Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

hexagon: add support for CONCAT op (#23648)

Browse Source 
* hexagon: add support for CONCAT with optimized concat_2d_transposed

qwen3.5 models are quite heavy on the CONCAT with large and transposed src1.

* hex-concat: use fastdiv in generic version

* hex-concat: make checks for transposed a bit more readable

* hex-concat: reoder dma ops for better pipelining

* hex-cont/cpy: optimize CPY and CONT ops

The primary change is to avoid scalar divs in the inner loops.
We were calling hvx_copy_uu(... type_size) where type_size is non a constexpr.
This causes runtime divs by that value which is normally just 4 or 2 (f32/f16).

* hex-get-rows: optimize GET_ROWS for large rows

We now use DMA for larger rows and also split them into chunks to improve perf for Qwen3.5 and other models
that do lots of GET_ROWS with huge (2MB+ rows).

Also bump the DMA queue depth now that we can take advantage of it.

* hex-concat: unroll the inner loops of concat_2d

* hex-concat: more updates to concat_2d to improve perf a bit further

* hex-cpy: fixed n_rows per thread checks in the copy ops

* hmx-fa: fix alignment issues while computing dma sizes

* hex-set-rows: add early returns for idle threads

* hvx-rope: minor optimization to replace loops with fastdiv logic

* hex-rope: replace scalar tail processing with HVX

* hex-rope: optimize rope cache init with HVX

Add hvx-utils sin/cos helpers that use an aprox method (similar to rsqrt, inverse, etc)
Use the helpers to optimize ROPE.
This commit is contained in:
Max Krasnyansky
2026-05-26 06:20:05 -07:00
committed by GitHub
parent 678d43d720
commit ef66bfab68
14 changed files with 867 additions and 230 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-hexagon/htp/rope-ops.c
+180 -60
View File
@@ -7,6 +7,7 @@
#include <math.h>
#include <string.h>
#include <stdlib.h>
#include "hex-dma.h"
#include "hvx-utils.h"
@@ -75,6 +76,9 @@ struct htp_rope_context {
size_t theta_cache_offset;
uint32_t src0_nrows;
struct fastdiv_values div_ne2_ne1;
struct fastdiv_values div_ne1;
uint64_t t_start;
};
@@ -117,13 +121,84 @@ static __attribute__((noinline)) void rope_cache_init(const float theta_base,
float * cache,
const float theta_scale) {
// ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
float theta = theta_base;
#if __HVX_ARCH__ >= 79
const bool is_v79_or_newer = true;
#else
const bool is_v79_or_newer = false;
#endif
for (uint32_t i0 = 0; i0 < ne0; i0 += 2) {
const float ff = freq_factors ? freq_factors[i0 / 2] : 1.0f;
rope_yarn_one(theta / ff, freq_scale, corr_dims, i0, ext_factor, mscale, cache);
if (is_v79_or_newer && ext_factor == 0.0f) {
// Fast path: fully vectorized
// We process 32 pairs (64 elements) per iteration.
const uint32_t n_blocks = ne0 / 64;
theta *= theta_scale;
// Initialize theta scale powers: [1.0f, theta_scale, theta_scale^2, ..., theta_scale^31]
float __attribute__((aligned(128))) theta_powers[32];
theta_powers[0] = 1.0f;
for (int j = 1; j < 32; j++) {
theta_powers[j] = theta_powers[j - 1] * theta_scale;
}
HVX_Vector v_theta_powers = hvx_vmem(theta_powers);
HVX_Vector v_freq_scale = hvx_vec_splat_f32(freq_scale);
HVX_Vector v_mscale = hvx_vec_splat_f32(mscale);
// Base theta starts at theta_base
float theta_block = theta_base;
// The scale factor for the next block is theta_scale^32
float theta_scale_32 = 1.0f;
for (int j = 0; j < 32; j++) {
theta_scale_32 *= theta_scale;
}
for (uint32_t b = 0; b < n_blocks; b++) {
uint32_t i0 = b * 64;
HVX_Vector v_theta_base = hvx_vec_splat_f32(theta_block);
