ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Vulkan Optimizations and Fixes (#8959)

Browse source 
* Optimize Vulkan REPEAT performance

* Use Vulkan GLSL fused multiply-add instruction where possible

* Add GGML_VULKAN_PERF option to output performance data per operator

* Rework and fix Vulkan descriptor set and descriptor pool handling

* Fix float32 concat f16 shader validation error

* Add Vulkan GROUP_NORM eps parameter

* Fix validation error with transfer queue memory barrier flags

* Remove trailing whitespaces
This commit is contained in:
0cc4m 2024-08-14 18:32:53 +02:00 • committed by GitHub
parent 98a532d474
commit 5fd89a70ea
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
16 changed files with 781 additions and 851 deletions
Download patch file Download diff file Expand all files Collapse all files

43
ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp
View file

@ -67,17 +67,17 @@ void main() {
const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] >> 4);
const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] >> 4);
const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1 + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * q4_3);
const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_5 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7);
const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx]) * q4_8 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_9 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * q4_10 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * q4_11);
const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_12 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_13 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * q4_14 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15);
const FLOAT_TYPE smin = FLOAT_TYPE(
FLOAT_TYPE(data_b[b_offset + y1_idx ]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx ]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * sc7
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7
);
tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin);
const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx]), q4_0, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), q4_1, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]), q4_2, FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * q4_3)));
const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_4, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), q4_5, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), q4_6, FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7)));
const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx]), q4_8, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), q4_9, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]), q4_10, FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * q4_11)));
const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_12, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), q4_13, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), q4_14, FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), sc7,
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), sc7,
fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 3]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 35]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 3]), sc6, FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7)))))))))))))));
const uint tmp_idx = 16 * ix + tid;
tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
#else
const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf);
const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf);
@ -88,16 +88,19 @@ void main() {
const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4);
const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4);
const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx ]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1);
const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx ]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_5);
const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
const FLOAT_TYPE smin = FLOAT_TYPE(
FLOAT_TYPE(data_b[b_offset + y1_idx]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7
+ FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7
);
const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]), q4_0, FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1);
const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_2, FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]), q4_4, FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_5);
const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(data_b[b_offset + y1_idx ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
+ fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7)))))));
tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) + sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) +
sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
const uint tmp_idx = 16 * ix + tid;
tmp[tmp_idx] = fma(dall, (fma(sx, FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f), fma(sy, FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f),
fma(sz, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)), fma(sw, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))))))), fma(-dmin, smin, tmp[tmp_idx]));
#endif
}