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

Add q4_1, q5_0, q5_1 and q8_0 dequant mat mat mul shaders

Browse source
This commit is contained in:
0cc4m 2024-03-02 12:33:10 +01:00
parent 2acb281105
commit a8eeab2d58
3 changed files with 16262 additions and 1547 deletions
Download patch file Download diff file Expand all files Collapse all files

17592
ggml-vulkan-shaders.hpp
View file

File diff suppressed because it is too large Load diff

127
ggml-vulkan.cpp
View file

@ -988,6 +988,10 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
ctx->device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>(); ctx->device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>(); ctx->device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>(); ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
if (device->fp16) { if (device->fp16) {
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_len, matmul_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_len, matmul_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
@ -1017,6 +1021,34 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_0_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_0_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_0_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
} else { } else {
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_fp32_len, matmul_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
@ -1045,6 +1077,34 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_mmq_l, l_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_mmq_m, m_align);
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_mmq_s, s_align);
} }
ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32", mul_mat_vec_f16_f32_len, mul_mat_vec_f16_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1); ggml_vk_create_pipeline(ctx, ctx->device->pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32", mul_mat_vec_f16_f32_len, mul_mat_vec_f16_f32_data, "main", 3, 3 * sizeof(uint32_t), {1, 1, 1}, { device->subgroup_size }, 1);
@ -4068,7 +4128,7 @@ static void ggml_vk_quantize_data(const float * from, void * to, size_t ne, ggml
ggml_quantize_q5_0(from, to, ne, ne, hist_cur.data()); ggml_quantize_q5_0(from, to, ne, ne, hist_cur.data());
break; break;
case GGML_TYPE_Q5_1: case GGML_TYPE_Q5_1:
ggml_quantize_q4_1(from, to, ne, ne, hist_cur.data()); ggml_quantize_q5_1(from, to, ne, ne, hist_cur.data());
break; break;
case GGML_TYPE_Q8_0: case GGML_TYPE_Q8_0:
ggml_quantize_q8_0(from, to, ne, ne, hist_cur.data()); ggml_quantize_q8_0(from, to, ne, ne, hist_cur.data());
@ -4226,7 +4286,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
float * d_chk = (float *) malloc(d_sz); float * d_chk = (float *) malloc(d_sz);
for (size_t i = 0; i < x_ne; i++) { for (size_t i = 0; i < x_ne; i++) {
x[i] = rand() / (float)RAND_MAX; x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
} }
ggml_vk_quantize_data(x, qx, x_ne, quant); ggml_vk_quantize_data(x, qx, x_ne, quant);
@ -4503,17 +4563,24 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
ggml_vk_test_transfer(ctx, 8192 * 1000, false); ggml_vk_test_transfer(ctx, 8192 * 1000, false);
ggml_vk_test_transfer(ctx, 8192 * 1000, true); ggml_vk_test_transfer(ctx, 8192 * 1000, true);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_F32); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q4_0); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q4_1); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q5_0); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_0);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q5_1); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_1);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q8_0); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q8_0);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q2_K); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q2_K);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q3_K); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q3_K);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q4_K); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_K);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q5_K); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_K);
// ggml_vk_test_dequant(ctx, 2560 * 7680, GGML_TYPE_Q6_K); ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q6_K);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 0);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 1);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 2);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 0);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 1);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 2);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_0); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_0); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_0);
@ -4521,12 +4588,34 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_0); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_0); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_0); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_0);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 0);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 1); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_1);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 2); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_1);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 0); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_1);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 1); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_1);
ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 2); ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_1);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_1);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_1);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_1);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_1);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_1);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_1);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_1);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q8_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q8_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q8_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q8_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q8_0);
ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q8_0);
std::cerr << std::endl; std::cerr << std::endl;

90
ggml_vk_generate_shaders.py
View file

@ -64,6 +64,7 @@ struct block_q5_0
#define A_TYPE block_q5_0 #define A_TYPE block_q5_0
""" """
shader_q5_1_defines = """ shader_q5_1_defines = """
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#define QUANT_K 32 #define QUANT_K 32
#define QUANT_R 2 #define QUANT_R 2
@ -71,7 +72,7 @@ struct block_q5_1
{ {
float16_t d; float16_t d;
float16_t m; float16_t m;
uint16_t qh[2]; uint qh;
uint8_t qs[16]; uint8_t qs[16];
}; };
@ -187,7 +188,7 @@ v = (v - 16.0f) * d;
shader_q5_1_dequant_func = """ shader_q5_1_dequant_func = """
#define DEQUANT_FUNC const float d = float(data_a[ib].d); \ #define DEQUANT_FUNC const float d = float(data_a[ib].d); \
const float m = float(data_a[ib].m); \ const float m = float(data_a[ib].m); \
const uint uint_qh = uint(data_a[ib].qh[1]) << 16 | uint(data_a[ib].qh[0]); \ const uint uint_qh = data_a[ib].qh; \
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); \ const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); \
const uint vui = uint(data_a[ib].qs[iqs]); \ const uint vui = uint(data_a[ib].qs[iqs]); \
vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y); \ vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y); \
@ -336,6 +337,7 @@ mulmat_load_scalar = """
mulmat_load_q4_0 = """ mulmat_load_q4_0 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
const uint ib = idx / 16; const uint ib = idx / 16;
const uint iqs = idx & 0xF; const uint iqs = idx & 0xF;
@ -343,11 +345,71 @@ mulmat_load_q4_0 = """
const float d = float(data_a[ib].d); const float d = float(data_a[ib].d);
const uint vui = uint(data_a[ib].qs[iqs]); const uint vui = uint(data_a[ib].qs[iqs]);
const vec2 v = (vec2(vui & 0xF, vui >> 4) - 8.0f) * d; const vec2 v = (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);""" buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);"""
mulmat_load_q4_1 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
const uint ib = idx / 16;
const uint iqs = idx & 0xF;
const float d = float(data_a[ib].d);
const float m = float(data_a[ib].m);
const uint vui = uint(data_a[ib].qs[iqs]);
const vec2 v = vec2(vui & 0xF, vui >> 4) * d + m;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);"""
mulmat_load_q5_0 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
const uint ib = idx / 16;
const uint iqs = idx & 0xF;
const float d = float(data_a[ib].d);
const uint uint_qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0];
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const uint vui = uint(data_a[ib].qs[iqs]);
const vec2 v = (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);"""
mulmat_load_q5_1 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
const uint ib = idx / 16;
const uint iqs = idx & 0xF;
const float d = float(data_a[ib].d);
const float m = float(data_a[ib].m);
const uint uint_qh = data_a[ib].qh;
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const uint vui = uint(data_a[ib].qs[iqs]);
const vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);"""
mulmat_load_q8_0 = """
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
const uint ib = idx / 16;
const uint iqs = (idx & 0xF) * 2;
const float d = float(data_a[ib].d);
const vec2 v = vec2(int(data_a[ib].qs[iqs]), int(data_a[ib].qs[iqs + 1])) * d;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);"""
mulmat_body2 = """ mulmat_body2 = """
} }
[[unroll]] for (uint l = 0; l < BN; l += loadstride_b) { [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
@ -614,7 +676,7 @@ void main() {
const float d = float(data_a[ib].d); const float d = float(data_a[ib].d);
const float m = float(data_a[ib].m); const float m = float(data_a[ib].m);
const uint qh = uint(data_a[ib].qh[1]) << 16 | uint(data_a[ib].qh[0]); const uint qh = data_a[ib].qh;
const uint q_idx = 8*il; const uint q_idx = 8*il;
@ -2336,6 +2398,26 @@ async def main():
tasks.append(string_to_spv("matmul_q4_0_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q4_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16)) tasks.append(string_to_spv("matmul_q4_0_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q4_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q4_0_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16)) tasks.append(string_to_spv("matmul_q4_0_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q4_1_defines, mulmat_body1, mulmat_load_q4_1, mulmat_body2))
tasks.append(string_to_spv("matmul_q4_1_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q4_1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q4_1_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q4_1", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_0_defines, mulmat_body1, mulmat_load_q5_0, mulmat_body2))
tasks.append(string_to_spv("matmul_q5_0_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q5_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q5_0_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q5_1_defines, mulmat_body1, mulmat_load_q5_1, mulmat_body2))
tasks.append(string_to_spv("matmul_q5_1_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q5_1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q5_1_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q5_1", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
stream.clear()
stream.extend((mulmat_head, shader_int8_ext, shader_float_type, shader_q8_0_defines, mulmat_body1, mulmat_load_q8_0, mulmat_body2))
tasks.append(string_to_spv("matmul_q8_0_f32", "".join(stream), {"LOAD_VEC_A": 2, "A_TYPE": "block_q8_0", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
tasks.append(string_to_spv("matmul_q8_0_f32_aligned", "".join(stream), {"LOAD_VEC_A": 2, "LOAD_VEC_B": load_vec, "A_TYPE": "block_q8_0", "B_TYPE": vec_type, "D_TYPE": "float"}, fp16))
# Shaders where precision is needed, so no fp16 version # Shaders where precision is needed, so no fp16 version
# mul mat vec # mul mat vec