary(sess, op);
|
||||
break;
|
||||
|
||||
@@ -5481,6 +5484,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
|
||||
case GGML_UNARY_OP_SIGMOID:
|
||||
case GGML_UNARY_OP_SOFTPLUS:
|
||||
case GGML_UNARY_OP_TANH:
|
||||
case GGML_UNARY_OP_ABS:
|
||||
case GGML_UNARY_OP_SILU:
|
||||
case GGML_UNARY_OP_GELU:
|
||||
case GGML_UNARY_OP_GELU_QUICK:
|
||||
|
||||
@@ -62,6 +62,8 @@ enum htp_op_code {
|
||||
HTP_OP_UNARY_NEG,
|
||||
HTP_OP_UNARY_SOFTPLUS,
|
||||
HTP_OP_UNARY_TANH,
|
||||
HTP_OP_UNARY_ABS,
|
||||
HTP_OP_UNARY_LOG,
|
||||
HTP_OP_GLU_SWIGLU,
|
||||
HTP_OP_GLU_SWIGLU_OAI,
|
||||
HTP_OP_GLU_GEGLU,
|
||||
|
||||
@@ -358,6 +358,34 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
|
||||
}
|
||||
}
|
||||
|
||||
//
|
||||
// Abs
|
||||
//
|
||||
|
||||
static inline void hvx_abs_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src % 128 == 0);
|
||||
|
||||
HVX_Vector * restrict vdst = (HVX_Vector *) dst;
|
||||
HVX_Vector * restrict vsrc = (HVX_Vector *) src;
|
||||
|
||||
const uint32_t elem_size = sizeof(float);
|
||||
const uint32_t epv = 128 / elem_size;
|
||||
const uint32_t nvec = n / epv;
|
||||
const uint32_t nloe = n % epv;
|
||||
|
||||
uint32_t i = 0;
|
||||
|
||||
_Pragma("unroll(4)")
|
||||
for (; i < nvec; i++) {
|
||||
vdst[i] = hvx_vec_abs_f32(vsrc[i]);
|
||||
}
|
||||
if (nloe) {
|
||||
HVX_Vector v = hvx_vec_abs_f32(vsrc[i]);
|
||||
hvx_vec_store_a((void *) &vdst[i], nloe * elem_size, v);
|
||||
}
|
||||
}
|
||||
|
||||
//
|
||||
// Square
|
||||
//
|
||||
|
||||
@@ -62,4 +62,28 @@ static inline HVX_Vector hvx_vec_log_f32(HVX_Vector x) {
|
||||
return hvx_vec_add_f32_f32(term_e, res);
|
||||
}
|
||||
|
||||
static inline void hvx_log_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
|
||||
assert((unsigned long) dst % 128 == 0);
|
||||
assert((unsigned long) src % 128 == 0);
|
||||
|
||||
HVX_Vector * restrict vdst = (HVX_Vector *) dst;
|
||||
HVX_Vector * restrict vsrc = (HVX_Vector *) src;
|
||||
|
||||
const uint32_t elem_size = sizeof(float);
|
||||
const uint32_t epv = 128 / elem_size;
|
||||
const uint32_t nvec = n / epv;
|
||||
const uint32_t nloe = n % epv;
|
||||
|
||||
uint32_t i = 0;
|
||||
|
||||
_Pragma("unroll(4)")
|
||||
for (; i < nvec; i++) {
|
||||
vdst[i] = hvx_vec_log_f32(vsrc[i]);
|
||||
}
|
||||
if (nloe) {
|
||||
HVX_Vector v = hvx_vec_log_f32(vsrc[i]);
|
||||
hvx_vec_store_a((void *) &vdst[i], nloe * elem_size, v);
|
||||
}
|
||||
}
|
||||
|
||||
#endif /* HVX_LOG_H */
|
||||
|
||||
@@ -777,6 +777,8 @@ static int execute_op(struct htp_ops_context * octx) {
|
||||
case HTP_OP_UNARY_NEG:
|
||||
case HTP_OP_UNARY_EXP:
|
||||
case HTP_OP_UNARY_TANH:
|
||||
case HTP_OP_UNARY_ABS:
|
||||
case HTP_OP_UNARY_LOG:
|
||||
case HTP_OP_L2_NORM:
|
||||
return op_unary(octx);
|
||||
|
||||
|
||||
@@ -443,6 +443,34 @@ static void tanh_f32(const float * restrict src,
|
||||
}
|
||||
}
|
||||
|
||||
static void abs_f32(const float * restrict src,
|
||||
float * restrict dst,
|
||||
const uint32_t num_rows,
|
||||
const struct htp_unary_context * uctx) {
|
||||
htp_unary_op_preamble;
|
||||
|
||||
for (uint32_t ir = 0; ir < num_rows; ir++) {
|
||||
const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
|
||||
uint8_t * restrict dst_local = (uint8_t *)dst + (ir * dst_row_size_aligned);
|
||||
|
||||
hvx_abs_f32_aa(dst_local, src_local, ne0);
|
||||
}
|
||||
}
|
||||
|
||||
static void log_f32(const float * restrict src,
|
||||
float * restrict dst,
|
||||
const uint32_t num_rows,
|
||||
const struct htp_unary_context * uctx) {
|
||||
htp_unary_op_preamble;
|
||||
|
||||
for (uint32_t ir = 0; ir < num_rows; ir++) {
|
||||
const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
|
||||
uint8_t * restrict dst_local = (uint8_t *)dst + (ir * dst_row_size_aligned);
|
||||
|
||||
hvx_log_f32_aa(dst_local, src_local, ne0);
|
||||
}
|
||||
}
|
||||
|
||||
