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

vulkan: further optimize mul_mat_vec using larger loads (#10387)

Browse source 
* vulkan: Use pipeline_robustness to disable robustness in mul_mat_vec.

Add some early returns for nonexistent rows in mul_mat_vec shaders. These
can only be hit when dispatching a 2D grid of workgroups. Fix the logic
for the 2D grid of workgroups to round up.

Enable the pipeline robustness extension if it's available, and use it to
disable robustness for these pipelines. The instructions to do the bounds
checking contend for the same ALU resources as the bit twiddling dequant
instructions.

* vulkan: Add GLSL structure aliases for quant types to allow larger loads

In Vulkan it's not possible to cast pointer types, so instead you have to
declare an aliased binding for the memory with a different type. This
commit adds aliases for the quant formats using 16b ints, and in a few
places where the struct size is a multiple of 4 also using 32b ints.
Currently only q4_k's aliases are used, but others will be used in
subsequent commits.

* vulkan: use larger loads in q5_k and q6_k shaders.

Similar to the optimization I did in q4_k recently, this vectorizes some loads
and reduces the number of bit twiddling instructions.

* vulkan: use larger K step per iteration in mul_mat_vec.

Add vec4 dequantization functions, and use them to do K=8 per iteration in
mul_mat_vec. This uses 16b loads for the quant values and 128b loads for B
which helps reduce the load on the memory system.

The K_PER_ITER==2 logic is still there, just for F16/F32, and really only
because they support unaligned sizes.

Tweak the num_iters/unrolling logic to be simpler and catch a couple missed
unrolling opportunities.
This commit is contained in:
Jeff Bolz 2024-11-20 01:11:00 -06:00 • committed by GitHub
parent ad21c9e1f1
commit 1bacb9f625
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
11 changed files with 457 additions and 147 deletions
Download patch file Download diff file Expand all files Collapse all files

69
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
View file

@ -1,5 +1,7 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types : require
#include "mul_mat_vec_base.comp"
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
@ -9,6 +11,10 @@ shared FLOAT_TYPE tmp[32];
void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
if (row >= p.stride_d) {
return;
}
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
@ -36,41 +42,66 @@ void main() {
const uint s_offset = 8*v_im + is;
const uint y_offset = 128*v_im + l0;
tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const uint y_idx = i * QUANT_K + y_offset;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
#if K_QUANTS_PER_ITERATION == 1
const uint tmp_idx = 16 * ix + tid;
tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx + 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x03) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x0c) << 2)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 0] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0x30) >> 0)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 0] & 0xc0) >> 2)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx]))))))));
#else
FLOAT_TYPE scales[4];
scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
uint32_t ql0_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
uint32_t q0_u32 = ql0_u32_lo4 | qh0_u32;
uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
uint32_t q2_u32 = ql0_u32_hi4 | qh4_u32;
uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
uvec4 q0 = uvec4(unpack8(q0_u32));
uvec4 q1 = uvec4(unpack8(q1_u32));
uvec4 q2 = uvec4(unpack8(q2_u32));
uvec4 q3 = uvec4(unpack8(q3_u32));
B_TYPE_VEC4 by0 = data_b_v4[(b_offset + y_idx) / 4];
B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 4; ++l) {
sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32),
fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum))));
sum = fma(FLOAT_TYPE(by0[l]) * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
}
tmp[16 * ix + tid] += sum;
#endif
temp += sum * d;
}
tmp[gl_LocalInvocationID.x] = temp;
// sum up partial sums and write back result
barrier();
[[unroll]] for (uint s = 16; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
}
barrier();
}
if (tid == 0) {