public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-09 15:55:37 +02:00
Code Issues Packages Projects Releases Wiki Activity
Files
982533fc0c38dabc6f7fa9155b7e33e5f565e223
whisper.cpp/ggml/src/ggml-vulkan/vulkan-shaders
T
History
Matt Corallo 982533fc0c vulkan: Block-load Q3_K/Q6_K block data and subtract on 32b ints (llama/23056) 
Q2_K/Q3_K/Q6_K do much better when using MMVQ on Intel BMG even
though they're only 2-byte aligned, and Q3_K still wins on
NVIDIA as well.

mesa isn't all that great at coalescing back-to-back loads from
alternating arrays, so we force it instead. Further, we can do
subtraction directly on a full int32_t rather than an i8vec4
with bit twiddling because the high bit is always free to start.

On Intel BMG on mesa, the switch to MMVQ provides an immediate
~57% perf increase in tg128 for unsloth/Qwen3.5-9B-GGUF:Q3_K and
~78% perf increase in tg128 for unsloth/Qwen3.5-9B-GGUF:Q6_K.

The futher switch to block loads leads to a ~24% perf increase in
tg128 for unsloth/Qwen3.5-9B-GGUF:Q3_K and a ~48% perf increase in
tg128 for unsloth/Qwen3.5-9B-GGUF:Q6_K.

Finally, Xe2 wins on MMVQ even for small k, so we take the NVIDIA
override for K quants on Xe2 as well.
2026-06-08 14:36:36 +03:00
..
feature-tests
vulkan: use GL_NV_cooperative_matrix_decode_vector for faster matmul (llama/23541)
2026-05-29 09:47:30 +03:00
abs.comp
vulkan: implement ABS and NEG (llama/17245)
2025-11-17 21:05:46 +02:00
acc.comp
vulkan: support GGML_OP_SET (llama/19584)
2026-02-15 21:44:37 +02:00
add1.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
add_id.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
add.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
arange.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
argmax.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
argsort_large.comp
vulkan: support larger argsort (llama/17313)
2025-12-12 17:53:04 +02:00
argsort.comp
vulkan: support larger argsort (llama/17313)
2025-12-12 17:53:04 +02:00
ceil.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
clamp.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
CMakeLists.txt
vulkan: use GL_NV_cooperative_matrix_decode_vector for faster matmul (llama/23541)
2026-05-29 09:47:30 +03:00
concat.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
contig_copy.comp
vulkan: add cpy bf16 -> f32 pipelines (llama/22677)
2026-05-25 12:26:07 +03:00
conv2d_dw.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
conv2d_mm.comp
vulkan: optimize conv2d and implement coopmat1 support (llama/22620)
2026-05-29 09:47:30 +03:00
conv_transpose_1d.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
copy_from_quant.comp
vulkan: Support GGML_TYPE_NVFP4 (llama/21455)
2026-04-30 11:29:07 +03:00
copy_to_quant.comp
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
copy_transpose.comp
vulkan: Add copy_transpose shader (llama/17371)
2025-12-12 17:53:04 +02:00
copy.comp
vulkan: add cpy bf16 -> f32 pipelines (llama/22677)
2026-05-25 12:26:07 +03:00
cos.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
count_equal.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
count_experts.comp
vulkan: preprocess mul_mat_id experts and discard workgroups more quickly (llama/18352)
2025-12-31 17:52:09 +02:00
cumsum_multipass1.comp
vulkan: Optimize GGML_OP_CUMSUM (llama/18417)
2026-01-14 09:11:59 +02:00
cumsum_multipass2.comp
vulkan: Optimize GGML_OP_CUMSUM (llama/18417)
2026-01-14 09:11:59 +02:00
cumsum.comp
vulkan: Optimize GGML_OP_CUMSUM (llama/18417)
2026-01-14 09:11:59 +02:00
dequant_f32.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_funcs_cm2.glsl
vulkan: use GL_NV_cooperative_matrix_decode_vector for faster matmul (llama/23541)
2026-05-29 09:47:30 +03:00
dequant_funcs.glsl
vulkan: Switch MUL_MAT_VEC to 4 K per iteration for F16/32 (llama/22887)
2026-05-29 09:47:30 +03:00
dequant_head.glsl
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq1_m.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq1_s.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq2_s.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq2_xs.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq2_xxs.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq3_s.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq3_xxs.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq4_nl.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_iq4_xs.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_mxfp4.comp
Vulkan MMQ Integer Dot Refactor and K-Quant support (llama/16536)
2025-11-09 23:38:03 +02:00
dequant_nvfp4.comp
vulkan: Support GGML_TYPE_NVFP4 (llama/21455)
2026-04-30 11:29:07 +03:00
dequant_q1_0.comp
vulkan: Support Q1_0 (llama/21539)
2026-04-30 11:29:05 +03:00
dequant_q2_k.comp
Vulkan MMQ Integer Dot Refactor and K-Quant support (llama/16536)
2025-11-09 23:38:03 +02:00
dequant_q3_k.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_q4_0.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_q4_1.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_q4_k.comp
Vulkan MMQ Integer Dot Refactor and K-Quant support (llama/16536)
2025-11-09 23:38:03 +02:00
dequant_q5_0.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_q5_1.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_q5_k.comp
Vulkan MMQ Integer Dot Refactor and K-Quant support (llama/16536)
2025-11-09 23:38:03 +02:00
dequant_q6_k.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
dequant_q8_0.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
diag_mask_inf.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
diag.comp
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
div.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
elu.comp
ggml-vulkan: Add ELU op support (llama/20183)
2026-03-16 13:10:15 +02:00
exp.comp
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
fill.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
flash_attn_base.glsl
vulkan: add Flash Attention support for BFloat16 KV cache (llama/23420)
2026-06-08 14:36:36 +03:00
flash_attn_cm1.comp
vulkan: add Flash Attention support for BFloat16 KV cache (llama/23420)
2026-06-08 14:36:36 +03:00
flash_attn_cm2.comp
vulkan: add Flash Attention support for BFloat16 KV cache (llama/23420)
2026-06-08 14:36:36 +03:00
flash_attn_dequant.glsl
vulkan: add Flash Attention support for BFloat16 KV cache (llama/23420)
2026-06-08 14:36:36 +03:00
flash_attn_mask_opt.comp
vulkan: fix OOB check in flash_attn_mask_opt (llama/20296)
2026-03-16 13:10:15 +02:00
flash_attn_mmq_funcs.glsl
vulkan: Support asymmetric FA in scalar/mmq/coopmat1 paths (llama/22589)
2026-05-14 21:26:48 +03:00
flash_attn_split_k_reduce.comp
vulkan: support flash attention GQA/split_k with small batches (llama/18938)
2026-01-30 15:56:40 +02:00
flash_attn.comp
vulkan: Support asymmetric FA in scalar/mmq/coopmat1 paths (llama/22589)
2026-05-14 21:26:48 +03:00
floor.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
fwht.comp
vulkan: fast path for walsh-hadamard transform (llama/23687)
2026-05-29 09:47:30 +03:00
gated_delta_net.comp
llama + spec: MTP Support (llama/22673)
2026-05-25 12:26:07 +03:00
geglu_erf.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
geglu_quick.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
geglu.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
gelu_erf.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
gelu_quick.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
gelu.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
generic_binary_head.glsl
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
generic_head.glsl
vulkan: support GGML_UNARY_OP_XIELU (llama/18062)
2025-12-31 17:52:09 +02:00
generic_unary_head.glsl
Vulkan: MMVQ Integer Dot K-Quant and MUL_MAT_ID support (llama/16900)
2025-12-12 17:53:12 +02:00
get_rows_quant.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
get_rows.comp
vulkan: support get_rows for i32 (llama/17941)
2025-12-18 08:20:56 +02:00
glu_head.glsl
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
glu_main.glsl
vulkan: add noncontiguous GLU support (llama/21081)
2026-03-29 15:04:36 +03:00
group_norm.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
hardsigmoid.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
hardswish.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
im2col_3d.comp
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
