#include "ggml-vulkan.h"
#include <vulkan/vulkan_core.h>
-#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_PERF)
+#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_PERF) || defined(GGML_VULKAN_CHECK_RESULTS)
#include <chrono>
+#include "ggml-cpu.h"
#endif
#include <vulkan/vulkan.hpp>
bool uma;
bool coopmat2;
+ bool coopmat_support;
+ bool coopmat_acc_f32_support;
+ bool coopmat_acc_f16_support;
+ uint32_t coopmat_m;
+ uint32_t coopmat_n;
+ uint32_t coopmat_k;
+
size_t idx;
+ bool mul_mat_l;
+ bool mul_mat_m;
+ bool mul_mat_s;
+ bool mul_mat_id_l;
+ bool mul_mat_id_m;
+ bool mul_mat_id_s;
+
vk_matmul_pipeline pipeline_matmul_f32;
vk_matmul_pipeline pipeline_matmul_f32_f16;
vk_matmul_pipeline2 pipeline_matmul_f16;
vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
vk_matmul_pipeline pipeline_matmul_id_f32;
- vk_matmul_pipeline pipeline_matmul_id_f16;
- vk_matmul_pipeline pipeline_matmul_id_f16_f32;
+ vk_matmul_pipeline2 pipeline_matmul_id_f16;
+ vk_matmul_pipeline2 pipeline_matmul_id_f16_f32;
- vk_matmul_pipeline pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
+ vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT];
return {64, 64};
};
+static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector<uint32_t>& warptile, bool mul_mat_id) {
+ // Needs to be kept up to date on shader changes
+ const uint32_t bank_conflict_offset = device->coopmat_support ? 8 : 1;
+ const uint32_t type_size = device->fp16 ? sizeof(ggml_fp16_t) : sizeof(float);
+ const uint32_t warps = warptile[0] / device->subgroup_size;
+
+ const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size;
+ const uint32_t mmid_row_ids = mul_mat_id ? 3072 * sizeof(uint32_t) : 0;
+ const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0;
+
+ return (load_bufs + mmid_row_ids + coopmat_stage) <= device->properties.limits.maxComputeSharedMemorySize;
+}
static void ggml_vk_load_shaders(vk_device& device) {
VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");
m_align = 64;
s_align = 32;
} else {
- l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
- m_warptile = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
- s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
- l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
- m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
- s_warptile_mmq = { subgroup_size_16, 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
+ // Matrix cores require different warp group sizes
+ const uint32_t tm_l = device->coopmat_support ? device->coopmat_m : 4;
+ const uint32_t tm_m = device->coopmat_support ? device->coopmat_m : 4;
+ const uint32_t tm_s = device->coopmat_support ? device->coopmat_m : 2;
+ const uint32_t tn_l = device->coopmat_support ? device->coopmat_n : 4;
+ const uint32_t tn_m = device->coopmat_support ? device->coopmat_n : 2;
+ const uint32_t tn_s = device->coopmat_support ? device->coopmat_n : 2;
+ const uint32_t tk_l = device->coopmat_support ? device->coopmat_k : 1;
+ const uint32_t tk_m = device->coopmat_support ? device->coopmat_k : 1;
+ const uint32_t tk_s = device->coopmat_support ? device->coopmat_k : 1;
+
+ l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size };
+ m_warptile = { 128, 64, 64, 16, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size };
+ s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size };
+
+ l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size };
+ m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size };
+ s_warptile_mmq = { subgroup_size_16, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size };
+
l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 };
s_mmq_wg_denoms = s_wg_denoms = { 32, 32, 1 };
// assert mul_mat_mat_id shaders will fit.
