ggml : add MUL_MAT src1 precision tests and gating (#31896)
This commit is contained in:
1 parent
281e7dad7c
commit
494d0f47bd
19 files changed
+192
-61
No files matched your search
@@ -271,46 +271,6 @@ mtmd_subcontext::support_info_graph mtmd_subcontext::reserve_compute_meta(const
|
||||
return res;
|
||||
}
|
||||
|
||||
// TODO: remove this and use ggml_mul_mat with GGML_PREC_F32 once all backends honor it for F32 x F32
|
||||
// ggml_mul_mat for 2D F32 tensors, always in F32 on CPU: some GPU backends round the inputs to F16 or TF32, which loses the quiet bins of a spectrum
|
||||
static void mtmd_mul_mat_prec_op(ggml_tensor * dst, int ith, int nth, void * /*userdata*/) {
|
||||
const ggml_tensor * a = dst->src[0];
|
||||
const ggml_tensor * b = dst->src[1];
|
||||
const int64_t n_k = a->ne[0];
|
||||
|
||||
for (int64_t n = ith; n < b->ne[1]; n += nth) {
|
||||
const float * b_row = (const float *) ((const char *) b->data + n * b->nb[1]);
|
||||
float * dst_row = (float *) ((char *) dst->data + n * dst->nb[1]);
|
||||
for (int64_t m = 0; m < a->ne[1]; m++) {
|
||||
const float * a_row = (const float *) ((const char *) a->data + m * a->nb[1]);
|
||||
// independent partial sums, so that the compiler can vectorize the loop
|
||||
float acc[16] = {};
|
||||
int64_t k = 0;
|
||||
for (; k + 16 <= n_k; k += 16) {
|
||||
for (int l = 0; l < 16; l++) {
|
||||
acc[l] += a_row[k + l] * b_row[k + l];
|
||||
}
|
||||
}
|
||||
float sum = 0.0f;
|
||||
for (; k < n_k; k++) {
|
||||
sum += a_row[k] * b_row[k];
|
||||
}
|
||||
for (int l = 0; l < 16; l++) {
|
||||
sum += acc[l];
|
||||
}
|
||||
dst_row[m] = sum;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
static ggml_tensor * mtmd_mul_mat_prec(ggml_context * ctx, ggml_tensor * a, ggml_tensor * b) {
|
||||
GGML_ASSERT(a->type == GGML_TYPE_F32 && b->type == GGML_TYPE_F32);
|
||||
GGML_ASSERT(ggml_n_dims(a) <= 2 && ggml_n_dims(b) <= 2 && a->ne[0] == b->ne[0]);
|
||||
GGML_ASSERT(a->nb[0] == sizeof(float) && b->nb[0] == sizeof(float));
|
||||
ggml_tensor * args[] = { a, b };
|
||||
return ggml_custom_4d(ctx, GGML_TYPE_F32, a->ne[1], b->ne[1], 1, 1, args, 2, mtmd_mul_mat_prec_op, GGML_N_TASKS_MAX, nullptr);
|
||||
}
|
||||
|
||||
bool mtmd_audio_log_mel(const mtmd_audio_mel_config & config, const float * samples, const float * window, int64_t n_frames, float * out) {
|
||||
ggml_backend_sched_t sched = config.sched;
|
||||
ggml_backend_t backend_cpu = ggml_backend_sched_get_backend(sched, ggml_backend_sched_get_n_backends(sched) - 1);
|
||||
@@ -356,7 +316,8 @@ bool mtmd_audio_log_mel(const mtmd_audio_mel_config & config, const float * samp
|
||||
ggml_set_input(inp_filters);
|
||||
|
||||
ggml_tensor * cur = ggml_cont(ctx0, ggml_view_2d(ctx0, inp_samples, frame_size, n_cur, config.hop_length * ggml_element_size(inp_samples), 0));
|
||||
cur = mtmd_mul_mat_prec(ctx0, inp_dft, cur); // [2 * n_fft_bins, n_cur]
|
||||
cur = ggml_mul_mat(ctx0, inp_dft, cur); // [2 * n_fft_bins, n_cur]
|
||||
ggml_prec_set_src(cur, GGML_PREC_F32, 1);
|
||||
|
||||
cur = ggml_sqr(ctx0, cur);
|
||||
cur = ggml_add(ctx0,
|
||||
@@ -366,7 +327,8 @@ bool mtmd_audio_log_mel(const mtmd_audio_mel_config & config, const float * samp
|
||||
cur = ggml_sqrt(ctx0, cur);
|
||||
}
|
||||
|
||||
cur = mtmd_mul_mat_prec(ctx0, cur, inp_filters); // [n_cur, n_mel]
|
||||
cur = ggml_mul_mat(ctx0, cur, inp_filters); // [n_cur, n_mel]
|
||||
ggml_prec_set_src(cur, GGML_PREC_F32, 1);
|
||||
cur = config.mel_floor_add
|
||||
? ggml_scale_bias(ctx0, cur, 1.0f, config.mel_floor)
|
||||
: ggml_clamp(ctx0, cur, config.mel_floor, FLT_MAX);
|
||||
|
||||
Reference in new issue
Block a user