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

vulkan: optimize mul_mat for small values of N (#10991)

Browse source 
Make the mul_mat_vec shaders support N>1 (as a spec constant, NUM_COLS) where
the batch_strides are overloaded to hold the row strides. Put the loads from the
B matrix in the innermost loop because it should cache better.

Share some code for reducing the result values to memory in mul_mat_vec_base.
This commit is contained in:
Jeff Bolz 2024-12-30 11:27:11 -06:00 • committed by GitHub
parent c250ecb315
commit 716bd6dec3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 288 additions and 349 deletions
Download patch file Download diff file Expand all files Collapse all files

122
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
View file

@ -9,9 +9,6 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
#define K_PER_ITER 8
#else
@ -21,70 +18,70 @@ layout (constant_id = 1) const uint NUM_ROWS = 1;
uint a_offset, b_offset, d_offset, y_offset;
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
{
const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
const uint iybs = col - col%QUANT_K; // y block start index
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
const uint iybs = col - col%QUANT_K; // y block start index
#if K_PER_ITER == 8
#if QUANT_R == 2
const B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4];
const B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4];
const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
const B_TYPE_VEC4 bv02 = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4];
const B_TYPE_VEC4 bv13 = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs + y_offset) / 4];
const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
#else
const vec4 bv0 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4]);
const vec4 bv1 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4 + 1]);
const vec4 bv0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]);
const vec4 bv1 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4 + 1]);
#endif
#else
// Check if the second of the pair of elements is OOB, and don't fetch B or
// accumulate it. We still fetch a pair of elements for A, which is fine for
// quantized formats since they'll be within the same block. We should
// probably skip fetching the second element for F16/F32, but as of now we
// still do.
const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
// Check if the second of the pair of elements is OOB, and don't fetch B or
// accumulate it. We still fetch a pair of elements for A, which is fine for
// quantized formats since they'll be within the same block. We should
// probably skip fetching the second element for F16/F32, but as of now we
// still do.
const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
FLOAT_TYPE b0 = 0, b1 = 0;
b0 = FLOAT_TYPE(data_b[b_offset + iybs + iqs]);
if (!OOB) {
b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
}
FLOAT_TYPE b0 = 0, b1 = 0;
b0 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]);
if (!OOB) {
b1 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]);
}
#endif
uint ibi = first_row*p.ncols;
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib = (ibi + col)/QUANT_K; // block index
ibi += p.ncols;
uint ibi = first_row*p.ncols;
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib = (ibi + col)/QUANT_K; // block index
ibi += p.ncols;
#if K_PER_ITER == 8
vec4 v = dequantize4(ib, iqs, a_offset);
vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
vec4 v = dequantize4(ib, iqs, a_offset);
vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
const vec2 dm = get_dm(ib, a_offset);
if (dm.y != 0) { // quant has min component
v = v * dm.x + dm.y;
v2 = v2 * dm.x + dm.y;
}
const vec2 dm = get_dm(ib, a_offset);
if (dm.y != 0) { // quant has min component
v = v * dm.x + dm.y;
v2 = v2 * dm.x + dm.y;
}
// matrix multiplication
FLOAT_TYPE rowtmp = dot(bv0, v);
rowtmp += dot(bv1, v2);
// matrix multiplication
FLOAT_TYPE rowtmp = dot(bv0, v);
rowtmp += dot(bv1, v2);
if (dm.y == 0)
rowtmp *= dm.x;
if (dm.y == 0)
rowtmp *= dm.x;
temp[n] += rowtmp;
temp[j][n] += rowtmp;
#else
const vec2 v = dequantize(ib, iqs, a_offset);
const vec2 v = dequantize(ib, iqs, a_offset);
// matrix multiplication
temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
if (!OOB) {
temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
}
// matrix multiplication
temp[j][n] = fma(FLOAT_TYPE(v.x), b0, temp[j][n]);
if (!OOB) {
temp[j][n] = fma(FLOAT_TYPE(v.y), b1, temp[j][n]);
}
#endif
}
}
}
@ -96,10 +93,12 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
FLOAT_TYPE temp[NUM_ROWS];
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
@ -131,24 +130,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
i++;
}
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {

33
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
View file

@ -83,3 +83,36 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
batch_idx * p.batch_stride_d;
#endif
}
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
layout (constant_id = 2) const uint NUM_COLS = 1;
shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
// sum up partial sums and write back result
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[j][n][tid] = temp[j][n];
}
}
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[j][n][tid] += tmpsh[j][n][tid + s];
}
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
}
}
}
}

92
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
View file

@ -5,11 +5,6 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
@ -32,24 +27,17 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint s_offset = 8*v_im;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp[NUM_ROWS];
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
@ -74,48 +62,42 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uvec2 qs0 = uvec2(unpack8(qs0_u16));
uvec2 qs16 = uvec2(unpack8(qs16_u16));
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
}
temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
}
temp[n] = fma(dall, sum1, fma(-dmin, sum2, temp[n]));
}
}
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {

75
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
View file

@ -5,11 +5,6 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
@ -33,10 +28,12 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint q_offset = 32*v_im + l0;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp[NUM_ROWS];
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
const uint s_shift = 4 * v_im;
@ -44,15 +41,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
@ -70,39 +58,34 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
u8vec2 s8 = unpack8(s8_16);
u8vec2 s10 = unpack8(s10_16);
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
}
temp[j][n] = fma(d, sum, temp[j][n]);
}
temp[n] = fma(d, sum, temp[n]);
}
}
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {

64
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
View file

@ -6,11 +6,6 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
@ -36,21 +31,18 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint q_offset = 32*v_im + l0;
const uint y_offset = 64*v_im + l0;
FLOAT_TYPE temp[NUM_ROWS];
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
B_TYPE_VEC4 by10 = data_b_v4[(b_offset + y1_idx) / 4];
B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
B_TYPE_VEC4 by20 = data_b_v4[(b_offset + y2_idx) / 4];
B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
@ -103,37 +95,27 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint32_t q4_14 = qs64_hi4.z;
const uint32_t q4_15 = qs64_hi4.w;
const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3)));
const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7)));
const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11)));
const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7)))))))))))))));
temp[n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[n]));
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC4 by10 = data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4];
B_TYPE_VEC4 by132 = data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8];
B_TYPE_VEC4 by20 = data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4];
B_TYPE_VEC4 by232 = data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8];
const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3)));
const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7)));
const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x), q4_8, fma(FLOAT_TYPE(by20.y), q4_9, fma(FLOAT_TYPE(by20.z), q4_10, FLOAT_TYPE(by20.w) * q4_11)));
const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x), q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6, FLOAT_TYPE(by232.w) * sc7)))))))))))))));
temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
}
}
}
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {

104
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
View file

@ -6,11 +6,6 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
@ -33,25 +28,18 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint q_offset = 32*v_im + l0;
const uint y_offset = 64*v_im + l0;
FLOAT_TYPE temp[NUM_ROWS];
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y1_idx = i * QUANT_K + y_offset;
const uint y2_idx = y1_idx + 128;
B_TYPE_VEC2 by10 = data_b_v2[(b_offset + y1_idx) / 2];
B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8];
B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16];
B_TYPE_VEC2 by148 = data_b_v2[(b_offset + y1_idx) / 2 + 24];
B_TYPE_VEC2 by20 = data_b_v2[(b_offset + y2_idx) / 2];
B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8];
B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
@ -116,53 +104,47 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint32_t q4_14 = qs64_80_hi4.z;
const uint32_t q4_15 = qs64_80_hi4.w;
const FLOAT_TYPE sx =
fma(FLOAT_TYPE(by10.x), q4_0,
fma(FLOAT_TYPE(by10.y), q4_1,
fma(FLOAT_TYPE(by116.x), q4_2,
FLOAT_TYPE(by116.y) * q4_3)));
const FLOAT_TYPE sy =
fma(FLOAT_TYPE(by132.x), q4_4,
fma(FLOAT_TYPE(by132.y), q4_5,
fma(FLOAT_TYPE(by148.x), q4_6,
FLOAT_TYPE(by148.y) * q4_7)));
const FLOAT_TYPE sz =
fma(FLOAT_TYPE(by20.x), q4_8,
fma(FLOAT_TYPE(by20.y), q4_9,
fma(FLOAT_TYPE(by216.x), q4_10,
FLOAT_TYPE(by216.y) * q4_11)));
const FLOAT_TYPE sw =
fma(FLOAT_TYPE(by232.x), q4_12,
fma(FLOAT_TYPE(by232.y), q4_13,
fma(FLOAT_TYPE(by248.x), q4_14,
FLOAT_TYPE(by248.y) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
(FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
temp[n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[n]));
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC2 by10 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2];
B_TYPE_VEC2 by116 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 8];
B_TYPE_VEC2 by132 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16];
B_TYPE_VEC2 by148 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24];
B_TYPE_VEC2 by20 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2];
B_TYPE_VEC2 by216 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 8];
B_TYPE_VEC2 by232 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16];
B_TYPE_VEC2 by248 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24];
const FLOAT_TYPE sx =
fma(FLOAT_TYPE(by10.x), q4_0,
fma(FLOAT_TYPE(by10.y), q4_1,
fma(FLOAT_TYPE(by116.x), q4_2,
FLOAT_TYPE(by116.y) * q4_3)));
const FLOAT_TYPE sy =
fma(FLOAT_TYPE(by132.x), q4_4,
fma(FLOAT_TYPE(by132.y), q4_5,
fma(FLOAT_TYPE(by148.x), q4_6,
FLOAT_TYPE(by148.y) * q4_7)));
const FLOAT_TYPE sz =
fma(FLOAT_TYPE(by20.x), q4_8,
fma(FLOAT_TYPE(by20.y), q4_9,
fma(FLOAT_TYPE(by216.x), q4_10,
FLOAT_TYPE(by216.y) * q4_11)));
const FLOAT_TYPE sw =
fma(FLOAT_TYPE(by232.x), q4_12,
fma(FLOAT_TYPE(by232.y), q4_13,
fma(FLOAT_TYPE(by248.x), q4_14,
FLOAT_TYPE(by248.y) * q4_15)));
const FLOAT_TYPE smin =
fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
(FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
}
}
}
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {

58
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
View file

@ -6,11 +6,6 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
@ -36,20 +31,17 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint s_offset = 8*v_im + is;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp[NUM_ROWS];
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;
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];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
@ -84,35 +76,25 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uvec4 q2 = uvec4(unpack8(q2_u32));
uvec4 q3 = uvec4(unpack8(q3_u32));
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 4; ++l) {
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))));
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC4 by0 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4];
B_TYPE_VEC4 by32 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 8];
B_TYPE_VEC4 by64 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 16];
B_TYPE_VEC4 by96 = data_b_v4[(j*p.batch_stride_b + 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(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))));
}
temp[j][n] += sum * d;
}
temp[n] += sum * d;
}
}
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {