Fixed the issue of OpenBLAS not being enabled on Windows.

[bobqianic]

Fixed the issue of not being able to find OpenBLAS on the Windows platform. Even though the name of the previously released binary file was whisper-blas-bin-x64.zip, BLAS was actually not enabled. After enabling, the inference speed can increase by 3-4 times. When compiling on Windows, all you need to do is download the OpenBLAS binary package, unzip it to any directory, and add the package's path to a system variable called OPENBLAS_PATH.

Before:

C:\Users\qianp>C:\Users\qianp\Downloads\whisper-blas-bin-x64\bench.exe -m C:\Users\qianp\Downloads\ggml-m
odel-largev2.bin -w 0 -t 4
whisper_init_from_file_no_state: loading model from 'C:\Users\qianp\Downloads\ggml-model-largev2.bin'
whisper_model_load: loading model
whisper_model_load: n_vocab       = 51865
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 1280
whisper_model_load: n_audio_head  = 20
whisper_model_load: n_audio_layer = 32
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 1280
whisper_model_load: n_text_head   = 20
whisper_model_load: n_text_layer  = 32
whisper_model_load: n_mels        = 80
whisper_model_load: ftype         = 1
whisper_model_load: type          = 5
whisper_model_load: mem required  = 3557.00 MB (+   71.00 MB per decoder)
whisper_model_load: adding 1608 extra tokens
whisper_model_load: model ctx     = 2950.97 MB
whisper_model_load: model size    = 2950.66 MB
whisper_init_state: kv self size  =   70.00 MB
whisper_init_state: kv cross size =  234.38 MB

system_info: n_threads = 4 / 20 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 | COREML = 0 |

whisper_print_timings:     load time =  2369.90 ms
whisper_print_timings:     fallbacks =   0 p /   0 h
whisper_print_timings:      mel time =     0.00 ms
whisper_print_timings:   sample time =     0.00 ms /     1 runs (    0.00 ms per run)
whisper_print_timings:   encode time = 48837.34 ms /     1 runs (48837.34 ms per run)
whisper_print_timings:   decode time =     0.00 ms /     1 runs (    0.00 ms per run)
whisper_print_timings:    total time = 51340.71 ms

If you wish, you can submit these results here:

  https://github.com/ggerganov/whisper.cpp/issues/89

Please include the following information:

  - CPU model
  - Operating system
  - Compiler

C:\Users\qianp>C:\Users\qianp\Downloads\whisper-blas-bin-x64\bench.exe -m C:\Users\qianp\Downloads\ggml-model-largev2.bin -w 2 -t 4
ggml_mul_mat:   64 x   64: Q4_0     2.9 GFLOPS (128 runs) / Q4_1     3.3 GFLOPS (128 runs) / F16     3.5 GFLOPS (128 runs) / F32     3.3 GFLOPS (128 runs)
ggml_mul_mat:  128 x  128: Q4_0    15.4 GFLOPS (128 runs) / Q4_1    20.7 GFLOPS (128 runs) / F16    23.9 GFLOPS (128 runs) / F32    15.1 GFLOPS (128 runs)
ggml_mul_mat:  256 x  256: Q4_0    33.9 GFLOPS (128 runs) / Q4_1    77.0 GFLOPS (128 runs) / F16    96.2 GFLOPS (128 runs) / F32    31.8 GFLOPS (128 runs)
ggml_mul_mat:  512 x  512: Q4_0    63.9 GFLOPS (128 runs) / Q4_1   107.9 GFLOPS (128 runs) / F16   128.5 GFLOPS (128 runs) / F32    36.8 GFLOPS (128 runs)
ggml_mul_mat: 1024 x 1024: Q4_0    80.9 GFLOPS ( 38 runs) / Q4_1   125.8 GFLOPS ( 59 runs) / F16   131.1 GFLOPS ( 62 runs) / F32    35.7 GFLOPS ( 17 runs)
ggml_mul_mat: 2048 x 2048: Q4_0    86.3 GFLOPS (  6 runs) / Q4_1   129.8 GFLOPS (  8 runs) / F16   135.5 GFLOPS (  8 runs) / F32    33.3 GFLOPS (  3 runs)
ggml_mul_mat: 4096 x 4096: Q4_0    89.2 GFLOPS (  3 runs) / Q4_1    97.2 GFLOPS (  3 runs) / F16    62.3 GFLOPS (  3 runs) / F32    20.6 GFLOPS (  3 runs)

After:

C:\Users\qianp>C:\Users\qianp\Downloads\whisper_build\bin\Release\bench.exe -m C:\Users\qianp\Downloads\ggml-model-largev2.bin -w 0 -t 4
whisper_init_from_file_no_state: loading model from 'C:\Users\qianp\Downloads\ggml-model-largev2.bin'
whisper_model_load: loading model
whisper_model_load: n_vocab       = 51865
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 1280
whisper_model_load: n_audio_head  = 20
whisper_model_load: n_audio_layer = 32
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 1280
whisper_model_load: n_text_head   = 20
whisper_model_load: n_text_layer  = 32
whisper_model_load: n_mels        = 80
whisper_model_load: ftype         = 1
whisper_model_load: qntvr         = 0
whisper_model_load: type          = 5
whisper_model_load: mem required  = 3557.00 MB (+   71.00 MB per decoder)
whisper_model_load: adding 1608 extra tokens
whisper_model_load: model ctx     = 2951.27 MB
whisper_model_load: model size    = 2950.66 MB
whisper_init_state: kv self size  =   70.00 MB
whisper_init_state: kv cross size =  234.38 MB

system_info: n_threads = 4 / 20 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | VSX = 0 | COREML = 0 | OPENVINO = 0 |

whisper_print_timings:     load time =  1739.12 ms
whisper_print_timings:     fallbacks =   0 p /   0 h
whisper_print_timings:      mel time =     0.00 ms
whisper_print_timings:   sample time =     0.00 ms /     1 runs (    0.00 ms per run)
whisper_print_timings:   encode time = 11481.76 ms /     1 runs (11481.76 ms per run)
whisper_print_timings:   decode time =     0.00 ms /     1 runs (    0.00 ms per run)
whisper_print_timings:    total time = 13400.80 ms

If you wish, you can submit these results here:

  https://github.com/ggerganov/whisper.cpp/issues/89

Please include the following information:

  - CPU model
  - Operating system
  - Compiler
  
C:\Users\qianp>C:\Users\qianp\Downloads\whisper_build\bin\Release\bench.exe -m C:\Users\qianp\Downloads\ggml-model-largev2.bin -w 2 -t 4
  64 x   64: Q4_0     3.2 GFLOPS (128 runs) | Q4_1     3.4 GFLOPS (128 runs)
  64 x   64: Q5_0     3.3 GFLOPS (128 runs) | Q5_1     3.3 GFLOPS (128 runs) | Q8_0     3.3 GFLOPS (128 runs)
  64 x   64: F16      3.1 GFLOPS (128 runs) | F32      2.9 GFLOPS (128 runs)
 128 x  128: Q4_0     6.3 GFLOPS (128 runs) | Q4_1     5.7 GFLOPS (128 runs)
 128 x  128: Q5_0     5.8 GFLOPS (128 runs) | Q5_1     5.7 GFLOPS (128 runs) | Q8_0     5.3 GFLOPS (128 runs)
 128 x  128: F16      6.6 GFLOPS (128 runs) | F32      6.5 GFLOPS (128 runs)
 256 x  256: Q4_0    41.3 GFLOPS (128 runs) | Q4_1    39.1 GFLOPS (128 runs)
 256 x  256: Q5_0    38.0 GFLOPS (128 runs) | Q5_1    37.4 GFLOPS (128 runs) | Q8_0    35.9 GFLOPS (128 runs)
 256 x  256: F16     47.3 GFLOPS (128 runs) | F32     46.4 GFLOPS (128 runs)
 512 x  512: Q4_0   190.8 GFLOPS (128 runs) | Q4_1   188.3 GFLOPS (128 runs)
 512 x  512: Q5_0   187.6 GFLOPS (128 runs) | Q5_1   175.6 GFLOPS (128 runs) | Q8_0   183.0 GFLOPS (128 runs)
 512 x  512: F16    204.0 GFLOPS (128 runs) | F32    180.5 GFLOPS (128 runs)
1024 x 1024: Q4_0   349.2 GFLOPS (128 runs) | Q4_1   345.7 GFLOPS (128 runs)
1024 x 1024: Q5_0   337.9 GFLOPS (128 runs) | Q5_1   331.3 GFLOPS (128 runs) | Q8_0   342.0 GFLOPS (128 runs)
1024 x 1024: F16    350.1 GFLOPS (128 runs) | F32    245.8 GFLOPS (115 runs)
2048 x 2048: Q4_0   446.7 GFLOPS ( 27 runs) | Q4_1   430.4 GFLOPS ( 26 runs)
2048 x 2048: Q5_0   433.3 GFLOPS ( 26 runs) | Q5_1   440.8 GFLOPS ( 26 runs) | Q8_0   447.1 GFLOPS ( 27 runs)
2048 x 2048: F16    447.5 GFLOPS ( 27 runs) | F32    335.9 GFLOPS ( 20 runs)
4096 x 4096: Q4_0   534.0 GFLOPS (  4 runs) | Q4_1   493.6 GFLOPS (  4 runs)
4096 x 4096: Q5_0   498.4 GFLOPS (  4 runs) | Q5_1   502.5 GFLOPS (  4 runs) | Q8_0   513.1 GFLOPS (  4 runs)
4096 x 4096: F16    520.9 GFLOPS (  4 runs) | F32    414.5 GFLOPS (  4 runs)