GGML_ASSERT(shmem_needed + 3072*4 <= device->properties.limits.maxComputeSharedMemorySize);
}
+ // Disable medium and large matrix multiplication if not enough shared memory is available
+ // Check mmq warptiles as the largest configuration
+ // Throw an error if not enough for any matrix multiplication is available
+ if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmq, false)) {
+ std::cerr << "ggml_vulkan: Error: Shared memory size too small for matrix multiplication." << std::endl;
+ throw std::runtime_error("Shared memory size too small for matrix multiplication.");
+ } else if (!ggml_vk_matmul_shmem_support(device, m_warptile_mmq, false)) {
+ device->mul_mat_m = false;
+ device->mul_mat_l = false;
+ } else if (!ggml_vk_matmul_shmem_support(device, l_warptile_mmq, false)) {
+ device->mul_mat_l = false;
+ }
+
+ // Disable mul_mat_id if not enough shared memory is available
+ if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmq, true)) {
+ device->mul_mat_id_s = false;
+ device->mul_mat_id_m = false;
+ device->mul_mat_id_l = false;
+ } else if (!ggml_vk_matmul_shmem_support(device, m_warptile_mmq, true)) {
+ device->mul_mat_id_m = false;
+ device->mul_mat_id_l = false;
+ } else if (!ggml_vk_matmul_shmem_support(device, l_warptile_mmq, true)) {
+ device->mul_mat_id_l = false;
+ }
}
device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>();
device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_matmul_id_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_matmul_id_f16 = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
- device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
std::vector<std::future<void>> compiles;
auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align, bool disable_robustness = false) {
CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
-
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
#undef CREATE_MM
#undef CREATE_MM2
} else
-#endif
- if (device->fp16) {
+#endif // defined(VK_NV_cooperative_matrix2)
+ if (device->coopmat_support) {
// Create 6 variants, {s,m,l}x{unaligned,aligned}
-#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \
-
- CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
-
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+ if (device->mul_mat ## ID ## _l) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _m) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _s) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _l) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \
+ if (device->mul_mat ## ID ## _m) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \
+ if (device->mul_mat ## ID ## _s) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \
+
+ // Create 2 variants, {f16,f32} accumulator
+#define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+ CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+ CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+
+ CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
// If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
- if (device->properties.limits.maxComputeSharedMemorySize >= 32768) {
- CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
- CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
- CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
-
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+ if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
+ CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+ CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+ CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ }
+#undef CREATE_MM
+ } else if (device->fp16) {
+ // Create 6 variants, {s,m,l}x{unaligned,aligned}
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+ if (device->mul_mat ## ID ## _l) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _m) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _s) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _l) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \
+ if (device->mul_mat ## ID ## _m) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \
+ if (device->mul_mat ## ID ## _s) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \
+
+ CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+ // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
+ if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
+ CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+ CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+ CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
}
#undef CREATE_MM
} else {
// Create 6 variants, {s,m,l}x{unaligned,aligned}
-#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \
- ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \
-
- CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
-
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
- CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+ if (device->mul_mat ## ID ## _l) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _m) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _s) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \
+ if (device->mul_mat ## ID ## _l) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \
+ if (device->mul_mat ## ID ## _m) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \
+ if (device->mul_mat ## ID ## _s) \
+ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \
+
+ CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+ CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
// If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
- if (device->properties.limits.maxComputeSharedMemorySize >= 32768) {
- CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
- CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
- CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
-
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
- CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+ if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
+ CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+ CREATE_MM(pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+ CREATE_MM(pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+ CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
}
+#undef CREATE_MM2
#undef CREATE_MM
}
bool fp16_compute = false;
bool maintenance4_support = false;
bool sm_builtins = false;
+ bool amd_shader_core_properties2 = false;
bool pipeline_robustness = false;
bool coopmat2_support = false;
+ device->coopmat_support = false;
// Check if maintenance4 is supported
for (const auto& properties : ext_props) {
fp16_compute = true;
} else if (strcmp("VK_NV_shader_sm_builtins", properties.extensionName) == 0) {
sm_builtins = true;
+ } else if (strcmp("VK_AMD_shader_core_properties2", properties.extensionName) == 0) {
+ amd_shader_core_properties2 = true;
} else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) {
pipeline_robustness = true;
+ } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
+ !getenv("GGML_VK_DISABLE_COOPMAT")) {
+ device->coopmat_support = true;
+ device->coopmat_m = 0;
+ device->coopmat_n = 0;
+ device->coopmat_k = 0;
} else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 &&
- !getenv("GGML_VULKAN_DISABLE_COOPMAT2")) {
+ !getenv("GGML_VK_DISABLE_COOPMAT2")) {
coopmat2_support = true;
}
}
vk::PhysicalDeviceMaintenance3Properties props3;
vk::PhysicalDeviceMaintenance4Properties props4;
vk::PhysicalDeviceSubgroupProperties subgroup_props;
+ vk::PhysicalDeviceDriverProperties driver_props;
vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props;
+ vk::PhysicalDeviceShaderCoreProperties2AMD amd_shader_core_properties2_props;
props2.pNext = &props3;
props3.pNext = &subgroup_props;
+ subgroup_props.pNext = &driver_props;
- VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&subgroup_props;
+ VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&driver_props;
if (maintenance4_support) {
last_struct->pNext = (VkBaseOutStructure *)&props4;
last_struct->pNext = (VkBaseOutStructure *)&sm_props;
last_struct = (VkBaseOutStructure *)&sm_props;
}
+ if (amd_shader_core_properties2) {
+ last_struct->pNext = (VkBaseOutStructure *)&amd_shader_core_properties2_props;
+ last_struct = (VkBaseOutStructure *)&amd_shader_core_properties2_props;
+ }
#if defined(VK_NV_cooperative_matrix2)
vk::PhysicalDeviceCooperativeMatrix2PropertiesNV coopmat2_props;
const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE");
if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) {
- device->max_memory_allocation_size = std::stoi(GGML_VK_FORCE_MAX_ALLOCATION_SIZE);
+ device->max_memory_allocation_size = std::stoul(GGML_VK_FORCE_MAX_ALLOCATION_SIZE);
} else if (maintenance4_support) {
device->max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize);
} else {
device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
if (sm_builtins) {
device->shader_core_count = sm_props.shaderSMCount;
+ } else if (amd_shader_core_properties2) {
+ device->shader_core_count = amd_shader_core_properties2_props.activeComputeUnitCount;
} else {
device->shader_core_count = 0;
}
- const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16");
- const bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr;
+ const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;
device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
+ if (device->vendor_id == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && driver_props.driverID == vk::DriverId::eAmdProprietary)) {
+ // Intel drivers don't support coopmat properly yet
+ // Only RADV supports coopmat properly on AMD
+ device->coopmat_support = false;
+ }
+
std::vector<vk::QueueFamilyProperties> queue_family_props = device->physical_device.getQueueFamilyProperties();
// Try to find a non-graphics compute queue and transfer-focused queues
device_extensions.push_back("VK_EXT_pipeline_robustness");
}
+ VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features;
+ coopmat_features.pNext = nullptr;
+ coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR;
+ coopmat_features.cooperativeMatrix = VK_FALSE;
+
+ if (device->coopmat_support) {
+ last_struct->pNext = (VkBaseOutStructure *)&coopmat_features;
+ last_struct = (VkBaseOutStructure *)&coopmat_features;
+ }
+
#if defined(VK_NV_cooperative_matrix2)
VkPhysicalDeviceCooperativeMatrix2FeaturesNV coopmat2_features {};
coopmat2_features.pNext = nullptr;
device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
+ device->coopmat_support = device->coopmat_support && coopmat_features.cooperativeMatrix;
+
if (coopmat2_support) {
#if defined(VK_NV_cooperative_matrix2)
if (coopmat2_features.cooperativeMatrixWorkgroupScope &&
if (device->fp16) {
device_extensions.push_back("VK_KHR_shader_float16_int8");
}
+
+ if (device->coopmat_support) {
+ // Query supported shapes
+ std::vector<VkCooperativeMatrixPropertiesKHR> cm_props;
+
+ PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR =
+ (PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR)vkGetInstanceProcAddr(vk_instance.instance, "vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR");
+
+ uint32_t cm_props_num;
+
+ pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR(device->physical_device, &cm_props_num, nullptr);
+
+ cm_props.resize(cm_props_num);
+
+ for (auto& prop : cm_props) {
+ prop.sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR;
+ }
+
+ pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR(device->physical_device, &cm_props_num, cm_props.data());
+
+ VK_LOG_DEBUG("ggml_vulkan: Cooperative Matrix Shapes: " << cm_props.size());
+
+ for (auto& prop : cm_props) {
+ VK_LOG_DEBUG("ggml_vulkan: M: " << prop.MSize << " N: " << prop.NSize << " K: " << prop.KSize << " A: " << vk::to_string((vk::ComponentTypeKHR)prop.AType) << " B: " << vk::to_string((vk::ComponentTypeKHR)prop.BType) << " C: " << vk::to_string((vk::ComponentTypeKHR)prop.CType) << " Result: " << vk::to_string((vk::ComponentTypeKHR)prop.ResultType) << " saturatingAccumulation: " << prop.saturatingAccumulation << " scope: " << vk::to_string((vk::ScopeKHR)prop.scope));
+
+ if ((vk::ComponentTypeKHR)prop.AType == vk::ComponentTypeKHR::eFloat16 &&
+ (vk::ComponentTypeKHR)prop.BType == vk::ComponentTypeKHR::eFloat16 &&
+ (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup
+ ) {
+ if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat32 &&
+ (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat32) {
+ // coopmat sizes not set yet
+ if (device->coopmat_m == 0) {
+ device->coopmat_acc_f32_support = true;
+ device->coopmat_m = prop.MSize;
+ device->coopmat_n = prop.NSize;
+ device->coopmat_k = prop.KSize;
+ } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) {
+ // Only enable if shape is identical
+ device->coopmat_acc_f32_support = true;
+ }
+ } else if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat16 &&
+ (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat16) {
+ // coopmat sizes not set yet
+ if (device->coopmat_m == 0) {
+ device->coopmat_acc_f16_support = true;
+ device->coopmat_m = prop.MSize;
+ device->coopmat_n = prop.NSize;
+ device->coopmat_k = prop.KSize;
+ } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) {
+ // Only enable if shape is identical
+ device->coopmat_acc_f16_support = true;
+ }
+ }
+ }
+ }
+
+ if (device->coopmat_m == 0) {
+ // No suitable matmul mode found
+ GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n");
+ device->coopmat_support = false;
+ }
+ }
+
+ if (device->coopmat_support) {
+ device_extensions.push_back("VK_KHR_cooperative_matrix");
+ }
+
device->name = GGML_VK_NAME + std::to_string(idx);
device_create_info = {
ggml_vk_create_queue(device, device->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false);
// Shaders
+ // Disable matmul tile sizes early if performance low or not supported
+ switch (device->vendor_id) {
+#ifndef GGML_VULKAN_RUN_TESTS
+ case VK_VENDOR_ID_AMD:
+ case VK_VENDOR_ID_INTEL:
+ device->mul_mat_l = false;
+ device->mul_mat_m = true;
+ device->mul_mat_s = true;
+ device->mul_mat_id_l = false;
+ device->mul_mat_id_m = true;
+ device->mul_mat_id_s = true;
+ break;
+ case VK_VENDOR_ID_APPLE:
+ device->mul_mat_l = false;
+ device->mul_mat_m = true;
+ device->mul_mat_s = false;
+ device->mul_mat_id_l = false;
+ device->mul_mat_id_m = true;
+ device->mul_mat_id_s = false;
+ break;
+#endif
+ default:
+ device->mul_mat_l = true;
+ device->mul_mat_m = true;
+ device->mul_mat_s = true;
+ device->mul_mat_id_l = true;
+ device->mul_mat_id_m = true;
+ device->mul_mat_id_s = true;
+ break;
+ }
+
ggml_vk_load_shaders(device);
if (!device->single_queue) {
bool fp16_storage = false;
bool fp16_compute = false;
+ bool coopmat_support = false;
for (auto properties : ext_props) {
if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
fp16_storage = true;
} else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
fp16_compute = true;
+ } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0) {
+ coopmat_support = true;
}
}
+ if (props2.properties.vendorID == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && driver_props.driverID == vk::DriverId::eAmdProprietary)) {
+ // Intel drivers don't support coopmat properly yet
+ // Only RADV supports coopmat properly on AMD
+ coopmat_support = false;
+ }
+
const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16");
bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr;
vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES;
vk11_features.pNext = &vk12_features;
+ // Pointer to the last chain element
+ VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_features;
+
+ VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features;
+ coopmat_features.pNext = nullptr;
+ coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR;
+ coopmat_features.cooperativeMatrix = VK_FALSE;
+
+ if (coopmat_support) {
+ last_struct->pNext = (VkBaseOutStructure *)&coopmat_features;
+ last_struct = (VkBaseOutStructure *)&coopmat_features;
+ }
+
vkGetPhysicalDeviceFeatures2(physical_device, &device_features2);
fp16 = fp16 && vk12_features.shaderFloat16;
+ coopmat_support = coopmat_support && coopmat_features.cooperativeMatrix;
+
std::string device_name = props2.properties.deviceName.data();
- GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu\n",
- idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size);
+ GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | matrix cores: %d\n",
+ idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size, coopmat_support);
if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
GGML_LOG_DEBUG("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n");
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
return ctx->device->pipeline_matmul_f32_f16;
}
- if (prec == GGML_PREC_DEFAULT && ctx->device->coopmat2) {
+ if (prec == GGML_PREC_DEFAULT && ctx->device->fp16) {
if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
return ctx->device->pipeline_matmul_f16_f32.f16acc;
}
assert(src1_type == GGML_TYPE_F16);
return ctx->device->pipeline_dequant_mul_mat_mat_f16[src0_type].f16acc;
}
- return ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
+ return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
}
static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type];
}
-static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) {
+static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_id_pipeline()");
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
return ctx->device->pipeline_matmul_id_f32;
}
- if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
- return ctx->device->pipeline_matmul_id_f16_f32;
- }
- if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
- return ctx->device->pipeline_matmul_id_f16;
+ if (prec == GGML_PREC_DEFAULT && ctx->device->fp16) {
+ if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+ return ctx->device->pipeline_matmul_id_f16_f32.f16acc;
+ }
+ if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+ return ctx->device->pipeline_matmul_id_f16.f16acc;
+ }
+ } else {
+ if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+ return ctx->device->pipeline_matmul_id_f16_f32.f32acc;
+ }
+ if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+ return ctx->device->pipeline_matmul_id_f16.f32acc;
+ }
}
GGML_ASSERT(src1_type == GGML_TYPE_F32);
return nullptr;
}
- return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type];
+ return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
}
static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
return split_k;
}
-static vk_pipeline ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
- if (m <= 32 || n <= 32) {
- return aligned ? mmp->a_s : mmp->s;
- }
- return aligned ? mmp->a_m : mmp->m;
-
- GGML_UNUSED(ctx);
-}
-
-static vk_pipeline ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) {
- return aligned ? mmp->a_m : mmp->m;
-
- GGML_UNUSED(ctx);
-}
-
-static vk_pipeline ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) {
- return aligned ? mmp->a_s : mmp->s;
-
- GGML_UNUSED(ctx);
-}
-
-static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
+static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned, ggml_type type_a) {
VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")");
- switch (ctx->device->vendor_id) {
- case VK_VENDOR_ID_AMD:
- return ggml_vk_guess_matmul_pipeline_amd(ctx, mmp, m, n, aligned);
- case VK_VENDOR_ID_APPLE:
- return ggml_vk_guess_matmul_pipeline_apple(ctx, mmp, aligned);
- case VK_VENDOR_ID_INTEL:
- return ggml_vk_guess_matmul_pipeline_intel(ctx, mmp, aligned);
- default:
- break;
- }
if (ctx->device->coopmat2) {
- if ((m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) {
+ if ((ctx->device->mul_mat_l && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_m && !ctx->device->mul_mat_s)) {
return aligned ? mmp->a_l : mmp->l;
}
- if ((m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) {
+ if ((ctx->device->mul_mat_m && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_s) {
return aligned ? mmp->a_m : mmp->m;
}
return aligned ? mmp->a_s : mmp->s;
}
- if (m <= 32 || n <= 32) {
+ if ((ctx->device->mul_mat_s && (m <= 32 || n <= 32)) || (!ctx->device->mul_mat_m && !ctx->device->mul_mat_l)) {
return aligned ? mmp->a_s : mmp->s;
}
- if (m <= 64 || n <= 64) {
+ if ((ctx->device->mul_mat_m && (m <= 64 || n <= 64)) || !ctx->device->mul_mat_l) {
return aligned ? mmp->a_m : mmp->m;
}
return aligned ? mmp->a_l : mmp->l;
}
-static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n) {
+static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, ggml_type type_a) {
VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")");
- return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true)->align;
+ return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true, type_a)->align;
}
static void ggml_vk_matmul(
ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 });
}
+static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
+ VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")");
+
+ if (ctx->device->coopmat2) {
+ if ((ctx->device->mul_mat_id_l && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_id_m && !ctx->device->mul_mat_id_s)) {
+ return aligned ? mmp->a_l : mmp->l;
+ }
+ if ((ctx->device->mul_mat_id_m && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_id_s) {
+ return aligned ? mmp->a_m : mmp->m;
+ }
+ return aligned ? mmp->a_s : mmp->s;
+ }
+
+ if ((ctx->device->mul_mat_id_s && (m <= 32 || n <= 32)) || (!ctx->device->mul_mat_id_m && !ctx->device->mul_mat_id_l)) {
+ return aligned ? mmp->a_s : mmp->s;
+ }
+ if ((ctx->device->mul_mat_id_m && (m <= 64 || n <= 64)) || !ctx->device->mul_mat_id_l) {
+ return aligned ? mmp->a_m : mmp->m;
+ }
+ return aligned ? mmp->a_l : mmp->l;
+}
+
+static uint32_t ggml_vk_guess_matmul_id_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n) {
+ VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")");
+ return ggml_vk_guess_matmul_id_pipeline(ctx, mmp, m, n, true)->align;
+}
+
static void ggml_vk_matmul_id(
ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline,
vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids,
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
- const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11));
+ const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, src0->type));
const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8;
- vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned);
+ vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, src0->type);
const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline);
const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
- vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type);
+ vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
const uint64_t y_ne = ne11 * ne10;
const uint64_t d_ne = ne21 * ne20;
- const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, nei1));
+ const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1));
const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
- vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, nei1, aligned);
+ vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned);
const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
for (size_t i = 0; i < x_ne; i++) {
if (std::is_same<float, X_TYPE>()) {
x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
+ // x[i] = 1.0f;
+ // x[i] = i + 1;
+ // x[i] = (i % k == i / k) ? 1.0f : 0.0f;
} else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
+ // x[i] = ggml_fp32_to_fp16(1.0f);
+ // x[i] = ggml_fp32_to_fp16(i + 1);
+ // x[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
} else {
GGML_ABORT("fatal error");
}
}
for (size_t i = 0; i < y_ne; i++) {
if (std::is_same<float, Y_TYPE>()) {
- // y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
- y[i] = (i % k == i / k) ? 1.0f : 0.0f;
+ y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
+ // y[i] = (i % k == i / k) ? 1.0f : 0.0f;
+ // y[i] = i + 1;
} else if (std::is_same<ggml_fp16_t, Y_TYPE>()) {
- // y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
- y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
+ y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
+ // y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
+ // y[i] = ggml_fp32_to_fp16(i + 1);
} else {
GGML_ABORT("fatal error");
}
double err = std::fabs(d[i] - d_chk[i]);
avg_err += err;
- if (err > 0.05f && first_err_n == -1) {
+ if ((err > 0.05f || std::isnan(err)) && first_err_n == -1) {
first_err_b = i / (m * n);
first_err_n = (i % (m * n)) / m;
first_err_m = (i % (m * n)) % m;
std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl;
- if (avg_err > 0.1) {
+ if (avg_err > 0.1 || std::isnan(avg_err)) {
std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
std::cerr << "Actual result: " << std::endl << std::endl;
ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
- std::cerr << std::endl;
- ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n + 15, first_err_b);
std::cerr << "Expected result: " << std::endl << std::endl;
ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
vk_pipeline p;
std::string shname;
if (shader_size == 0) {
- p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s;
+ p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_s : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s;
shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S";
} else if (shader_size == 1) {
- p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m;
+ p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_m : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m;
shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M";
} else if (shader_size == 2) {
- p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l;
+ p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_l : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l;
shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L";
} else {
GGML_ASSERT(0);
if (k != kpad) {
if (shader_size == 0) {
- p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s;
+ p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->s : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s;
shname = std::string(ggml_type_name(quant)) + "_S";
} else if (shader_size == 1) {
- p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m;
+ p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->m : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m;
shname = std::string(ggml_type_name(quant)) + "_M";
} else if (shader_size == 2) {
- p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l;
+ p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->l : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l;
shname = std::string(ggml_type_name(quant)) + "_L";
} else {
GGML_ASSERT(0);
static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
#if defined(GGML_VULKAN_RUN_TESTS)
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_0);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_1);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q8_0);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q2_K);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q3_K);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_K);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_K);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q6_K);
- ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_IQ4_NL);
-
- ggml_vk_test_matmul<ggml_fp16_t, ggml_fp16_t>(ctx, 512, 512, 100, 32, 100, 1, 2);
-
- 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, 1, GGML_TYPE_Q4_0);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, 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, 2, GGML_TYPE_Q4_0);
-
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_1);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_1);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_1);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_1);
- // 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);
-
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q2_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q2_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q2_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q2_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q2_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q2_K);
-
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q3_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q3_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q3_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q3_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q3_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q3_K);
-
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_K);
-
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_K);
-
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q6_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q6_K);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q6_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q6_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q6_K);
- // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q6_K);
-
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_IQ4_NL);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_IQ4_NL);
- ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_IQ4_NL);
-
- std::cerr << std::endl;
-
const std::vector<size_t> vals {
+ 512, 512, 128,
+ 128, 512, 512,
+ 4096, 512, 4096,
+ 11008, 512, 4096,
+ 4096, 512, 11008,
+ 32000, 512, 4096,
8, 8, 8,
100, 46, 576,
623, 111, 128,
49, 49, 128,
128, 49, 49,
4096, 49, 4096,
- 11008, 49, 4096,
- 4096, 49, 11008,
- 32000, 49, 4096,
- 512, 512, 128,
- 128, 512, 512,
- 4096, 512, 4096,
- 11008, 512, 4096,
- 4096, 512, 11008,
- 32000, 512, 4096,
};
const size_t num_it = 100;
ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0);
ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1);
ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2);
- // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0);
- // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1);
- // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2);
- std::cerr << std::endl;
+ std::cerr << '\n';
+ ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0);
+ ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1);
+ ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2);
+ std::cerr << '\n';
+ ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0);
+ ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1);
+ ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2);
+ std::cerr << '\n' << std::endl;
+
+ if (vals[i + 2] % 32 == 0) {
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0, GGML_TYPE_Q4_0);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1, GGML_TYPE_Q4_0);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2, GGML_TYPE_Q4_0);
+ std::cerr << '\n';
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0, GGML_TYPE_Q4_0);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1, GGML_TYPE_Q4_0);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2, GGML_TYPE_Q4_0);
+ std::cerr << '\n';
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0, GGML_TYPE_Q4_0);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1, GGML_TYPE_Q4_0);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2, GGML_TYPE_Q4_0);
+ std::cerr << '\n' << std::endl;
+ }
+
+ if (vals[i + 2] % 256 == 0) {
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0, GGML_TYPE_Q4_K);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1, GGML_TYPE_Q4_K);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2, GGML_TYPE_Q4_K);
+ std::cerr << '\n';
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0, GGML_TYPE_Q4_K);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1, GGML_TYPE_Q4_K);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2, GGML_TYPE_Q4_K);
+ std::cerr << '\n';
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0, GGML_TYPE_Q4_K);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1, GGML_TYPE_Q4_K);
+ ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2, GGML_TYPE_Q4_K);
+ std::cerr << '\n' << std::endl;
+ }
}
GGML_ABORT("fatal error");
case GGML_OP_MUL_MAT_ID:
{
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
- if (op->op == GGML_OP_MUL_MAT_ID &&
- ggml_vk_get_device(ctx->device)->properties.limits.maxComputeSharedMemorySize < 32768) {
+ const vk_device& device = ggml_vk_get_device(ctx->device);
+ if (op->op == GGML_OP_MUL_MAT_ID && !device->mul_mat_id_s && !device->mul_mat_id_m && !device->mul_mat_id_l) {
// If there's not enough shared memory for row_ids and the result tile, fallback to CPU
return false;
}
overview / pkg / ggml / sources / whisper.cpp / commitdiff