Skip to content

Vulkan: Add Integer Dot Product mul_mat_vec shader for legacy quants #14903

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 1 commit into
base: master
Choose a base branch
from
+310 −14
Draft

Vulkan: Add Integer Dot Product mul_mat_vec shader for legacy quants #14903

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
86 changes: 73 additions & 13 deletions ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -344,6 +344,7 @@ struct vk_device_struct {
bool float_controls_rte_fp16;
bool subgroup_add;
bool subgroup_shuffle;
bool subgroup_clustered;

bool integer_dot_product;

Expand Down Expand Up @@ -409,6 +410,8 @@ struct vk_device_struct {
vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];

vk_pipeline pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];

vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio];
vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
Expand Down Expand Up @@ -2752,6 +2755,24 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_s_f16_f32_len, mul_mat_vec_iq3_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_xs_f16_f32_len, mul_mat_vec_iq4_xs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);

#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
if (device->integer_dot_product) {
if (device->subgroup_clustered && device->vendor_id != VK_VENDOR_ID_INTEL) {
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_q8_1_f32_subgroup_len, mul_mat_vec_q4_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_q8_1_f32_subgroup_len, mul_mat_vec_q4_1_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_q8_1_f32_subgroup_len, mul_mat_vec_q5_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_q8_1_f32_subgroup_len, mul_mat_vec_q5_1_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_q8_1_f32_subgroup_len, mul_mat_vec_q8_0_q8_1_f32_subgroup_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true, true);
} else {
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_q8_1_f32_len, mul_mat_vec_q4_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_q8_1_f32_len, mul_mat_vec_q4_1_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_q8_1_f32_len, mul_mat_vec_q5_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_q8_1_f32_len, mul_mat_vec_q5_1_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_q8_1_f32_len, mul_mat_vec_q8_0_q8_1_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
}
}
#endif // GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT
}

ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
Expand Down Expand Up @@ -3275,9 +3296,10 @@ static vk_device ggml_vk_get_device(size_t idx) {

device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);

device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
device->subgroup_clustered = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eClustered);

const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;

Expand Down Expand Up @@ -4236,9 +4258,22 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte

static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols) {
VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16 || b_type == GGML_TYPE_Q8_1);
GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols);

if (b_type == GGML_TYPE_Q8_1) {
switch (a_type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
return ctx->device->pipeline_dequant_mul_mat_vec_q8_1_f32[a_type][num_cols-1];
default:
return nullptr;
}
}

switch (a_type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
Expand Down Expand Up @@ -4325,7 +4360,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
}

static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec_id()");
GGML_ASSERT(b_type == GGML_TYPE_F32);

switch (a_type) {
Expand Down Expand Up @@ -5507,12 +5542,6 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&

const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;

const bool qx_needs_dequant = x_non_contig;
const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;

// Not implemented
GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT

const uint64_t x_ne = ne01 * ne00;
const uint64_t y_ne = ne11 * ne10;
const uint64_t d_ne = ne11 * ne01;
Expand All @@ -5533,7 +5562,29 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
} else {
to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
}
vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11);

bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0;

// Check for mmq first
vk_pipeline dmmv = quantize_y ? ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, GGML_TYPE_Q8_1, ne11) : nullptr;
vk_pipeline to_q8_1 = nullptr;

if (dmmv == nullptr) {
// Fall back to f16 dequant mul mat
dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11);
quantize_y = false;
}

if (quantize_y) {
to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
}

const bool qx_needs_dequant = x_non_contig;
const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig);

// Not implemented
GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT

GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT
GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT
GGML_ASSERT(dmmv != nullptr);
Expand All @@ -5549,7 +5600,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
ctx->prealloc_size_x = x_sz_upd;
}
if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
ctx->prealloc_size_y = y_sz_upd;
}

Expand All @@ -5560,6 +5611,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
if (qy_needs_dequant) {
ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
}
if (quantize_y) {
ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
}
ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
return;
}
Expand Down Expand Up @@ -5590,6 +5644,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
}
if (qy_needs_dequant) {
d_Y = ctx->prealloc_y;
} else if (quantize_y) {
d_Y = ctx->prealloc_y;
GGML_ASSERT(d_Y->size >= y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1));
} else {
d_Y = d_Qy;
y_buf_offset = qy_buf_offset;
Expand All @@ -5604,6 +5661,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
}
if (quantize_y) {
ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13);
}

// For batch_n, the A matrix is the same for each batch, and B/D use the row stride as the batch stride
uint32_t stride_batch_x = batch_n ? 0 : ne00*ne01;
Expand Down Expand Up @@ -11285,7 +11345,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
} else if (tensor->op == GGML_OP_CONCAT) {
tensor_clone = ggml_concat(ggml_ctx, src_clone[0], src_clone[1], *(int *)tensor->op_params);
} else if (tensor->op == GGML_OP_UPSCALE) {
tensor_clone = ggml_upscale_ext(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], (ggml_scale_mode) tensor->op_params[0]);
tensor_clone = ggml_interpolate(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], (ggml_scale_mode) tensor->op_params[0]);
} else if (tensor->op == GGML_OP_SCALE) {
const float * params = (const float *)tensor->op_params;
tensor_clone = ggml_scale(ggml_ctx, src_clone[0], params[0]);
Expand Down Expand Up @@ -11400,7 +11460,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]);
}
} else if (tensor->op == GGML_OP_SET_ROWS) {
tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1]);
tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
} else if (tensor->op == GGML_OP_CONT) {
tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
} else if (tensor->op == GGML_OP_RESHAPE) {
Expand Down
11 changes: 11 additions & 0 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,10 +8,19 @@

#include "types.comp"

#ifndef MMQ
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
#else
layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
#endif

layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
#ifdef B_TYPE_VEC2
layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
#endif
#ifdef B_TYPE_VEC4
layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
#endif

layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
#ifdef MUL_MAT_ID
Expand Down Expand Up @@ -88,6 +97,7 @@ 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;

#if !defined(MMQ) || !defined(USE_SUBGROUPS)
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) {
Expand Down Expand Up @@ -116,3 +126,4 @@ void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32
}
}
}
#endif
Loading
Loading