HVX_Vector v_theta = hvx_vec_mul_f32_f32(v_theta_base, v_theta_powers);
if (freq_factors) {
// Load 32 elements of freq_factors
HVX_Vector v_ff = hvx_vmemu(freq_factors + i0 / 2);
HVX_Vector v_inv_ff = hvx_vec_inverse_f32(v_ff);
v_theta = hvx_vec_mul_f32_f32(v_theta, v_inv_ff);
}
HVX_Vector v_theta_final = hvx_vec_mul_f32_f32(v_theta, v_freq_scale);
HVX_Vector vcos = hvx_vec_cos_f32(v_theta_final);
HVX_Vector vsin = hvx_vec_sin_f32(v_theta_final);
vcos = hvx_vec_mul_f32_f32(vcos, v_mscale);
vsin = hvx_vec_mul_f32_f32(vsin, v_mscale);
HVX_VectorPair vstore = Q6_W_vshuff_VVR(vsin, vcos, -4);
if (((uintptr_t)cache) % 128 == 0) {
hvx_vmem(cache + i0 + 0) = Q6_V_lo_W(vstore);
hvx_vmem(cache + i0 + 32) = Q6_V_hi_W(vstore);
} else {
hvx_vec_store_u(cache + i0 + 0, 32 * sizeof(float), Q6_V_lo_W(vstore));
hvx_vec_store_u(cache + i0 + 32, 32 * sizeof(float), Q6_V_hi_W(vstore));
}
theta_block *= theta_scale_32;
}
// Leftovers
float theta = theta_block;
for (uint32_t i0 = n_blocks * 64; i0 < ne0; i0 += 2) {
const float ff = freq_factors ? freq_factors[i0 / 2] : 1.0f;
rope_yarn_one(theta / ff, freq_scale, corr_dims, i0, ext_factor, mscale, cache);
theta *= theta_scale;
}
} else {
// Fallback to original scalar loop
float theta = theta_base;
for (uint32_t i0 = 0; i0 < ne0; i0 += 2) {
const float ff = freq_factors ? freq_factors[i0 / 2] : 1.0f;
rope_yarn_one(theta / ff, freq_scale, corr_dims, i0, ext_factor, mscale, cache);
theta *= theta_scale;
}
}
}
@@ -195,24 +270,18 @@ static void rope_corr_dims(int n_dims,
}
static inline void hvx_rope_neox_f32_aa(float * restrict dst, const float * restrict src0, uint32_t ne, const float * restrict theta_cache) {
const HVX_Vector * restrict vsrc = (const HVX_Vector *) src0;
const HVX_Vector * restrict vtheta = (const HVX_Vector *) theta_cache;
HVX_Vector * restrict vdst = (HVX_Vector *) dst;
const uint32_t he = ne / 2;
const uint32_t nvec = he / 32;
const uint32_t nloe = he % 32;
uint32_t nvec = (ne / (VLEN_FP32 * 2) * 2); // 2 vecs per loop, step of 2
for (uint32_t i = 0; i < nvec; i++) {
HVX_Vector v0 = ((const HVX_Vector *) src0)[i];
HVX_Vector v1 = hvx_vmemu(src0 + he + i * 32);
uint32_t he = ne / 2; // half_dims offset in elements
uint32_t hv = he / VLEN_FP32; // half_dims offset in vectors
HVX_Vector v2 = ((const HVX_Vector *) theta_cache)[i * 2 + 0];
HVX_Vector v3 = ((const HVX_Vector *) theta_cache)[i * 2 + 1];
#pragma unroll(2)
for (uint32_t i = 0; i < nvec; i += 2) {
HVX_Vector v0 = vsrc[i/2+0];
HVX_Vector v1 = vsrc[i/2+hv];
HVX_Vector v2 = vtheta[i+0];
HVX_Vector v3 = vtheta[i+1];
HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4); // vcos_sin[0] = cos_theta, vcos_sin[1] = sin_theta
HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);
HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_lo_W(vcos_sin));
HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_hi_W(vcos_sin));
@@ -222,37 +291,45 @@ static inline void hvx_rope_neox_f32_aa(float * restrict dst, const float * rest
HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
vdst[i/2+0] = Q6_Vsf_equals_Vqf32(v4);
vdst[i/2+hv] = Q6_Vsf_equals_Vqf32(v5);
((HVX_Vector *) dst)[i] = Q6_Vsf_equals_Vqf32(v4);
hvx_vmemu(dst + he + i * 32) = Q6_Vsf_equals_Vqf32(v5);
}
for (uint32_t i = nvec * VLEN_FP32; i < ne; i += 2) {
const float cos_theta = theta_cache[i+0];
const float sin_theta = theta_cache[i+1];
float x0 = src0[i/2];
float x1 = src0[i/2 + he];
dst[i/2] = x0 * cos_theta - x1 * sin_theta;
dst[i/2 + he] = x0 * sin_theta + x1 * cos_theta;
if (nloe > 0) {
HVX_Vector v0 = hvx_vmemu(src0 + nvec * 32);