im2col.comp
vulkan: optimize operations in the IM2COL shader (llama/22685)
2026-05-25 12:26:07 +03:00
l2_norm.comp
vulkan: fix l2_norm epsilon handling (llama/20350)
2026-03-16 13:10:15 +02:00
leaky_relu.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
log.comp
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
mul_mat_split_k_reduce.comp
vulkan: optimize and reenable split_k (llama/10637)
2024-12-08 20:14:35 +02:00
mul_mat_vec_base.glsl
vulkan: split mul_mat into multiple dispatches to avoid overflow (llama/19509)
2026-02-27 20:57:58 +02:00
mul_mat_vec_iface.glsl
vulkan: unify type macros to use Vx instead of _VECx (llama/21605)
2026-04-30 11:29:04 +03:00
mul_mat_vec_iq1_m.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec_iq1_s.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec_iq2_s.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec_iq2_xs.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec_iq2_xxs.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec_iq3_s.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec_iq3_xxs.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec_nc.comp
vulkan: Fuse mul_mat_id+add_id+mul and mul_mat+add+add. (llama/17287)
2025-11-17 21:05:46 +02:00
mul_mat_vec_p021.comp
vulkan: Fuse mul_mat_id+add_id+mul and mul_mat+add+add. (llama/17287)
2025-11-17 21:05:46 +02:00
mul_mat_vec_q2_k.comp
vulkan: unify type macros to use Vx instead of _VECx (llama/21605)
2026-04-30 11:29:04 +03:00
mul_mat_vec_q3_k.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec_q4_k.comp
vulkan: unify type macros to use Vx instead of _VECx (llama/21605)
2026-04-30 11:29:04 +03:00
mul_mat_vec_q5_k.comp
vulkan: unify type macros to use Vx instead of _VECx (llama/21605)
2026-04-30 11:29:04 +03:00
mul_mat_vec_q6_k.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mat_vec.comp
vulkan: Switch MUL_MAT_VEC to 4 K per iteration for F16/32 (llama/22887)
2026-05-29 09:47:30 +03:00
mul_mat_vecq_funcs.glsl
vulkan: Block-load Q3_K/Q6_K block data and subtract on 32b ints (llama/23056)
2026-06-08 14:36:36 +03:00
mul_mat_vecq.comp
vulkan: Use VK_EXT_shader_64bit_indexing to handle large mat_mul(_id) (llama/18678)
2026-01-14 09:11:59 +02:00
mul_mm_cm2.comp
vulkan: use GL_NV_cooperative_matrix_decode_vector for faster matmul (llama/23541)
2026-05-29 09:47:30 +03:00
mul_mm_funcs.glsl
vulkan: Coalesce Q4_K/Q5_K scale loads (llama/21751)
2026-04-30 11:29:22 +03:00
mul_mm_id_funcs.glsl
vulkan: fix data race in mul_mat_id shader (llama/19790)
2026-02-27 20:57:58 +02:00
mul_mm.comp
vulkan: unify type macros to use Vx instead of _VECx (llama/21605)
2026-04-30 11:29:04 +03:00
mul_mmq_funcs.glsl
vulkan: unify type macros to use Vx instead of _VECx (llama/21605)
2026-04-30 11:29:04 +03:00
mul_mmq_shmem_types.glsl
vulkan: Support asymmetric FA in scalar/mmq/coopmat1 paths (llama/22589)
2026-05-14 21:26:48 +03:00
mul_mmq.comp
vulkan: fix MMQ shader push constants and multi-dispatch (llama/19732)
2026-02-27 20:57:58 +02:00
mul.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
multi_add.comp
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
neg.comp
vulkan: implement ABS and NEG (llama/17245)
2025-11-17 21:05:46 +02:00
norm.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
opt_step_adamw.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
opt_step_sgd.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
pad.comp
ggml : add circular tiling support to pad, for Vulkan, CUDA, and CPU (used for making seamless textures) (llama/16985)
2025-12-12 17:53:20 +02:00
pool2d.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
quantize_q8_1.comp
vulkan: handle quantize_q8_1 overflowing the max workgroup count (llama/18515)
2026-01-14 09:11:59 +02:00
reglu.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
relu.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
repeat_back.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
repeat.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
rms_norm_back.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
rms_norm_partials.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
rms_norm.comp
vulkan: fix non-contig rope (llama/19299)