#define DEFINE_UNARY_TASK(NAME, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR) \
|
||||
static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * data) { \
|
||||
const struct htp_unary_context * uctx = (const struct htp_unary_context *) data; \
|
||||
@@ -611,6 +639,8 @@ DEFINE_UNARY_TASK(unary_silu, false, false, silu_f32(src0_vtcm, dst_vtcm, bl
|
||||
DEFINE_UNARY_TASK(unary_gelu, false, false, gelu_f32(src0_vtcm, dst_vtcm, block_size, uctx))
|
||||
DEFINE_UNARY_TASK(unary_softplus, false, false, softplus_f32(src0_vtcm, dst_vtcm, block_size, uctx))
|
||||
DEFINE_UNARY_TASK(unary_tanh, false, false, tanh_f32(src0_vtcm, dst_vtcm, block_size, uctx))
|
||||
DEFINE_UNARY_TASK(unary_abs, false, false, abs_f32(src0_vtcm, dst_vtcm, block_size, uctx))
|
||||
DEFINE_UNARY_TASK(unary_log, false, false, log_f32(src0_vtcm, dst_vtcm, block_size, uctx))
|
||||
DEFINE_UNARY_TASK(l2_norm, false, false, l2_norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
|
||||
DEFINE_UNARY_TASK(tri, false, true, tri_f32(src0_vtcm, dst_vtcm, block_size, ir, uctx))
|
||||
|
||||
@@ -858,6 +888,8 @@ DEFINE_UNARY_TILED_TASK(unary_silu, false, tile_silu_f32(dst_vtcm, src_vtcm,
|
||||
DEFINE_UNARY_TILED_TASK(unary_gelu, false, tile_gelu_f32(dst_vtcm, src_vtcm, tw))
|
||||
DEFINE_UNARY_TILED_TASK(unary_softplus, false, tile_unary_softplus_f32(dst_vtcm, src_vtcm, tw))
|
||||
DEFINE_UNARY_TILED_TASK(unary_tanh, false, hvx_tanh_f32_aa(dst_vtcm, src_vtcm, tw))
|
||||
DEFINE_UNARY_TILED_TASK(unary_abs, false, hvx_abs_f32_aa(dst_vtcm, src_vtcm, tw))
|
||||
DEFINE_UNARY_TILED_TASK(unary_log, false, hvx_log_f32_aa(dst_vtcm, src_vtcm, tw))
|
||||
DEFINE_UNARY_TILED_TASK(tri, true, tri_apply_tile_f32(src_vtcm, dst_vtcm, tw, col, i01, ne0, tri_ttype))
|
||||
|
||||
static int execute_op_unary_f32(struct htp_ops_context * octx) {
|
||||
@@ -883,6 +915,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
|
||||
case HTP_OP_UNARY_GELU: op_type = "gelu-f32"; break;
|
||||
case HTP_OP_UNARY_SOFTPLUS: op_type = "softplus-f32"; break;
|
||||
case HTP_OP_UNARY_TANH: op_type = "tanh-f32"; break;
|
||||
case HTP_OP_UNARY_ABS: op_type = "abs-f32"; break;
|
||||
case HTP_OP_UNARY_LOG: op_type = "log-f32"; break;
|
||||
case HTP_OP_L2_NORM: op_type = "l2norm-f32"; break;
|
||||
case HTP_OP_TRI: op_type = "tri-f32"; break;
|
||||
|
||||
@@ -981,6 +1015,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
|
||||
case HTP_OP_UNARY_GELU: task_func = unary_task_f32_tiled_unary_gelu; break;
|
||||
case HTP_OP_UNARY_SOFTPLUS: task_func = unary_task_f32_tiled_unary_softplus; break;
|
||||
case HTP_OP_UNARY_TANH: task_func = unary_task_f32_tiled_unary_tanh; break;
|
||||
case HTP_OP_UNARY_ABS: task_func = unary_task_f32_tiled_unary_abs; break;
|
||||
case HTP_OP_UNARY_LOG: task_func = unary_task_f32_tiled_unary_log; break;
|
||||
case HTP_OP_TRI: task_func = unary_task_f32_tiled_tri; break;
|
||||
default: break;
|
||||
}
|
||||
@@ -1000,6 +1036,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
|
||||
case HTP_OP_UNARY_GELU: task_func = unary_task_f32_unary_gelu; break;
|
||||
case HTP_OP_UNARY_SOFTPLUS: task_func = unary_task_f32_unary_softplus; break;
|
||||
case HTP_OP_UNARY_TANH: task_func = unary_task_f32_unary_tanh; break;
|
||||
case HTP_OP_UNARY_ABS: task_func = unary_task_f32_unary_abs; break;
|
||||
case HTP_OP_UNARY_LOG: task_func = unary_task_f32_unary_log; break;
|
||||
case HTP_OP_L2_NORM: task_func = unary_task_f32_l2_norm; break;
|
||||
case HTP_OP_TRI: task_func = unary_task_f32_tri; break;
|
||||
default: break;
|
||||
|
||||
@@ -55,6 +55,8 @@ static inline bool htp_op_is_unary(uint32_t opcode) {
|
||||
case HTP_OP_UNARY_GELU:
|
||||
case HTP_OP_UNARY_SOFTPLUS:
|
||||
case HTP_OP_UNARY_TANH:
|
||||
case HTP_OP_UNARY_ABS:
|
||||
case HTP_OP_UNARY_LOG:
|
||||
case HTP_OP_L2_NORM:
|
||||
case HTP_OP_TRI:
|
||||
return true;
|
||||
|
||||
Reference in New Issue
Block a user