HVX_Vector v1 = hvx_vmemu(src0 + he + nvec * 32);
HVX_Vector v2 = ((const HVX_Vector *) theta_cache)[nvec * 2 + 0];
HVX_Vector v3 = ((const HVX_Vector *) theta_cache)[nvec * 2 + 1];
HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);
HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_lo_W(vcos_sin));
HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(v0, Q6_V_hi_W(vcos_sin));
HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(v1, Q6_V_lo_W(vcos_sin));
HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(v1, Q6_V_hi_W(vcos_sin));
HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
hvx_vec_store_u(dst + nvec * 32, nloe * sizeof(float), Q6_Vsf_equals_Vqf32(v4));
hvx_vec_store_u(dst + he + nvec * 32, nloe * sizeof(float), Q6_Vsf_equals_Vqf32(v5));
}
}
static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict src0, uint32_t ne, const float * restrict theta_cache) {
const HVX_Vector * restrict vsrc = (const HVX_Vector *) src0;
const HVX_Vector * restrict vtheta = (const HVX_Vector *) theta_cache;
HVX_Vector * restrict vdst = (HVX_Vector *) dst;
const uint32_t nvec = ne / 64;
const uint32_t nloe = ne % 64;
uint32_t nvec = (ne / (VLEN_FP32 * 2)) * 2; // 2 vecs per loop, step of two
for (uint32_t i = 0; i < nvec; i++) {
HVX_Vector v0 = ((const HVX_Vector *) src0)[i * 2 + 0];
HVX_Vector v1 = ((const HVX_Vector *) src0)[i * 2 + 1];
#pragma unroll(2)
for (uint32_t i = 0; i < nvec; i+=2) {
HVX_Vector v0 = vsrc[i+0];
HVX_Vector v1 = vsrc[i+1];
HVX_Vector v2 = ((const HVX_Vector *) theta_cache)[i * 2 + 0];
HVX_Vector v3 = ((const HVX_Vector *) theta_cache)[i * 2 + 1];
HVX_Vector v2 = vtheta[i+0];
HVX_Vector v3 = vtheta[i+1];
HVX_VectorPair vx0_x1 = Q6_W_vdeal_VVR(v1, v0, -4); // vx0_x1[0] = x0, vx0_x1[1] = x1
HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4); // vcos_sin[0] = cos_theta, vcos_sin[1] = sin_theta
HVX_VectorPair vx0_x1 = Q6_W_vdeal_VVR(v1, v0, -4);
HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);
HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_lo_W(vcos_sin));
HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_hi_W(vcos_sin));
@@ -264,17 +341,52 @@ static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict
HVX_VectorPair vstore = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(v5), Q6_Vsf_equals_Vqf32(v4), -4);
vdst[i+0] = Q6_V_lo_W(vstore);
vdst[i+1] = Q6_V_hi_W(vstore);
((HVX_Vector *) dst)[i * 2 + 0] = Q6_V_lo_W(vstore);
((HVX_Vector *) dst)[i * 2 + 1] = Q6_V_hi_W(vstore);
}
for (uint32_t i = nvec * VLEN_FP32; i < ne; i += 2) {
const float cos_theta = theta_cache[i+0];
const float sin_theta = theta_cache[i+1];
float x0 = src0[i+0];
float x1 = src0[i+1];
dst[i+0] = x0 * cos_theta - x1 * sin_theta;
dst[i+1] = x0 * sin_theta + x1 * cos_theta;
if (nloe > 0) {
if (nloe <= 32) {
HVX_Vector v0 = hvx_vmemu(src0 + nvec * 64);
HVX_Vector v2 = hvx_vmemu(theta_cache + nvec * 64);
HVX_VectorPair vx0_x1 = Q6_W_vdeal_VVR(Q6_V_vzero(), v0, -4);
HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(Q6_V_vzero(), v2, -4);
HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_lo_W(vcos_sin));
HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_hi_W(vcos_sin));
HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_lo_W(vcos_sin));
HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_hi_W(vcos_sin));
HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
HVX_VectorPair vstore = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(v5), Q6_Vsf_equals_Vqf32(v4), -4);
hvx_vec_store_u(dst + nvec * 64, nloe * sizeof(float), Q6_V_lo_W(vstore));
} else {