2026-02-08 09:29:10 +02:00
roll.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
rope_funcs.glsl
vulkan: Support unaligned tensors for ROPE (llama/22637)
2026-05-25 12:26:07 +03:00
rope_head.glsl
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
rope_multi.comp
vulkan: fix non-contig rope (llama/19299)
2026-02-08 09:29:10 +02:00
rope_neox.comp
vulkan: fix non-contig rope (llama/19299)
2026-02-08 09:29:10 +02:00
rope_norm.comp
vulkan: fix non-contig rope (llama/19299)
2026-02-08 09:29:10 +02:00
rope_params.glsl
vulkan: Support unaligned tensors for ROPE (llama/22637)
2026-05-25 12:26:07 +03:00
rope_vision.comp
vulkan: fix non-contig rope (llama/19299)
2026-02-08 09:29:10 +02:00
round.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
scale.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
sgn.comp
ggml-vulkan: add SGN operator, auto-generate Vulkan.csv and ops.md (llama/20219)
2026-03-16 13:10:15 +02:00
sigmoid.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
silu_back.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
silu.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
sin.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
snake.comp
vulkan: fuse snake activation (mul, sin, sqr, mul, add) (llama/22855)
2026-05-25 12:26:07 +03:00
soft_max_back.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
soft_max_large1.comp
vulkan: Multi-pass softmax for large number of cols (llama/17892)
2025-12-18 08:20:56 +02:00
soft_max_large2.comp
vulkan: Multi-pass softmax for large number of cols (llama/17892)
2025-12-18 08:20:56 +02:00
soft_max_large3.comp
vulkan: Multi-pass softmax for large number of cols (llama/17892)
2025-12-18 08:20:56 +02:00
soft_max_large_common.glsl
vulkan: Multi-pass softmax for large number of cols (llama/17892)
2025-12-18 08:20:56 +02:00
soft_max.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
softplus.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
solve_tri.comp
vulkan: support solve_tri with larger N/K values (llama/17781)
2025-12-12 17:53:20 +02:00
sqrt.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
square.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
ssm_conv.comp
vulkan: fuse SSM_CONV + BIAS + SILU (llama/22653)
2026-05-25 12:26:07 +03:00
ssm_scan.comp
vulkan: optimize ssm_scan (llama/18630)
2026-01-14 09:11:59 +02:00
step.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
sub.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
sum_rows.comp
vulkan: Implement GGML_OP_CUMSUM (llama/17479)
2025-12-12 17:53:08 +02:00
sum_rows.glsl
vulkan: Implement GGML_OP_CUMSUM (llama/17479)
2025-12-12 17:53:08 +02:00
swiglu_oai.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
swiglu.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
tanh.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
timestep_embedding.comp
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
topk_argsort.comp
vulkan: Reduce temporary memory usage for TOP_K (llama/17623)
2025-12-12 17:53:15 +02:00
topk_moe.comp
vulkan: fix topk_moe_sigmoid_norm_bias failures in GLM-4.6 (llama/18582)
2026-01-14 09:11:59 +02:00
topk_nary_search.comp
vulkan: fix top_k bug when there are ties in the input (llama/17659)
2025-12-12 17:53:19 +02:00
tri.comp
vulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (llama/21572)
2026-04-30 11:29:08 +03:00
trunc.comp
vulkan: implement ADD1, ARANGE, FILL, SOFTPLUS, STEP, ROUND, CEIL, FLOOR, TRUNC (llama/17319)
2025-12-12 17:53:04 +02:00
types.glsl
vulkan: Switch MUL_MAT_VEC to 4 K per iteration for F16/32 (llama/22887)
2026-05-29 09:47:30 +03:00
upscale.comp
vulkan: Support UPSCALE w/antialias (llama/18327)
2025-12-31 17:52:09 +02:00
utils.glsl
vulkan : incremental shader builds (llama/16341)
2025-10-12 11:16:23 +03:00
vulkan-shaders-gen.cpp
vulkan: add Flash Attention support for BFloat16 KV cache (llama/23420)
2026-06-08 14:36:36 +03:00
wkv6.comp
rwkv6: add wkv6 support for Vulkan backend (llama/10829)
2024-12-18 12:52:16 +02:00
wkv7.comp
llama: Add support for RWKV v7 architecture (llama/12412)
2025-03-27 11:06:03 +02:00
xielu.comp
vulkan: support GGML_UNARY_OP_XIELU (llama/18062)
2025-12-31 17:52:09 +02:00