HVX_Vector v0 = hvx_vmemu(src0 + nvec * 64);
HVX_Vector v1 = hvx_vmemu(src0 + nvec * 64 + 32);
HVX_Vector v2 = hvx_vmemu(theta_cache + nvec * 64);
HVX_Vector v3 = hvx_vmemu(theta_cache + nvec * 64 + 32);
HVX_VectorPair vx0_x1 = Q6_W_vdeal_VVR(v1, v0, -4);
HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);
HVX_Vector vx0_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_lo_W(vcos_sin));
HVX_Vector vx0_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(vx0_x1), Q6_V_hi_W(vcos_sin));
HVX_Vector vx1_c = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_lo_W(vcos_sin));
HVX_Vector vx1_s = Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(vx0_x1), Q6_V_hi_W(vcos_sin));
HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
HVX_VectorPair vstore = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(v5), Q6_Vsf_equals_Vqf32(v4), -4);
((HVX_Vector *) dst)[nvec * 2 + 0] = Q6_V_lo_W(vstore);
hvx_vec_store_u(dst + nvec * 64 + 32, (nloe - 32) * sizeof(float), Q6_V_hi_W(vstore));
}
}
}
@@ -348,13 +460,19 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
const int32_t * pos = (const int32_t *) src1->data;
const float * freq_factors = src2 ? (const float *) src2->data : NULL;
uint32_t ir = 0;
const uint32_t i3_start = fastdiv(src0_start_row, &rctx->div_ne2_ne1);
const uint32_t rem = fastmodulo(src0_start_row, ne2 * ne1, &rctx->div_ne2_ne1);
const uint32_t i2_start = fastdiv(rem, &rctx->div_ne1);
const uint32_t i1_start = fastmodulo(rem, ne1, &rctx->div_ne1);
uint32_t ir = src0_start_row;
uint32_t prev_i2 = (uint32_t) -1;
for (uint32_t i3 = 0; i3 < ne3; i3++) { // batch
for (uint32_t i2 = 0; i2 < ne2; i2++) { // seq-len
for (uint32_t i1 = 0; i1 < ne1; ) { // attn-heads
if (ir < src0_start_row) { ir++; i1++; continue; }
for (uint32_t i3 = i3_start; i3 < ne3; i3++) { // batch
const uint32_t i2_init = (i3 == i3_start) ? i2_start : 0;
for (uint32_t i2 = i2_init; i2 < ne2; i2++) { // seq-len
const uint32_t i1_init = (i3 == i3_start && i2 == i2_start) ? i1_start : 0;
for (uint32_t i1 = i1_init; i1 < ne1; ) { // attn-heads
if (ir >= src0_end_row) goto done;
// Rows in this block
@@ -407,9 +525,6 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
ne0, rctx->ext_factor, rctx->attn_factor,
theta_cache, rctx->theta_scale);
}
// FARF(HIGH, "rope-theta %u: ir %u i1 %u i2 %u i3 %u cache %p : usec %u", ith, ir, i1, i2, i3, theta_cache,
// (unsigned) HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - rctx->t_start));
}
// Skip output DMA transactions from prev block (if any)
@@ -489,7 +604,7 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
// Aligned row sizes for VTCM
const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
const size_t dst_row_size_aligned = hex_round_up(dst_row_size, VLEN);
const size_t theta_cache_size_aligned = hex_round_up(src0->ne[0] * sizeof(float), 128);
const size_t theta_cache_size_aligned = hex_round_up(src0->ne[0] * sizeof(float), 256);
// Calculate spad sizes per thread
size_t src0_spad_per_thread = theta_cache_size_aligned + HTP_ROPE_SPAD_NROWS * src0_row_size_aligned;
@@ -546,6 +661,11 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
rctx.src0_nrows = src0_nrows;
rctx.src0_nrows_per_thread = (src0_nrows + n_threads - 1) / n_threads;
if (src0_nrows > 0) {
rctx.div_ne2_ne1 = init_fastdiv_values(dst->ne[2] * dst->ne[1]);
rctx.div_ne1 = init_fastdiv_values(dst->ne[1]);
}
FARF(HIGH, "rope-f32 n-rows %u n-dims %d ne0 %u ext-factor %.6f theta-scale %.6f attn-factor %.6f\n", rctx.src0_nrows, rctx.n_dims, ne0,
rctx.ext_factor, rctx.theta_scale, rctx.attn_factor);